thesis about discourse

319
DISCOURSE SEMANTICS OF S-MODIFYING ADVERBIALS Katherine M. Forbes A DISSERTATION in Linguistics Presented to the Faculties of the University of Pennsylvania in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy 2003 Bonnie Webber, Supervisor of Dissertation Ellen Prince, Supervisor of Dissertation Donald A. Ringe, Graduate Group Chair Aravind Joshi, Committee Member Robin Clark, Committee Member

Upload: blessedkkr

Post on 06-May-2015

2.699 views

Category:

Technology


0 download

TRANSCRIPT

Page 1: Thesis about discourse

DISCOURSESEMANTICSOFS-MODIFYING ADVERBIALS

KatherineM. Forbes

A DISSERTATION

in

Linguistics

Presentedto theFacultiesof theUniversityof Pennsylvaniain Partial

Fulfillment of theRequirementsfor theDegreeof Doctorof Philosophy

2003

BonnieWebber, Supervisorof Dissertation Ellen Prince,Supervisorof Dissertation

DonaldA. Ringe,GraduateGroupChair Aravind Joshi,CommitteeMember

RobinClark,CommitteeMember

Page 2: Thesis about discourse

Acknowledgements

I wish to thankBonnieWebber. Without herpatienceandherseeminglyendlessdepthsof insight,

I might never have completedthis thesis.I amenormouslygratefulfor herguidance.

I alsoowe many thanksto Ellen Prince. Sheis an intellectualleaderat Pennwho hashelped

many, includingme,find away throughthejungleof discourseanalysis.

I amindebtedto every professorwho hastaughtme. Specialthanksto RobinClark for beinga

memberof my dissertationcommittee.

I amverylucky to haveworkedwith Aravind Joshi.Heis acontinualsourceof knowledgein the

DLTAG meetings.Thefield of computationallinguisticshasalreadybenefitedfrom his sentence-

level work; I fully expectheandBonniewill producesimilarly usefulresultswith DLTAG.

Also in DLTAG, Eleni Miltsakaki andRashmiPrasad,andlaterCassandreCreswellandJason

Teepleall providedstimulationandsolace.Theirgreatcompany andgreateffort onDLTAGprojects

taughtme to appreciatehow muchcanbe donewhenmindswork together. I look forward to the

chanceto work with themin thefuture.

I amalsothankfulto MarthaPalmer, PaulKingsbury, andScottCottonfor allowing meto work

with themon thePropbankprojectandsupplementbothmy incomeandmy work in discourse.

Onapersonalnote,theForbes,Finley, andRiley familiesdeserve thanksfor giving meloveand

diversionandbalanceandtalkingmethroughmy education.Mostof all, thanksto EnricoRiley, for

beingeverythingto me.

ii

Page 3: Thesis about discourse

ABSTRACT

DISCOURSESEMANTICSOFS-MODIFYING ADVERBIALS

KatherineM. Forbes

Supervisors:BonnieWebberandEllen Prince

In this thesis,we addressthequestionof why certainS-modifyingadverbialsareonly interpretable

with respectto thediscourseor spatio-temporalcontext, andnot just their own matrix clause.It is

not possibleto list theseadverbialsbecausethesetof adverbialsis compositionalandthereforein-

finite. Instead,we investigatethemechanismsunderlyingtheir interpretation.Wepresentacorpus-

basedanalysisof the predicateargumentstructureandinterpretationof over 13,000S-modifying

adverbials. We useprior researchon discoursedeixis andclause-level predicatesto studythe se-

manticsof theargumentsof S-modifyingadverbialsandthesyntacticconstituentsfrom which they

canbederived. We show thatmany S-modifyingadverbialscontainsemanticargumentsthatmay

notbesyntacticallyovert,but whoseinterpretationneverthelessrequiresanabstractobjectfrom the

discourseor spatio-temporalcontext. Prior work hasinvestigatedonly a smallsubsetof thesedis-

courseconnectives; at theclause-level theirsemanticshasbeenlargely ignoredandat thediscourse

level they areusuallytreatedas“signals”of predefinedlists of abstractdiscourserelations.Our in-

vestigationshedslight on thespaceof relationsimpartedby amuchwidervarietyof adverbials.We

furthershow how theirpredicateargumentstructureandinterpretationcanbeformalizedandincor-

poratedinto a rich intermediatemodelof discoursethataloneamongothermodelsviews discourse

connectivesaspredicateswhosesyntaxandsemanticsmustbespecifiedandrecoverableto interpret

discourse.It is notonly dueto theirargumentstructureandinterpretationthatadverbialshavebeen

treatedasdiscourseconnectives,however. Our corpuscontainsadverbialswhosesemanticsalone

doesnot causethemto beinterpretedwith respectto abstractobjectinterpretationsin thediscourse

or spatio-temporalcontext. Weexploreotherexplanationsfor why theseadverbialsevokediscourse

context for their interpretation;in particular, weshow how theinteractionof prosodywith theinter-

pretationof S-modifyingadverbialscancontribute to discoursecoherence,andwe alsoshow how

S-modifyingadverbialscanbeusedto convey implicatures.

iii

Page 4: Thesis about discourse

Contents

Acknowledgements ii

Abstract iii

Contents iv

List of Tables x

List of Figures xiv

1 Intr oduction 1

1.1 TheProblem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Contributionsof theThesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 ThesisOutline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Anaphora and DiscourseModels 6

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Descriptive Theoriesof DiscourseCoherence . . . . . . . . . . . . . . . . . . . . 8

2.2.1 An Early EncompassingDescription. . . . . . . . . . . . . . . . . . . . . 8

2.2.2 Alternative Descriptionsof PropositionalRelations . . . . . . . . . . . . . 10

2.2.3 DiscourseRelationsasConstraints. . . . . . . . . . . . . . . . . . . . . . 12

2.2.4 AbducingDiscourseRelationsby Applying theConstraints . . . . . . . . 14

2.2.5 Interactionof DiscourseInferenceandVP Ellipsis . . . . . . . . . . . . . 17

iv

Page 5: Thesis about discourse

2.2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.3 A Three-TieredModelof Discourse . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.3.1 TheThreeTiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.3.2 Coherencewithin DiscourseSegments. . . . . . . . . . . . . . . . . . . . 22

2.3.3 ModelingLinguisticStructureandAttentionalStateasaTree . . . . . . . 23

2.3.4 Introductionto DiscourseDeictic Reference. . . . . . . . . . . . . . . . . 25

2.3.5 Retrieving Antecedentsof DiscourseDeixis from theTree . . . . . . . . . 27

2.3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.4 A TreeStructurewith aSyntax-SemanticInterface . . . . . . . . . . . . . . . . . 33

2.4.1 ConstituentsandTreeConstruction . . . . . . . . . . . . . . . . . . . . . 33

2.4.2 TheSyntaxSemanticInterface . . . . . . . . . . . . . . . . . . . . . . . . 34

2.4.3 Retrieving Antecedentsof Anaphorafrom theTree . . . . . . . . . . . . . 36

2.4.4 TheNeedFor UpwardPercolation. . . . . . . . . . . . . . . . . . . . . . 36

2.4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.5 A Descriptive Theoryof DiscourseStructure. . . . . . . . . . . . . . . . . . . . . 38

2.5.1 AnalyzingText Structure. . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.5.2 TheNeedfor Multiple Levelsof DiscourseStructure . . . . . . . . . . . . 42

2.5.3 “Elaboration”asReference. . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

2.6 A SemanticTheoryof DiscourseCoherence. . . . . . . . . . . . . . . . . . . . . 48

2.6.1 AbstractObjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.6.2 A FormalLanguagefor Discourse. . . . . . . . . . . . . . . . . . . . . . 53

2.6.3 Retrieving Antecedentsof Anaphorafrom theDiscourseStructure. . . . . 57

2.6.4 A Systemfor InferringDiscourseRelations . . . . . . . . . . . . . . . . . 57

2.6.5 ExtendingtheTheoryto Cognitive States . . . . . . . . . . . . . . . . . . 62

2.6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

2.7 Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

2.7.1 Proliferationof DiscourseRelations . . . . . . . . . . . . . . . . . . . . . 64

v

Page 6: Thesis about discourse

2.7.2 Useof LinguisticCuesasSignals . . . . . . . . . . . . . . . . . . . . . . 66

2.7.3 StructuralandAnaphoricCuePhrases. . . . . . . . . . . . . . . . . . . . 69

2.7.4 Comparisonof DLTAG andOtherModels . . . . . . . . . . . . . . . . . . 71

2.7.5 RemainingQuestions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3 SemanticMechanismsin Adverbials 78

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

3.2 Linguistic BackgroundandDataCollection . . . . . . . . . . . . . . . . . . . . . 79

3.2.1 Functionof Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

3.2.2 Structureof PPandADVP . . . . . . . . . . . . . . . . . . . . . . . . . . 81

3.2.3 DataCollection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

3.3 AdverbialModificationTypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

3.3.1 Clause-Level Analysesof ModificationType . . . . . . . . . . . . . . . . 87

3.3.2 Problemswith CategoricalApproaches . . . . . . . . . . . . . . . . . . . 91

3.3.3 ModificationTypesasSemanticFeatures . . . . . . . . . . . . . . . . . . 92

3.3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

3.4 AdverbialSemanticArguments. . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

3.4.1 (Optional)Argumentsor Adjuncts? . . . . . . . . . . . . . . . . . . . . . 94

3.4.2 ExternalArgumentAttachmentAmbiguity . . . . . . . . . . . . . . . . . 98

3.4.3 SemanticRepresentationof ExternalArgument . . . . . . . . . . . . . . . 103

3.4.4 SemanticArgumentsasAbstractObjects . . . . . . . . . . . . . . . . . . 104

3.4.5 Numberof AbstractObjects . . . . . . . . . . . . . . . . . . . . . . . . . 107

3.4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

3.5 S-Modifying PPAdverbials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

3.5.1 ProperNouns,Possessives,andPronouns . . . . . . . . . . . . . . . . . . 111

3.5.2 Demonstrative andDefiniteDeterminers. . . . . . . . . . . . . . . . . . . 114

3.5.3 IndefiniteArticles,GenericandPluralNouns,andOptionalArguments . . 117

vi

Page 7: Thesis about discourse

3.5.4 PPandADJPModifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

3.5.5 OtherArguments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

3.5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

3.6 S-Modifying ADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

3.6.1 SyntacticallyOptionalArguments . . . . . . . . . . . . . . . . . . . . . . 135

3.6.2 Context-DependentADVP Adverbials. . . . . . . . . . . . . . . . . . . . 139

3.6.3 Comparative ADVP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

3.6.4 SetsandWorlds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

3.6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

4 Incorporating Adverbial Semanticsinto DLTAG 157

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

4.2 Syntax-SemanticInterfacesat theSentenceLevel . . . . . . . . . . . . . . . . . . 158

4.2.1 TheRoleof theSyntax-SemanticInterface . . . . . . . . . . . . . . . . . 158

4.2.2 LTAG: LexicalizedTreeAdjoining Grammar . . . . . . . . . . . . . . . . 159

4.2.3 A Syntax-SemanticInterfacefor LTAG DerivationTrees . . . . . . . . . . 161

4.2.4 A Syntax-SemanticInterfacefor LTAG ElementaryTrees . . . . . . . . . 166

4.2.5 Comparisonof Approaches. . . . . . . . . . . . . . . . . . . . . . . . . . 169

4.2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

4.3 Syntax-SemanticInterfacesat theDiscourseLevel . . . . . . . . . . . . . . . . . . 171

4.3.1 DLTAG: LexicalizedTreeAdjoining Grammarfor Discourse. . . . . . . . 171

4.3.2 Syntax-SemanticInterfacesfor DerivedTrees. . . . . . . . . . . . . . . . 179

4.3.3 A Syntax-SemanticInterfacefor DLTAG DerivationTrees . . . . . . . . . 190

4.3.4 Comparisonof Approaches. . . . . . . . . . . . . . . . . . . . . . . . . . 218

4.3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

4.4 DLTAG AnnotationProject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

4.4.1 Overview of Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

4.4.2 PreliminaryStudy1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222

vii

Page 8: Thesis about discourse

4.4.3 PreliminaryStudy2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

4.4.4 FutureWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

5 Other WaysAdverbials Contrib ute to DiscourseCoherence 229

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

5.2 Focus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230

5.2.1 ThePhenomena. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230

5.2.2 Information-StructureandTheoriesof StructuredMeanings . . . . . . . . 232

5.2.3 Alternative Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

5.2.4 Backgroundsor Alternatives? . . . . . . . . . . . . . . . . . . . . . . . . 237

5.2.5 Contrastive Themes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

5.2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240

5.3 FocusSensitivity of Modifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240

5.3.1 FocusParticles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241

5.3.2 OtherFocusSensitive Sub-ClausalModifiers . . . . . . . . . . . . . . . . 244

5.3.3 S-Modifying “FocusParticles” . . . . . . . . . . . . . . . . . . . . . . . . 249

5.3.4 FocusSensivity of S-Modifying Adverbials . . . . . . . . . . . . . . . . . 254

5.3.5 FocusingS-Modifying Adverbialsto Evoke Context . . . . . . . . . . . . 258

5.3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

5.4 Implicatures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262

5.4.1 GriceanImplicature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262

5.4.2 PragmaticandSemanticPresupposition. . . . . . . . . . . . . . . . . . . 267

5.4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270

5.5 UsingS-Modifying Adverbialsto Convey Implicatures . . . . . . . . . . . . . . . 270

5.5.1 Presupposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270

5.5.2 ConversationalImplicatures . . . . . . . . . . . . . . . . . . . . . . . . . 271

5.5.3 Interactionof FocusandImplicature . . . . . . . . . . . . . . . . . . . . . 277

5.5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277

viii

Page 9: Thesis about discourse

5.6 OtherContributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

5.6.1 DiscourseStructure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

5.6.2 Performatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

6 Conclusion 279

6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279

6.2 FutureDirections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281

Bibliography 285

ix

Page 10: Thesis about discourse

List of Tables

2.1 Main Categoriesof [HH76]’s RelationsbetweenPropositions. . . . . . . . . . . . 10

2.2 Main Categoriesof [Lon83]’s RelationsbetweenPropositions . . . . . . . . . . . 10

2.3 Main Categoriesof [Mar92]’s RelationsbetweenPropositions . . . . . . . . . . . 11

2.4 Main Categoriesof [Hob90]’s RelationsbetweenPropositions . . . . . . . . . . . 12

2.5 [Keh95]’sCause-Effect Relations. . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.6 [Keh95]’sResemblanceRelations . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.7 [GS86] Changesin DiscourseStructureIndicatedby LinguisticExpressions. . . . 21

2.8 CenteringTheoryTransitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.9 [Web91]’s Classificationof DiscourseDeicticReference . . . . . . . . . . . . . . 26

2.10 Organizationsof RSTRelationDefinitions . . . . . . . . . . . . . . . . . . . . . 39

2.11 Evidence:RSTRelationDefinition . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.12 Volitional-Cause:RSTRelationDefinition . . . . . . . . . . . . . . . . . . . . . . 42

2.13 Elaboration:RSTRelationDefinition . . . . . . . . . . . . . . . . . . . . . . . . 44

2.14 [Ven67]’s ImperfectandPerfectNominalizations . . . . . . . . . . . . . . . . . . 49

2.15 [Ven67]’s LooseandNarrow Containers. . . . . . . . . . . . . . . . . . . . . . . 50

2.16 DICE: discourserelationdefinitions . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.17 DICE: Indefeasibleaxioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2.18 DICE: Defeasiblelaws on world knowledge . . . . . . . . . . . . . . . . . . . . . 59

2.19 DICE: Defeasiblelaws on discourseprocesses. . . . . . . . . . . . . . . . . . . . 60

2.20 DICE: Deductionrules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

2.21 [Kno96]’s Featuresof DiscourseConnectives . . . . . . . . . . . . . . . . . . . . 68

x

Page 11: Thesis about discourse

3.1 Non-DerivedandDerivedAdverbs . . . . . . . . . . . . . . . . . . . . . . . . . . 82

3.2 tgrep Resultsfor S-AdjoinedADVP andPPin WSJandBrown Corpora. . . . . . 85

3.3 TotalS-AdjoinedAdverbialsin WSJandBrown Corpora . . . . . . . . . . . . . . 86

3.4 [Ale97]’s ModificationTypes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

3.5 [Ern84]’s ModificationTypes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3.6 [KP02]’sModificationTypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3.7 [Gre69]’s SyntacticTestsfor DistinguishingVP andSModification . . . . . . . . 99

3.8 SemanticInterpretationsof [Ern84]’s ModificationTypes . . . . . . . . . . . . . . 103

3.9 AbstractObjectInterpretations. . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

3.10 ApproximateCountsof TokensandTypesof someInternalPParguments . . . . . 111

3.11 PPAdverbialswith ProperNounor YearInternalArgument . . . . . . . . . . . . 111

3.12 PPAdverbialwith Possessive ProperNounInternalArgument . . . . . . . . . . . 112

3.13 PPAdverbialswith PronominalInternalArgument . . . . . . . . . . . . . . . . . 113

3.14 PPAdverbialwith Possessive Pronoun. . . . . . . . . . . . . . . . . . . . . . . . 114

3.15 ApproximateCountsof TokensandTypesof someInternalPParguments . . . . . 114

3.16 PPAdverbialswith DefiniteConcreteObjectInternalArgument . . . . . . . . . . 115

3.17 PPAdverbialswith DefiniteAO InternalArgument . . . . . . . . . . . . . . . . . 115

3.18 PPAdverbialswith Demonstrative ConcreteObjectInternalArgument. . . . . . . 116

3.19 PPAdverbialswith Demonstrative AO InternalArguments . . . . . . . . . . . . . 117

3.20 ApproximateCountsof TokensandTypesof someInternalPParguments . . . . . 118

3.21 PPAdverbialwith IndefiniteConcreteObjectInternalArgument . . . . . . . . . . 118

3.22 PPAdverbialwith IndefiniteAO InternalArgument . . . . . . . . . . . . . . . . . 118

3.23 PPAdverbialwith RelationalIndefiniteAO InternalArgument . . . . . . . . . . . 119

3.24 PPAdverbialswith Genericor PluralConcreteObjectInternalArguments. . . . . 122

3.25 PPAdverbialswith Genericor PluralAO InternalArguments. . . . . . . . . . . . 122

3.26 PPAdverbialswith RelationalGenericAO InternalArguments. . . . . . . . . . . 123

3.27 ApproximateCountsof TokensandTypesof someInternalArgumentModifiers . . 124

3.28 Binary DefiniteInternalArgumentwith OvertArgument . . . . . . . . . . . . . . 125

xi

Page 12: Thesis about discourse

3.29 Binary IndefiniteInternalArgumentwith Overt Argument . . . . . . . . . . . . . 125

3.30 Binary Genericor PluralInternalArgumentwith OvertArgument . . . . . . . . . 126

3.31 InternalArgumentwith aSpatio-TemporalADJ . . . . . . . . . . . . . . . . . . . 126

3.32 InternalArgumentwith ReferentialAdjective . . . . . . . . . . . . . . . . . . . . 127

3.33 InternalArgumentwith Non-ReferentialAdjective . . . . . . . . . . . . . . . . . 128

3.34 InternalArgumentwith DeterminerandNon-ReferentialAdjective . . . . . . . . . 128

3.35 InternalArgumentwith OrdinalAdjective . . . . . . . . . . . . . . . . . . . . . . 129

3.36 InternalArgumentwith Alternative Phrase. . . . . . . . . . . . . . . . . . . . . . 129

3.37 InternalArgumentwith DeterminerandAlternative Phrase . . . . . . . . . . . . . 129

3.38 InternalArgumentwith Comparative/Superlative Adjective . . . . . . . . . . . . . 130

3.39 InternalArgumentwith OtherSet-Evoking Adjectives. . . . . . . . . . . . . . . . 131

3.40 ApproximateCountsof TokensandTypesof someInternalPParguments . . . . . 131

3.41 PPAdverbialwith ReducedClauseInternalArgument . . . . . . . . . . . . . . . 132

3.42 PPAdverbialSummary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

3.43 ApproximateCountsof TokensandTypesof someADVP Adverbials . . . . . . . 135

3.44 Mis-TaggedPPAdverbials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

3.45 PP-like ADVP Adverbialswith Overt Arguments . . . . . . . . . . . . . . . . . . 136

3.46 PP-like ADVP Adverbialswith HiddenArgument . . . . . . . . . . . . . . . . . . 136

3.47 RelationalADJPwith Overt Argument. . . . . . . . . . . . . . . . . . . . . . . . 137

3.48 RelationalADVP Adverbialswith HiddenArgument . . . . . . . . . . . . . . . . 138

3.49 ApproximateCountsof TokensandTypesof someADVP Adverbials . . . . . . . 139

3.50 ADVP AdverbialConjunctions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

3.51 Mis-TaggedPPAdverbialConstructions. . . . . . . . . . . . . . . . . . . . . . . 142

3.52 Spatio-TemporalADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . 143

3.53 AnotherSpatio-TemporalADVP Adverbial . . . . . . . . . . . . . . . . . . . . . 143

3.54 OtherSpatio-TemporalADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . 143

3.55 Spatio-TemporalMannerADVP Adverbials . . . . . . . . . . . . . . . . . . . . . 144

3.56 Deictic ADVP Adverbials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

xii

Page 13: Thesis about discourse

3.57 Deictic-DerivedADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . . 145

3.58 ApproximateCountsof TokensandTypesof someADVP Adverbials . . . . . . . 146

3.59 Comparative AdverbModifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

3.60 Comparative ADVP Adverbials. . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

3.61 SpecifiedComparative ADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . 147

3.62 Comparative-DerivedADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . 148

3.63 Comparative Constructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

3.64 ApproximateCountsof TokensandTypesof someADVP Adverbials . . . . . . . 150

3.65 OrdinalADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

3.66 Ordinal-ly ADVP Adverbials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

3.67 Frequency ADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

3.68 EpistemicADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

3.69 DomainADVP Adverbials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

3.70 Non-SpecificSet-Evoking ADVP Adverbials . . . . . . . . . . . . . . . . . . . . 153

3.71 Multiply-FeaturedADVP Adverbials. . . . . . . . . . . . . . . . . . . . . . . . . 153

3.72 More Multiply-FeaturedADVP Adverbials . . . . . . . . . . . . . . . . . . . . . 153

3.73 Evaluative or Agent-OrientedADVP Adverbials. . . . . . . . . . . . . . . . . . . 154

3.74 ADVP AdverbialSummary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

4.1 Nine ConnectivesStudiedin [CFM�

02] . . . . . . . . . . . . . . . . . . . . . . . 223

4.2 AnnotationTagsfor theNineConnectivesStudiedin [CFM�

02] . . . . . . . . . . 224

4.3 LOC TagValues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

4.4 Inter-AnnotatorAgreement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

5.1 ADVP/PPAdverbialswith FocusParticleModifier . . . . . . . . . . . . . . . . . 260

5.2 Higher-OrderedEpistemicAdverbialsYielding Implicatures . . . . . . . . . . . . 271

5.3 Lower-OrderedEpistemicAdverbialsYielding Implicatures . . . . . . . . . . . . 273

5.4 Lower-OrderedQuantificationalAdverbialsYielding Implicatures . . . . . . . . . 274

5.5 Higher-OrderedQuantificationalAdverbialsYielding Implicatures . . . . . . . . . 274

xiii

Page 14: Thesis about discourse

List of Figures

2.1 [HH76]’s Typesof Cohesion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2 Illustrationof [GS86]’s DiscourseModel . . . . . . . . . . . . . . . . . . . . . . . 20

2.3 Illustrationof [Web91]’s AttachmentOperation . . . . . . . . . . . . . . . . . . . 24

2.4 Illustrationof [Web91]’s AdjunctionOperation . . . . . . . . . . . . . . . . . . . 24

2.5 LDM Right-AttachmentOperation . . . . . . . . . . . . . . . . . . . . . . . . . . 34

2.6 LDM TreeStructurefor Example(2.58) . . . . . . . . . . . . . . . . . . . . . . . 36

2.7 LDM TreeStructurefor Example(2.59) . . . . . . . . . . . . . . . . . . . . . . . 37

2.8 RSTSchemas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.9 EvidenceRelation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.10 RSTConditionandMotivationRelations . . . . . . . . . . . . . . . . . . . . . . 43

2.11 [KOOM01]’sDiscourseModel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

2.12 [Ash93]’s Classificationof AbstractObjects . . . . . . . . . . . . . . . . . . . . . 52

2.13 SampleDRSs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

2.14 SampleSDRS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

2.15 ElementaryDLTAG Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.1 S-AdjoiningPPandADVP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.2 S-AdjoinedDiscourseandClausalAdverbials . . . . . . . . . . . . . . . . . . . . 84

3.3 S-AdjoinedADVP andPPAdverbialsin PennTreebankI . . . . . . . . . . . . . . 84

3.4 [Ash93]’s Classificationof AbstractObjects . . . . . . . . . . . . . . . . . . . . . 105

3.5 SyntacticStructureof S-ModifyingPPAdverbials. . . . . . . . . . . . . . . . . . 110

xiv

Page 15: Thesis about discourse

3.6 SyntacticStructureof S-ModifyingADVP Adverbials. . . . . . . . . . . . . . . . 134

4.1 ElementaryLTAG Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

4.2 LTAG DerivedTreeafterSubstitions. . . . . . . . . . . . . . . . . . . . . . . . . 160

4.3 LTAG DerivedTreeAfter Adjunction . . . . . . . . . . . . . . . . . . . . . . . . 161

4.4 LTAG DerivationTree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

4.5 SemanticRepresentationsof ��������� and � ������� . . . . . . . . . . . . . . . . . . . 163

4.6 SemanticRepresentationsof Johnwalks . . . . . . . . . . . . . . . . . . . . . . . 163

4.7 SemanticRepresentationsof ��������� , ��������� , ��������� , and� ���� "!#� . . . . . . . . . . . 164

4.8 SemanticRepresentationsof JohnoftenwalksFido . . . . . . . . . . . . . . . . . 164

4.9 SimplifiedSemanticRepresentationof ���$�%�&�� and� ���� "!'� . . . . . . . . . . . . . . 165

4.10 TheElementaryTreefor slide . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

4.11 TheSyntax-SemanticInterfacefor (*),+'-/. . . . . . . . . . . . . . . . . . . . . . . . 168

4.12 DLTAG Initial Treesfor SubordinatingConjunctions . . . . . . . . . . . . . . . . 172

4.13 DLTAG Auxiliary Treefor andand 0 . . . . . . . . . . . . . . . . . . . . . . . . . 173

4.14 DLTAG Auxiliary Treesfor DiscourseAdverbials . . . . . . . . . . . . . . . . . . 174

4.15 DLTAG Initial Treefor AdverbialConstructions. . . . . . . . . . . . . . . . . . . 177

4.16 DLTAG DerivedTreefor Example(4.18) . . . . . . . . . . . . . . . . . . . . . . 178

4.17 DLTAG DerivationTreefor Example(4.18) . . . . . . . . . . . . . . . . . . . . . 178

4.18 Illustrationof [Web91]’s AttachmentandAdjunctionOperations. . . . . . . . . . 179

4.19 Webber’s Adjunctionata Leaf . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

4.20 DerivedTreefor Example(2.41) . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

4.21 Substitutionin FTAG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

4.22 Adjunctionin FTAG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

4.23 LDM ElementaryDCU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

4.24 DTAG ElementaryDCU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

4.25 LDM List Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

4.26 DTAG R Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

4.27 [Gar97b]’s - -Substitution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

xv

Page 16: Thesis about discourse

4.28 [Gar97b]’s - -Adjunction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

4.29 First DTAG Derivationof Example(4.21) . . . . . . . . . . . . . . . . . . . . . . 186

4.30 SecondDTAG Derivationof Example(4.21) . . . . . . . . . . . . . . . . . . . . 186

4.31 StepOnein theSecondDTAG Derivationof Example(4.21) . . . . . . . . . . . . 187

4.32 StepTwo in theSecondDTAG Derivationof Example(4.21) . . . . . . . . . . . . 187

4.33 StepThreein theSecondDTAG Derivationof Example(4.21) . . . . . . . . . . . 188

4.34 StepFour in theSecondDTAG Derivationof Example(4.21) . . . . . . . . . . . 188

4.35 StepFive in theSecondDTAG Derivationof Example(4.21) . . . . . . . . . . . . 189

4.36 StepSix in theSecondDTAG Derivationof Example(4.21) . . . . . . . . . . . . 189

4.37 StepSevenin theSecondDTAG Derivationof Example(4.21) . . . . . . . . . . . 190

4.38 DLTAG ElementaryTreesfor Example(4.22) . . . . . . . . . . . . . . . . . . . . 191

4.39 SemanticRepresentationof ��12!�34��5�'! , ���&�6��!� and �������&�� . . . . . . . . . . . . . . . 192

4.40 DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.22) . . . 192

4.41 DLTAG ElementaryTreesfor Example(4.24) . . . . . . . . . . . . . . . . . . . . 193

4.42 SemanticRepresentationof� ���7� , � '!�! and ���8�"# . . . . . . . . . . . . . . . . . . 193

4.43 DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.24) . . . 194

4.44 DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.26) . . . 194

4.45 DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.28) . . . 195

4.46 LTAG DerivedandDerivationTreesfor Example(4.30) . . . . . . . . . . . . . . . 196

4.47 Quantifiersin French . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

4.48 [Kal02]’s . -Edgesfor Quantifiersin French . . . . . . . . . . . . . . . . . . . . . 197

4.49 [Kal02]’s . -DerivationGraph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

4.50 DLTAG DerivedTreeand . -DerivationGraphfor Example(4.28) . . . . . . . . . 199

4.51 Additional SemanticRepresentationfor (4.28)dueto . -DerivationGraph . . . . . 199

4.52 DLTAG DerivedTreeand . -DerivationGraphfor Example(4.32) . . . . . . . . . 200

4.53 Flexible Compositionin LTAG . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202

4.54 DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.28) . . . 203

4.55 DLTAG ElementaryTreesfor Example(4.33) . . . . . . . . . . . . . . . . . . . . 204

xvi

Page 17: Thesis about discourse

4.56 SemanticRepresentationof� �6� , � "���� , � �$��9 , �;: , � �6�� "!'<�=���!#� and ��34���>4����!#< . . . . . 204

4.57 DLTAG DerivedandDerivationTrees,. -DerivationGraphandSemanticsfor (4.33) 205

4.58 DLTAG ElementaryTreesfor Example(4.35) . . . . . . . . . . . . . . . . . . . . 206

4.59 SemanticRepresentationof� 34���7#!�?#5*!'�� "�@9 , �A: , �����6��� and �B�&��'! . . . . . . . . . . . . 207

4.60 DLTAG DerivedandDerivationTrees,. -DerivationGraph,andSemanticsfor (4.35) 207

4.61 DLTAG ElementaryTreesfor Example(4.37) . . . . . . . . . . . . . . . . . . . . 208

4.62 SemanticRepresentationof� �% , � � , ��<�!�45C�& ED;��� , �� "���� , �;: , �����6��� and �����#! . . . . . 208

4.63 DLTAG DerivedandDerivationTrees,. -DerivationGraphandSemanticsfor (4.37) 209

4.64 DLTAG ElementaryTreeandSemanticRepresentationfor ��<�!�'5C�@ in (4.39) . . . . 210

4.65 DLTAG DerivedandDerivationTrees,. -DerivationGraphandSemanticsfor (4.39) 210

4.66 DLTAG DerivationTreeand . -DerivationGraphfor Example(4.41) . . . . . . . . 212

4.67 ElementaryLTAG TreesandSemanticRepresentationsof� !#=�!'<�9 , � !#=�!'<�9 , ������F , � 12��<��� 214

4.68 ElementaryDLTAG Treesfor Examplefor example . . . . . . . . . . . . . . . . . 215

4.69 DerivationTreesfor PPDiscourseAdverbialswith QuantifiedInternalArguments. 216

4.70 DLTAG DerivedandDerivationTreesfor (4.32) . . . . . . . . . . . . . . . . . . . 218

4.71 AnotherRepresentationof theR Treein Figure4.26 . . . . . . . . . . . . . . . . 219

5.1 GriceanFramework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

xvii

Page 18: Thesis about discourse

Chapter 1

Intr oduction

1.1 The Problem

Traditionally in linguistic theory, syntaxandsemanticsprovide mechanismsto build the interpre-

tationof a sentencefrom its parts;althoughit is non-controversiala sequenceof sentencessuchas

foundin (1.1) - (1.3)alsohasaninterpretation,themechanismswhichproduceit arenotdefined.

(1.1) Thereis a high degreeof stresslevel from the needto competeandsucceedin this ‘me

generation’.As a result,peoplehave becomemoreself-centeredover time.

(1.2)Johnhasfinally beenrewardedfor his greattalent.Specifically, hejust won a gold medal

for mogul-skiingin theOlympics.

(1.3)Thecompany interviewedeveryonewhoappliedfor theposition.In thisway, they consid-

eredall theiroptions.

Most discoursetheoriesgo beyond sentencelevel linguistic theory to explain how suchse-

quencesare put togetherto createa discourseinterpretation. Thesetheoriesevoke the notion

of abstractdiscourserelationsbetweendiscourseunits, provide lists of theserelationsof vary-

ing lengthand organization,andproposediscoursemodelsconstructedfrom theserelationsand

units. Someof thesemodelsproducecompositionalaccountsof discoursestructureand/orinter-

pretation([Pol96, Ash93, MT88, GS86];othersproduceaccountsfor how relationsbetweenunits

areinferred([Keh95, HSAM93, LA93]. Themajority make useof thepresenceof cuephrases, or

1

Page 19: Thesis about discourse

discourseconnectives, treatingthemas“signals” of thepresenceof particulardiscourserelations.

In (1.1),for example,therelevantcuephraseis theadverbialasa result, andthediscourserelationit

signalsis frequentlyclassifiedasa resultrelation. Along with certainadverbials,thesubordinating

andcoordinatingconjunctionsarealsoclassifiedasdiscourseconnectivesin thesetheories.

DLTAG [FMP�

01, CFM�

02, WJSK03, WKJ99,WJSK99,WJ98] is a theorythatbridgesthe

gapbetweenclause-level anddiscourse-level theories,providing amodelof arich intermediatelevel

betweenclausestructureandhigh-level discoursestructure,namely, thesyntaxandsemanticsasso-

ciatedwith discourseconnectives. In DLTAG, discourseconnectivesarepredicates, akin to verbs

at theclauselevel, exceptthatthey takediscourseunitsasarguments.DLTAG proposesto build the

interpretationof thesepredicatesdirectly on topof theclause,usingthesamesyntacticandseman-

tic mechanismsthatarealreadyusedto build theclause.Basedon considerationsof computational

economyandbehavioral evidence,DLTAG arguesthatbothargumentsof subordinatingandcoor-

dinatingconjunctionscanberepresentedstructurally, but only oneargumentof adverbialdiscourse

connectivescomesstructurally;theotherargumentmustberesolvedanaphorically. However, while

DLTAG hasshown thatcertainadverbialsfunctionasdiscourseconnectives,it hasnot isolatedthe

subsetof adverbialswhich functionasdiscourseconnectivesfrom thesetof all adverbials.

Thesetof all adverbialsis a largeset;in fact,it is compositionalandthereforeinfinite[Kno96].

Becauseit is thusnot possibleto list all of the adverbialsthat function asdiscourseconnectives,

in this thesiswe investigatehow semanticsand pragmaticscausean adverbial to function as a

discourseconnective.

1.2 Contrib utions of the Thesis

This thesisextendstheDLTAG model,investigatingthesemanticsandpragmaticsunderlyingthe

behavioral anaphoricityof adverbialdiscourseconnectives. We presenta corpus-basedanalysisof

over13,000S-modifyingadverb(ADVP)andpreposition(PP)adverbialsin thePennTreebankCor-

pus[PT]. Weshow thatcertainadverbials,whichwecall discourseadverbials, canbedistinguished

semanticallyfrom otheradverbials,whichwecall clausaladverbials. Someclausaladverbialsfrom

ourcorpusareshown in (1.4),andsomediscourseadverbialsfrom ourcorpusareshown in (1.5).

2

Page 20: Thesis about discourse

(1.4)Probably/Inmycity/In truth, womentake careof thehouseholdfinances.

(1.5)Asa result/Specifically/Inthisway, womentake careof thehouseholdfinances.

Themostfrequentlyoccurringclausalanddiscourseadverbialshave bothbeenclassifiedin the

literatureasdiscourseconnectives,dueto the fact that they seemto be interpretableonly with re-

spectto context. In this thesiswe will show that while syntaxcannotdistinguishthesetwo types

of adverbials,theirpredicateargumentstructureandinterpretationshows thatonly discourseadver-

bialsfunctionsemanticallyasdiscourseconnectives.

The syntaxandsemanticsof mostdiscourseadverbialshasnot beenwell studied. Generally,

only asmallsubset(thosethatoccurfrequently)havebeenaddressedatall. At theclauselevel these

areusuallydesignatedasthedomainof discourselevel research,andat thediscourselevel thefocus

is frequentlyon thediscourserelationthey “signal”. Our investigationshedslight on thespaceof

relationsimpartedby amuchwider varietyof adverbials.

In our analysiswe draw on clause-level researchinto the semanticsof adverbialsand other

sub-clausalconstituents. We useprior researchon discoursedeixis to study both the semantic

natureof theargumentsof adverbialsandthesyntacticconstituentsfrom whichthey canbederived.

We presenta wide varietyof discourseandclausaladverbials. We show thatdiscourseadverbials

functionsemanticallyasdiscourseconnectivesbecausethey containsemanticargumentsthatmayor

maynotbesyntacticallyovert,but whoseinterpretationrequiresanabstractobjectinterpretationof

acontextual constituent.Weshow thatclausaladverbialsdo not functionsemanticallyasdiscourse

connectivesbecausetheinterpretationsof theirsemanticargumentsdonotrequiretheabstractobject

interpretationof a contextual constituent,althoughthey may make anaphoricreferenceto other

contextual interpretations.Wefurthershow how thepredicateargumentstructureandinterpretation

of discourseadverbialscanbeformalizedandincorporatedinto thesyntaxof theDLTAG model.

It is not only dueto their predicateargumentstructureandinterpretationthat adverbialshave

beenclassifiedasdiscourseconnectives,however. We encounterin our corpusa numberof adver-

bials thathave beentreatedasdiscourseconnectivesdespitethe fact that their semanticsdoesnot

requireabstractobjectinterpretationsin thediscourseor spatio-temporalcontext. Weexploreother

explanationsfor how theseadverbialsevoke discoursecontext duringtheir interpretation;in partic-

3

Page 21: Thesis about discourse

ular, we investigatethe interactionof their semanticswith othersemanticandpragmaticdevices.

We show how focuseffects in S-modifyingadverbialscontribute to discoursecoherence,andwe

alsoshow how S-modifyingadverbialscanbeusedto convey Griceanimplicatures.

While thesemanticsandpragmaticsdiscussedherewill not provide a completeaccountof the

discoursefunctionsof all adverbials, it will show that the analysisof adverbialscan be viewed

modularly: certain functionscan be attributed to the semanticdomain,othersto the pragmatic

domain,andstill othersto largerissuesof discoursestructure.

Therearenumerousbenefitsof thisanalysis.First, it is economical,makinguseof pre-existing

clause-level mechanismsto build adverbial semanticsat the discourselevel, therebyreducingthe

loadon inferenceto accountfor discourseinterpretation(c.f. [Keh95]). Secondly, it providesa the-

oreticalgroundingfor [Kno96]’s empiricalapproachto studyingthelexical semanticsof discourse

connectives, in the processshowing that additionaladverbialsshouldbe includedin the classhe

isolatesbasedon intuition alone,andthatsomeof thoseincludedtheredon’t reallybelong.Thirdly,

it expandsanexistingmodelof discoursewhicharguesthatdiscoursestructurecanbebuilt directly

on top of clausestructureand therebybridgesthe gap betweenhigh-level discoursetheory and

clause-level theory.

1.3 ThesisOutline

In this chapter, we have givena brief overview of theanalysesthatwe presentin theremainderof

this thesis.Therestof this thesisis organizedasfollows:

In Chapter2 we survey a variety of existing discoursetheoriesandexaminethe similarities

anddifferencesbetweeneachtheory. We discusshow, takentogether, eachtheoryservesto distin-

guishdifferentmodulesrequiredto build a completeinterpretationof discourse.We thenoverview

DLTAG asanotherimportantmodulethataloneout of all theothersis capableof bridgingthegap

betweendiscourselevel theoriesandclauselevel theories,by treatingdiscourseconnectivesaspred-

icatesandusingthesamesyntaxandsemanticsthatbuilds theclauseto build anintermediatelevel

of discourse.

In Chapter3 we investigatethe semanticmechanismsthat causesomeadverbialsto function

4

Page 22: Thesis about discourse

asdiscourseconnectives. We discussprior researchinto the semanticsof adverbialsandpresent

an analysisof the S-modifying adverbials in the PennTreebankcorpusthat distinguishesthose

adverbialsthat function asdiscourseconnectives accordingto their predicate-argumentstructure

andinterpretation.

In Chapter4 we show how thesemanticsof adverbialsdiscussedin Chapter3 canbeincorpo-

ratedinto a syntaxsemanticinterfacefor DLTAG. We discusssyntax-semanticinterfacesthathave

beenproposedfor clause-level grammarsandrelateddiscoursegrammarsandshow how thesein-

terfacescanbeextendedto DLTAG. We furtherdiscusstheDLTAG annotationprojectwhosegoal

is to annotatetheargumentsof all discourseconnectives,bothstructuralandanaphoric.

In Chapter5 we continueour analysisof how adverbialsfunction as discourseconnectives.

investigatingotherwaysapartfrom their predicateargumentstructureandargumentresolutionin

whichanadverbialcanbeusedto contribute to discoursecoherence.

Weconcludein Chapter6 anddiscussdirectionsfor futurework.

5

Page 23: Thesis about discourse

Chapter 2

Anaphora and DiscourseModels

2.1 Intr oduction

Discoursemodelsexplain how sequencesof utterancesareput togetherto createa text. Building

a coherent discourseinvolves more than just concatenatingrandomutterances;in addition, the

contributionsof eachutteranceto the surroundingcontext mustbe established.Two major areas

of investigationhave beendistinguished.The first concernshow sub-clausalconstituentsobtain

their meaningthroughrelationshipsto entitiespreviously evoked in a discourse.Suchconstituents

includeNPs, suchasin (2.1) wherethe personalpronounhe refersto an entity mentionedin the

prior sentence,(2.2)wherethebeerrefersto oneof theelementsof thepicnic in theprior sentence,

and(2.3)wherethedemonstrative pronounthat refersto theinterpretationof theprior sentence.

(2.1)Bill talkedto Phillip. He got reallyupset.

(2.2)Bill andMary tookapicnic to thepark.Thebeerwaswarm.

(2.3)Bill talkedto Phillip. Thatmadememad.

OtherexamplesincludeVPs, suchasin (2.4)wheretheelidedVP ( G ) mustbedeterminedfrom

themeaningof theprior sentence,andin (2.5)wheretheuseof simplepasttensein bothsentences

createsanimpressionof forwardprogressionin time.

(2.4)Bill talkedto Phillip. I did G too.

(2.5)Bill entered theroom.He began to talk.

6

Page 24: Thesis about discourse

The secondmajor areaof investigationconcernshow clausal(andsuper-clausal)constituents

obtain their meaningthroughrelationshipsto clausalconstituentsin the surroundingcontext. To

illustratethenatureof theseinvestigations,considerthediscoursein (2.6).

(2.6a)Lastsummer, theKeatingstraveledin Zimbabwe.

(2.6b) Patstudiedflora in theChimanimanimountains.

In the absenceof any additionalcontext, onereadermight interpret(2.6), and/orthe writer’s

intention in producing(2.6), as a descriptionof what the Keatingson the onehand,andPat on

the other, did the prior summer. Another readermight interpret it as contrastingwhat the two

participantsdid theprior summer, e.g.theKeatings(just) traveled,whereasPatstudied.

Interactionsbetweenthesetwo areasof investigationhave also beenstudied. For example,

suppose(2.6) is precededandfollowedby othersentences,asin (2.7).

(2.7a)PatKeatingmarriedMaria Lopezlastspring.

(2.7b) Lastsummer, theKeatingstraveledin Zimbabwe.

(2.7c) Patstudiedflora in theChimanimanimountains.

(2.7d) Thatwasaspectacularcelebration.

Due to additionof (a), the readerwill likely determinethat Pat is a memberof the Keating

family. S/hemight thus interpretPat’s studyingasan elaborationof, or even asa causeof, the

Keatings’traveling, or s/hemight simply interpretPat’s studyingasoccurringafter the Keatings’

traveling. World knowledgeor inferencemayyield thebelief that theChimanimanimountainsare

locatedin Zimbabwe.Notethatthedemonstrative referencein (d) is hardto resolve to themarriage

describedin (a)unlesswe move it to apositionimmediatelyfollowing (a) in thediscourse.

A completemodelof discoursemustaccountfor all of theserelationships,andtheir interactions.

In particular, adiscoursemodelmustcharacterize:H thepropertiesof theconstituentsthatarebeingrelatedH thetypeof relationshipsthatcanexist betweentheseconstituentsH themechanismsunderlyingtheserelationships

7

Page 25: Thesis about discourse

H theconstraintson theapplicationof thesemechanisms

In the following sectionswe will survey a variety of existing discoursemodelsin termsof

their coverageof the above characterizations.By taking a roughly chronologicalapproach,and

examiningthe benefitsandlimitations of eachsubsequentmodel in termsof how it incorporates

thoseprior to it, thesecharacterizationswill befleshedout,andit will beshown that,takentogether,

eachtheoryservesto distinguishdifferentmodulesrequiredto build a completeinterpretationof

discourse.We thenintroduceDLTAG asanimportantmodulecapableof bridgingthegapbetween

discourselevel theoriesandclauselevel theories.

2.2 DescriptiveTheoriesof DiscourseCoherence

2.2.1 An Early EncompassingDescription

[HH76] earlyproposedthata singleunderlyingfactor, which they call cohesion, unifiessequences

of sentencestocreateadiscourse.Cohesionis definedasthe“semanticrelationsbetweensuccessive

linguistic devicesin a text, wherebythe interpretationof onepresupposesthe interpretationof the

otherin thesensethatit cannotbeeffectively decodedexceptby recourseto it”([HH76] p.4)1. They

distinguishfiveclassesof cohesion,shown in Figure2.1.

Figure2.1: [HH76]’s Typesof Cohesion

Referenceis asemanticrelationachievedby theuseof acataphoricor anaphoricreferenceitem

to signalthat theappropriateinstantialmeaningbesupplied.Personalreference(signaledby per-

1This useof the term“presupposition”is not equivalentto semanticpresupposition;the latterdependson truth val-uation and the former doesnot. Both [HH76] and [Sil76] define“discourse”,or “pragmatic”, presuppositionas therelationshipof a linguistic form to its prior context; Silversteinaddsthata pragmaticpresuppositionis whata languageusermustknow aboutthecontext of useof a linguistic signalin orderto interpretit [Sil76, 1]. SeeChapter5 for furtherdiscussionof presupposition.

8

Page 26: Thesis about discourse

sonalpronounsanddeterminers,e.g. I, my), demonstrative reference(signaledby demonstratives,

e.g. this, that), andcomparative reference(signaledby certainnominalmodifiers,e.g. same, and

verbaladjuncts,e.g. identically) aredistinguished,andexemplifiedin italics in (2.8).

(2.8)Johnsaw ablackcat,but that doesn’t meanit wasthesameblackcathesaw before.

Lexical cohesionis a semanticrelation achieved by the successive useof vocabulary items

referringto thesameentity or event, includingdefinitedescriptions,repetitions,synonyms,super-

ordinates,generalnouns,andcollocation.Everylexical itemcanbelexically cohesive; this function

is establishedby referenceto the text. In [HH76]’s example,shown in (2.9), thereare definite

descriptions:a pie...thepie, repetitions:pie...pie, generalnounsandsynonyms: a pie...adaintydish

andsuper-ordinates:blackbirds...birds.

(2.9)

Singasongof sixpence,apocket full of rye,

Four-and-twentyblackbirdsbakedin apie,

Whenthepie wasopened,thebird beganto sing,

Wasn’t thatadaintydishto setbeforeaking?

Substitution andEllipsis aregrammaticalrelations,which canbenominal,verbal,or clausal.

Thesubstitutemustbeof thesamegrammaticalclassastheitemfor whichit substitutes,andellipsis

is substitutionby zero([HH76, 89]). In (2.10),nominaloneis a substitute,andthereis ellipsisof

theembeddedpredicatein thefinal clause.

(2.10)Mary covetstwo things. Her money will be thefirst oneto leave her. Her husbandwill

bethenext 0.

Conjunction is asemanticrelationusuallyachievedby theuseof conjunctive elements,whose

meaningpresupposesthepresenceof otherpropositionsin thediscourseandspecifiestheway they

connectto the propositionthat follows. Italicized examplesareshown in (2.11). [HH76] distin-

guishfour main typesof relationsbetweenpropositions,shown in Table2.1. Theserelationsare

furthersubdivided,andanorthogonaldistinctionis madebetweenexternal and internal relations;

the formerhold betweenelementsin theworld (referredto in thetext), andthe latterbetweentext

elementsthemselves,suchasspeechacts.

9

Page 27: Thesis about discourse

(2.11)Becauseit snowedheavily, thebattlewasnot fought,so thesoldierswenthome.

Table2.1: Main Categoriesof [HH76]’sRelationsbetweenPropositions

ADDITIVE ADVERSATIVE CAUSAL TEMPORALcomplex contrastive specific sequentialapposition correction conditional simultaneouscomparison dismissal respective conclusive

correlative

2.2.2 Alter nativeDescriptionsof PropositionalRelations

In comparisonto [HH76], [Lon83]’s studyof discoursecoherencedistinguishesbetweenpredica-

tions expressedby clauses,which he modelswith predicatecalculus,and relationson the pred-

icationsexpressedby clauses,which he characterizesinto two main types,shown in Table 2.2:

the “basic” operationsof propositionalcalculus,supplementedby temporalrelations,anda setof

elaborative relations.Theserelationsarefurthersubdivided,andanorthogonaldistinctionis made

betweennon-frustratedandfrustratedrelations,thelatterbeingthecasewhenanexpectedrelation

is not satisfiedby theassertionsin thetext. Unlike [HH76], [Lon83] doesnot emphasizea correla-

tion betweentheserelationsandsurfacesignalsin thetext; rather, they aremeantto categorizethe

“deep” relationsunderlyingthesurfacestructureof discourse.

Table2.2: Main Categoriesof [Lon83]’s RelationsbetweenPropositions

BASIC ELABORATIVEconjoining( I ) paraphrasealternation( J ) illustrationimplication( K ) deixistemporal attribution

Morerecently, [Mar92] hasproposedanalternativesetof relationsbetweenpropositions,shown

in Table2.3,in whichfour maintypesaredistinguished.Theserelationsarefurthersubdivided,and

orthogonaldistinctionsaremadebetweeninternalandexternalrelations,andparatactic,hypotactic,

andneutral relations.Thefirst dimensionis taken from [HH76], andthe latterdimensionroughly

10

Page 28: Thesis about discourse

correspondsto coordinating,subordinating,andvariablycoordinatingandsubordinatingrelations,

respectively. Like [HH76], Martin usesexplicit signalsto derive his set of discourserelations,

but like [Lon83], he defendstheclaim that they represent“deep” relationsunderlyingthesurface

structure.He combinesthe two approachesby usingan insertiontest: a “deep” relationexists at

a placein thetext if anexplicit signalcanbe insertedthere.Nevertheless,his setis differentfrom

both[HH76] and[Lon83].

Table2.3: Main Categoriesof [Mar92]’sRelationsbetweenPropositions

ADDITIVE COMPARATIVE TEMPORAL CONSEQUENTIALaddition similarity simultaneous purpose concessionalternation contrast successive condition manner

consequence

[SSN93] take a psychologicalapproach,identifying the basiccognitive resourcesunderlying

theproductionof discourserelations.Four cognitive primitivesareidentified,accordingto which

discourserelationscanbeclassified,which they exemplify usingexplicit cuephrases.[SSN93]cite

anumberof psychologicalexperimentsto supportthesefeatures.H basicoperation: Eachrelationcreateseitheranadditive (and) or acausal(because) connec-

tion betweentherelatedconstituents.H source of coherence: Eachrelationcreateseithersemanticor pragmaticcoherence;in the

first casethepropositionalcontentof theconstituentsis related,in thesecondcasetheillocu-

tionaryforceof theconstituentsis related.H order of segments: Causalrelationsmayhave thecausingsegmentto theleft or theright of

theresult.H polarity : A relationis negative if it links thecontentof onesegmentto thenegationof the

contentof theothersegment(although), andpositive otherwise.

[Hob90] takesa computationalapproach,identifying relationsbetweenpropositionsaccording

to thekind of inferencethatis requiredto identify them.Themaincategoriesareshown in Table2.4.

11

Page 29: Thesis about discourse

Respectively, thesecategoriesdistinguishinferenceaboutcausalitybetweeneventsin the world,

inferenceaboutthespeaker’s goals,inferenceaboutwhat theheareralreadyknows, andinference

thata heareris expectedto be ableto make aboutrelationshipsbetweenobjectsandpredicatesin

theworld. [Hob90] suggeststhat inferenceshouldbeviewedasa recursive mechanism;whentwo

propositionsare linked by a relation, they form a unit which itself canbe relatedto otherunits,

therebybuilding aninterpretationof thediscourseasawhole.

Table2.4: Main Categoriesof [Hob90]’s RelationsbetweenPropositions

Occasion Evaluation Ground-Figure Explanationcause background parallel generalizationenablement explanation exemplification contrast

2.2.3 DiscourseRelationsasConstraints

[Keh95] reformulates[Hob90]’s relationsbetweenpropositionsinto threemaintypesof moregen-

eral “discourserelations”: Contiguity, Cause-Effect, andResemblance,which hedefinesin terms

of constraintsonbothclausalandsub-clausalpropertiesof discourseunitsSL andSM . Hethenspec-

ifies how an inferencemechanismcanbeusedto derive Cause-Effect andResemblancerelations,

andshows how they interactwith sub-clausalcoherence.Like [HH76] and[Mar92], hecorrelates

theserelationswith the presenceof cuephrases,suggestingthat they could be treatedasbearing

semanticfeaturesthatinteractwith thediscourseinferenceprocess.

Narration is theonly ContiguityrelationKehlerdefines.Exemplifiedin (2.12),theconstraint

on its derivation is that a changeof statefor a systemof entitiesfrom SM be inferred,wherethe

initial statefor this systemis inferredfrom SL .(2.12)Bill pickedup thespeech.Hebeganto read.

Kehlernotesthat the full setof constraintsgoverning the recognitionof a Narrationrelation

arenotwell understood,but herefutes([HH76, Lon83])’s treatments,whichequateit with temporal

progression,citing [Hob90]’sexample(2.13),whoseinterpretationrequirestheadditionalinference

thatBushis on thetrain,or thatthetrainarrival is somehow relevantto him.

12

Page 30: Thesis about discourse

(2.13)At 5:00a trainarrivedin Chicago.At 6:00GeorgeBushheldapressconference.

Kehlerdistinguishesfour typesof Cause-Effectrelations,all of which mustsatisfy the con-

straintthat a presupposedpathof implicationbe inferredbetweena propositionP from SL , anda

propositionQ from SM . Eachtypeandtheimplicationit requiresis shown in Table2.5,alongwith

correlatedcuephrases.

Table2.5: [Keh95]’s Cause-Effect Relations

Relation Presupposition ConjunctionsResult P K Q asa result,therefore,and

Explanation Q K P becauseViolatedExpectation P KON Q butDenialof Preventer Q KON P despite,eventhough

To take two examples,aResult relationis inferredwhenQ is recognizedasnormallyfollowing

from P. In (2.14),beinga politician normallyimpliesbeingdishonest.

(2.14)Bill is apolitician,andthereforehe’s dishonest.

Denial of Preventer relationsareinferredwhen N P is recognizedasnormally following from

Q (example(2.15)).

(2.15)Bill is honesteventhoughhe’s apolitician.

Kehlerdistinguishessix typesof Resemblancerelations,all having theconstraintthata com-

monor contrastingrelationP beinferredbetweenSL andSM , suchthat P subsumespL andp M , where

pL appliesover a setof entitiesaM ,...a� from SL , andp M appliesover a setof entitiesb M ,...b� from

SM . CertainResemblancerelationsalsohave theconstraintthatapropertyvector Q beinferred,such

that Q consistsof commonor contrastingpropertiesq� , which hold for a� andb� , for all + 2. Table

2.6providestheconstraintsfor eachResemblancerelationandits correlatedcuephrase.

For example,Exemplification holdsbetweena generalstatementfollowed by an exampleof

the generalization.In (2.16),aM andb M correspondto themeaningsof youngaspiring politicians

2KehlernotesthatElaborationrelationsarea limiting caseof Parallel relations,wherethesimilar entitiesaR andbRareidentical.

13

Page 31: Thesis about discourse

and SUT7VXW ; while pL andp M correspondto themeaningsof supportandcampaignfor respectively3.

Generalization is identicalto Exemplification,exceptthattheorderof theclausesis reversed.

(2.16)Youngaspiringpoliticiansoftensupporttheirparty’spresidentialcandidate.For instance,

Johncampaignedhardfor Clinton in 1992.

Table2.6: [Keh95]’sResemblanceRelations

Relation Constraints ConjunctionsElaboration pL = p M , a� = b� in otherwords

Exemplification pL = p M , b�UY a� or b�UZ a� for exampleGeneralization pL = p M , a� Y b� or a� Z b� in general

Parallel pL = p M , q� (a� ) andq� (b� ) andContrast(i) pL = N p M , q� (a� ) andq� (b� ) butContrast(ii) pL = p M , q� (a� ) and N q� (b� ) but

Parallel relationsrequiretherelationsexpressedby thesentenceandthecorrespondingentities

to be recognizedassharinga commonproperty. In (2.17),pL andp M correspondto the meanings

of organizedrallies for anddistributedpamphletsfor respectively; P correspondsto the meaning

of do somethingto support. aM andbM correspondto the meaningsof John andBill , which share

the commonpropertyq M that they arepeoplerelevant to the conversation.Contrast relationsre-

quireeithertherelationsexpressedby thesentences(example(2.18))or thecorrespondingentities

(example(2.19))to berecognizedascontrasting.

(2.17)Johnorganizedralliesfor Clinton,andFreddistributedpamphletsfor him.

(2.18)JohnsupportedClinton,but Mary opposedhim.

(2.19)JohnsupportedClinton,but Mary supportedBush.

2.2.4 Abducing DiscourseRelationsby Applying the Constraints

Kehler’s constraintsareformulatedin termsof two operationsfrom artificial intelligence:(1) iden-

tifying commonancestorsof setsof objectswith respectto a semantichierarchy(Resemblance

3Although not discussedby Kehler, the subsumingproperty[ is equatablewith p\ , andp ] canbe recognizedasamemberof p\ .

14

Page 32: Thesis about discourse

relations),and(2) computingimplication relationshipswith respectto a knowledgebase(Cause-

Effectsrelations).Kehlerdistinguishestwo stepsin thediscourseinferenceprocess:

(a) Identify andretrieve theargumentsto thediscourserelation

This stepis achieved via the sentenceinterpretation;Kehlerusesa formalism relatedto the

versionof Categorial Semanticsdescribedin [Per90], in which sentenceinterpretationresultsin a

syntacticstructureannotatedwith thesemanticrepresentationof eachconstituent.Thesesemantic

representationsareargumentsto thediscourserelation,andareidentifiedandretrievedvia theircor-

respondingsyntacticnodes.Cause-Effect relationsrequireonly theidentificationof thesentential-

level semanticsfor the clausesasa whole (i.e. P andQ). Resemblancerelationsrequirethat the

semanticsof sub-clausalconstituentsbeaccessed,in orderto identify pL andp M , anda� andb� .(b) Apply theconstraintsof therelationto thosearguments

Thesecondstep,Kehlersuggests,couldbeachievedfor Resemblancerelationsusingcompari-

sonandgeneralizationoperationssuchasproposedin [Hob90] andelsewhere,while [HSAM93]’s

logical abductioninterpretationmethodcouldbeusedto abducethepresuppositionfor theCause-

Effect relations.[HSAM93]’s methodcouldfurtherdeterminewith whatdegreeof plausibility the

constraintsaresatisfiedsuchthataparticularrelationholds.

In [HSAM93]’s framework, discourserelationsbetweendiscourseunits areproved (abduced)

usingworld anddomainknowledge,via aprocedureof axiomapplication.Eachdiscourseunit is a(�.�^`_a.�Wcb , asdefinedby axiom(2.20),whereif ( is a sentencecontaininga stringof words, d , and. is its assertionor topic, thenit is adiscoursesegment.

(2.20)( e w, e)s(w, e) f Segment(w, e)

Whenadiscourserelationholdsbetweentwo segments,theresultingstructureis alsoasegment,

yielding a hierarchicaldiscoursestructure,ascapturedby axiom (2.21),whereif w M andw g are

segmentswhoseassertionor topic arerespectively eM andeg , anda discourse(coherence)relation

holdsbetweenthecontentof w M andw g , thenthestringw M w g is alsoasegment.Theargument. of

CoherenceRelis theassertionor topicof thecomposedsegment,asdeterminedby thedefinitionof

thediscourserelation.

15

Page 33: Thesis about discourse

(2.21) (e w M , w g , eM , eg , e) Segment(wM , eM ) J Segment(wg , eg ) J CoherenceRel(eM , eg , e) fSegment(wM w g , e)

To interpretadiscourseW, therefore,onemustprove theexpression:

(2.22)( h e)Segment(W, e)

Weuseasanexampleavariantof thatfoundin [Keh95]:

(2.23)Johnis dishonest.He’sapolitician.

To interpretthis discourse,it mustbe proven a segment,by establishingthe threepremisesin

axiom (2.21). The first two premisesareestablishedby (2.20), it thereforeremainsto establish

a discourserelation. BecauseExplanationis a defineddiscourserelation,we have the following

axiom:

(2.24)( e eM , eg )Explanation(eM , eg ) f CoherenceRel(eM , eg , eM )In explanations,Hobbsnotes,it is thefirst segmentthatis explained;thereforeit is thedominant

segmentandits assertion,eM , will betheassertionof thecomposedsegment,i.e. thethird argument

of CoherenceRelin (2.24).

Recall that the constraintsdefinedby Kehler on Explanationrelationswere that Q K P be

presupposed;in Hobbs’ terms,thepresuppositioni8j/k$(�. (eM , eg ) mustbeabduced,asexpressedby

thefollowing axiom:

(e eM , eg )cause(eM , eg ) f Explanation(eM , eg )In otherwords,to abduceanExplanationrelation,whatis assertedby eg mustprovento bethe

causeof eM . In [HSAM93], utterances,like discourserelations,areinterpretedby abducingtheir

logical form, usingaxiomsthat arealreadyin the knowledgebase,arederivable from axiomsin

the knowledgebase,or canbe assumedat a costcorrespondingto somemeasureof plausibility.

Assumewehave abducedthefollowing axiom:

(2.25)( e x, eM )Politician(eM , x) f ( h eg )Dishonest(eg , x) J cause(eM , eg )Thatis, if eM isastateof x beingapolitician,thenthatwill causethestateeg of x beingdishonest.

Theplausibility measurethat is assignedto this formulawill beinverselyproportionateto thecost

16

Page 34: Thesis about discourse

assignedto anExplanationrelation.Assuming(2.25)hasahighplausibility in ourknowledgebase,

then in the logical forms of the two sentencesin (2.23), John (andhe) canbe identifiedwith x,

cause(eM , eg ) proventhereby, andExplanationwill beviewedasthelikely relationbetweenthetwo

sentences4.

2.2.5 Interaction of DiscourseInferenceand VP Ellipsis

[Keh95] shows how thediscourseinferenceprocessfor Resemblancerelationsinteractsdifferently

thanthediscourseinferenceprocessfor Cause-Effect relationswith VP ellipsis,basedon thedif-

ferentconstraintsthey requireto besatisfiedby theclausesthey areinferredbetween.In particular,

the argumentsto Resemblancerelationsaresetsof parallelentitiesandrelations. Therefore,the

discourseinferenceprocessmustaccesssub-clausalconstituentsin identifyingandretrieving those

arguments,including the missingconstituentin VP ellipsis. In contrast,the argumentsto Cause-

Effect relationsarepropositions.Thereforetheinferenceprocessneednot accesssub-clausalcon-

stituents.Thisdifferenceaccountsfor differentfelicity judgmentsconcerningVP ellipsisdisplayed

acrossthetwo typesof relations.

To exemplify his analysis,considerexample(2.26),in which a Parallelrelationcanbeinferred

betweenthetwo clauses:

(2.26)Bill becameupset,andHillary did too.

ToestablishaParallelrelation(seetheResemblanceRelationdefinitionsin Table2.6),p(aM ,ag ...)mustbeinferredfrom Sl , andp(bM ,bg ...) mustbeinferredfrom Sm , wherefor somepropertyvectorQ , q� (a� ) andq� (b� ), for all + . Theidentificationof theseargumentsrequirestheelidedmaterialto be

recovered j`WU- reconstructedin theelidedVP (see[Keh95] for detailsof theprocessof reconstruc-

tion). Compare(2.26),however, with (2.27).

(2.27)*The problemwaslookedinto by Billy, andHillary did too.

Again,to establishaParallelrelationbetweenthetwo clauses,theargumentsmustbeidentified,

requiringtheelidedmaterialto berecoveredandreconstructedin theelidedVP. But in thiscasethe

4This explanationis from [Keh95, 18] and[Lag98]. See[HSAM93] for furtherdetails

17

Page 35: Thesis about discourse

recovery of waslooked into createsa mismatchof syntacticform whenit is reconstructedin the

elidedVP, resultingin aninfelicitousdiscourse.

Suchinfelicity doesnot occur, however, when thereis a Cause-Effect relation betweentwo

similar clauses,asin example(2.28):

(2.28)Theproblemwasto have beenlookedinto, but obviously nobodydid.

KehlerarguesthatbecauseestablishingaViolationof Expectationrelation(seetheCause-Effect

Relationdefinitionsin Table2.5)requiresonly thatapropositionPbeinferredfrom Sl , andapropo-

sition Q be inferredfrom Sm (wherenormallyP KnN Q), theelidedVP neednot bereconstructed

in thesyntax,but canberecoveredthroughanaphoraresolution.Theresultis that thediscourseis

felicitous5.

2.2.6 Summary

In this section,we have seentheearlydelineationof differenttypesof coherenceproposedby Hal-

liday andHasanreflectedin subsequenttheoriesof discoursecoherence,which we will seefurther

below. Thecomparisonof thesetof propositionalrelationsproposedby Halliday andHasanwith

thoseproposedin otherdescriptive theorieshighlightsthelack of agreementin theliteratureabout

how animportantaspectof discoursecoherenceshouldbedescribed.As we will continueto seein

the following sections,thoughmostmodelsmake useof explicit signalsto characterizediscourse

relations,therestill existsconsiderablevariationin thenumberandtypeof discourserelationseach

modeldefines.Whatdistinguisheseachmodelis thedegreeandmannerwith which they associate

theirpostulatedsetof discourserelationsto mechanismsthatproducethemandhow they constraint

theapplicationof thesemechanisms.Kehler’s attemptto definerelationsbetweendiscourseunits

in termsof constraintswhich thoseunitsmustsatisfy, to demonstratehow their satisfactioncanbe

determinedusingthe logical abductionmethodof Hobbset al., andto show how this satisfaction

interactswith sub-clausalcoherence,is a first exemplificationof suchanassociation.We will see

othersbelow, andin thefinal sectionwe will seea way in which thesevariousapproachescanbe

5Kehlerdoesnotaddressthefactthat(2.26)is infelicitouswith a Cause-Effect relation,e.g.Theproblemwaslookedinto byBilly, but Hillary didn’t.

18

Page 36: Thesis about discourse

simplified.

2.3 A Thr ee-Tiered Model of Discourse

2.3.1 The ThreeTiers

[GS86]presenta theoryof discoursethatdistinguishesthreeinteractingcomponents:thelinguistic

structure,theintentionalstructure,andtheattentionalstate.

The linguistic structure representsthestructureof sequencesof utterances,i.e. thestructure

of segmentsinto which utterancesaggregate. This structureis not strictly compositional,because

a segmentmay consistof embeddedsubsegmentsaswell asutterancesnot in thosesubsegments.

This structureis viewedasakin to thesyntacticstructureof individual sentences([GS86],footnote

1), althoughtheboundariesof discoursesegmentsareharderto distinguish6.

The intentional structure representsthe structureof purposes,(DSPs),i.e. the functionsof

eachdiscoursesegment,whosefulfillment leadsto the fulfillment of anoverall discoursepurpose

(DP).DPsandDSPsaredistinguishedfrom otherintentionsby thefactthatthey areintendedto be

recognized.Non-DP/DSPintentions,suchasa speaker’s intentionto usecertainwords,or impress

or teachthehearer, areprivate, i.e. not intendedto contributeto discourseinterpretation.Examples

of DPsandDSPsinclude intendingthe hearer to performsomeaction, intendingthe hearer to

believe somefact, intendingthe hearer identify someobject or property of an object. As these

examplesimply, thesetof intentionsthatcanserveasDSPsandDPsis infinite, althoughit remains

anopenquestionof whetherthereis afinite descriptionof thisset.However, [GS86] arguethatthere

areonly two structuralrelationswhich canhold betweenDSPsandtheir correspondingdiscourse

segments.If thefulfillment of a DSPA providespartial fulfillment of a DSPB, thenB dominates

A. If a DSPA mustbefulfilled beforea DSPB, thenA satisfaction-precedesB. Becausea hearer

cannotknow the whole setof intentionsthat might serve asDSPs,what they recognize,[GS86]

argue,is therelevantstructuralrelationsbetweenthem.

Theattentional state is viewedasa componentof thecognitive state, which alsoincludesthe

6See[GS86,FM02] for referencesto studiesinvestigatingtheseboundaries.

19

Page 37: Thesis about discourse

knowledge,beliefs,desiresandprivateintentionsof thespeaker andhearers.Theattentionalstate

is inherentlydynamic,andis modeledby a stackof focusspaces, eachconsistingof the objects,

propertiesand relationsthat are salientin eachDSP, as well as the DSP itself. Changesin the

attentionalstatearisethroughtherecognitionof thestructuralrelationsbetweenDSPs.In general,

whentheDSPfor a new discoursesegmentcontributesto theDSPfor the immediatelypreceding

segment,it will bepushedontothestack;whenthenew DSPcontributesto someintentionhigher

in the dominancehierarchy, several focusspacesarepoppedfrom the stackbeforethe new one

is pushed. Onerole of the stackis to constrainthe possibleDSPsconsideredas candidatesfor

structuralrelationswith theincomingDSP;only DSPsin thestackandin oneof thetwo structural

relationsareavailable. Another role of the stackis to constrainthe hearer’s searchfor possible

referentsof referringexpressionsin anincomingutterance;thefocusspacecontainingtheutterance

will provide themostsalientreferents.Figure2.2 illustratesthemajoraspectsof themodel.

Figure2.2: Illustrationof [GS86]’sDiscourseModel

In the left of thefigure,a sequenceof five utterancesis divided into DSs,whereDS1includes

bothDS2andDS3,aswell asUtterance1andUtterance5,which arenot includedin eitherDS2or

DS3.As shown in (a),thefocusspaceFS1containingDSP1andtheobjects,propertiesandrelations

so far identifiedin DS1 is pushedon the stack. BecauseDSP1is identifiedasdominatingDSP2,

FS2 is alsopushedonto the stack. In (b), DSP2is identifiedasbeing in a satisfaction-precedes

relationshipwith DSP3;FS2is thuspoppedfrom thestackbeforeFS3is pushedontothestack.

[GS86] argue that the hearermakesuseof threepiecesof informationwhendeterminingthe

20

Page 38: Thesis about discourse

segments,their DSPs,andthestructuralrelationshipsbetweenthem. First, linguistic expressions,

including cue phrasesand referring expressionsas well as intonationand changesin tenseand

aspect,areviewed as primary indicatorsof discoursestructure,even as the attentionalstructure

constrainstheir interpretations.[GS86]arguethatwhile linguisticexpressionscannotindicatewhat

intentionis enteringinto focus,they canprovide partial informationaboutchangesin attentional

states,whetherthischangereturnsto apreviousfocusspaceor createsanew one,how theintention

in thecontainingdiscoursesegmentis relatedto otherintentions,andstructuralrelationsbetween

segments.They exemplify suchusesof linguisticexpressionasshown in Table2.7.

Table2.7: [GS86]Changesin DiscourseStructureIndicatedby LinguisticExpressions

AttentionalChange (push) now, next, thatremindsme,and,but(popto) anyway, but anyway, in any case,now backto(complete) theend,ok, fine,paragraphbreak

TrueInterruption I mustinterrupt,excusemeFlashbacks Oops,I forgotDigressions By theway, incidentally, speakingof

Did youhearabout...,thatremindsmeSatisfaction-precedes in thefirst place,first, second,finally

moreover, furthermoreNew Dominance for example,to wit, first, second,and

moreover, furthermore,therefore,finally

Second,thehearermakesuseof theutterance-level intentionsof eachutterance[Gri89] to de-

terminethe DSPof eachdiscoursesegment. The DSPmay be identical to someutterance-level

intentionin a segment,asin a rhetoricalquestion,whoseintentionis to causethehearerto believe

thepropositionconveyed in thequestion.Alternatively, theDSPmaybesomecombinationof the

utterance-level intentions,asin asetof instructions,wheretheintentionof thespeaker is thatall of

thembecompleted.

Third, sharedknowledgebetweenthe speaker andheareraboutthe objectsandactionsin the

stackcan help determinethe structuralrelationsbetweenutterancesand the intentionsunderly-

ing them. [GS86] proposetwo relationshipsconcerningobjectsandactionsthata heareruses.A

supports relationholding betweenpropositionsmay indicatedominancein onedirection,while

21

Page 39: Thesis about discourse

a generatesrelationholding betweenpropositionsmay indicatedominancein anotherdirection.

They leaveasanopenquestionhow theserelationsbetweenobjectsarecomputed,but view themas

morebasicversionsof thepossiblerelationsbetweenpropositionsproposedby [HH76] andothers.

Together, this informationenableahearerto reasonout theDSPsandDP in adiscourse.

2.3.2 Coherencewithin DiscourseSegments

Within eachdiscoursesegment,CenteringTheory(CT) [WJP81] is amodelof sub-clausaldiscourse

coherencewhichtracksto themovementof entitiesthrougheachfocusstateby oneof four possible

focusshifts. In CT, eachdiscoursesegmentconsistsof utterancesdesignatedasU � . Eachutterance

U � evokesa (�.�b of discourseentities,the forward-looking centers,Cb(U � ). The highest-ranked

entity in Cf(U �,D M ) that is op.�jq)r+#s`.�- in U � is thebackward-looking center, Cb. Thehighest-ranked

entity in Cf(U � ) is thepreferred center, Cp. Therealizerelationis definedin [WJP81] asfollows:

As utteranceU realizesa center i if i is an elementof the situationdescribedby U, or i is the

semanticinterpretationof somesubpartof U.

Rankingof themembersof theCf list is language-specific;in Englishtherankingis asfollows:

Subjectt IndirectObject t DirectObject t Other

Four typesof transitionsaredefinedto reflectvariationsin the degreeof topic continuity and

arecomputedaccordingto Table2.8:

Table2.8: CenteringTheoryTransitions

Cb(U� ) = Cb(U�,D M ) Cb(U� ) uv Cb(U�,D M )Cb(U� ) = Cp(U� ) Continue Smooth-ShiftCb(U� ) uv Cp(U� ) Retain Rough-Shift

Discoursecoherenceis thencomputedaccordingto thefollowing transitionorderingrule: Con-

tinueis preferredto Retain,which is preferredto Smooth-Shift,which is preferredto RoughShift.

CT modelsdiscourseprocessingfactorsthatexplain thedifferencein theperceived coherence

of discoursessuchas(2.29)and(2.30).

22

Page 40: Thesis about discourse

(2.29a)Jeff helpedDick washthecar.

(2.29b)He washedthewindows asDick waxedthecar.

(2.29c)He soapedapane.

(2.30a)Jeff helpedDick washthecar.

(2.30b)He washedthewindows asDick waxedthecar.

(2.30c)He buffed thehood.

CT predictsthat (2.30) is harderto processthan(2.29), becausethoughinitially in both dis-

coursestheentity realizedby Jeff is establishedastheCb,utterance(2.30c)causesa Smooth-Shift

in whichtheCbbecomestheentityrealizedby Dick, becausetheverbbuffing is asubsetof thewax-

ing event.Thepredictedpreferencefor aContinue(whichactuallyoccursin (2.29c))meansthatthe

hearerfirst interpretsthepronounshein (2.30c)astheCp(U�,D M ) andthenrevisesthis interpretation.

2.3.3 Modeling Linguistic Structure and Attentional Stateasa Tree

[Web91] arguesthata treestructureandinsertionalgorithmcanserve asa formal analogueof both

on-linerecognitionof discoursestructureandchangesin attentionstate,therebyremoving theneed

to postulateaseparatestackfor focusspaces,while retainingthedistinctionbetweentext structure,

intentionalstructure,andattentionalstate.

Webber’s modelassumesa one-to-onemappingbetweendiscoursesegmentsand treenodes,

with a clauseconstitutingtheminimal unit. In this way thelinguistic structureis representedcom-

positionally. Eachnodein thetreeis associatedwith theentities,propertiesandrelationsconveyed

by thediscoursesegmentit represents.Whentheinformationin anew clauseC is to beincorporated

into an existing discoursesegmentDS, C is incorporatedinto the treeby the operationof attach-

ment, which addstheC nodeasa child of theDS node,andaddstheinformationconveyedby C to

theDSnode.Thisoperationis illustratedin Figure2.3. (a)shows thetreebeforenode3 is attached,

while (b) shows thetreeafternode3 is attached.Notethat theinformationassociatedwith node3

is representedin node3 and incorporatedinto thediscoursesegment(1,2,3)it hasattachedto.

Whenthe informationin a new clauseC is combinedwith the informationin an existing dis-

coursesegmentto composea new discoursesegmentDS, C is incorporatedinto the treeby the

23

Page 41: Thesis about discourse

Figure2.3: Illustrationof [Web91]’s AttachmentOperation

operationof adjunction, which makesC andDS thechildrenof a new node,andaddstheinforma-

tion conveyedby C andDS to thenew node.This operationis illustratedin Figure2.4. (a) shows

thetreebeforenode3 is adjoined,while (b) shows thetreeafternode3 is adjoined.Note that the

informationassociatedwith node3 is incorporatedalongwith theinformationassociatedwith node

(1,2) (whichwasalsocreatedby adjunction)into thenew node(1,2),3).

Figure2.4: Illustrationof [Web91]’s AdjunctionOperation

Bothof theseoperationsarerestrictedto applyingto nodesontheright frontierof thediscourse

tree.Formally, theright frontier is thesmallestsetof nodescontainingtheroot suchthatwhenever

anodeis in theright frontier, sois its rightmostchild. In thisway, thetreenodesappearin thesame

linearorderasthecorrespondingsegmentsin thetext.

In Webber’s model, the treereplaces[GS86]’s linguistic structure,andthe right fr ontier re-

places[GS86]’s attentionalstate,i.e. the informationin eachnodeon theright frontier represents

the informationin eachfocusspacein thestack. Becausethemodelis strictly compositional,not

24

Page 42: Thesis about discourse

all nodes(discoursesegments)in thetreewill containdiscoursesegmentpurposes(DSPs)(e.g.Ut-

terance1andUtterance5in Figure2.2);however, all nodeson theright frontier exceptpossiblythe

leafwill containDSPsthatcontributeto theDPof theoverall discourse(whichwill becontainedin

therootof thetree.)

2.3.4 Intr oduction to DiscourseDeictic Reference

[Lak74] first usedtheterm“discoursedeixis” to referto usesof thedemonstrative likethosein (2.31)

- (2.33),wheretheantecedentof thedemonstrative canbethe interpretationof a verbalpredicates

(2.31),theinterpretationof aclause(2.32),or theinterpretationsof morethanoneclause(2.33).

(2.31)John[smiled]. He doesthat often.

(2.32)[Johntook Biology 101.] Thatmeanshecantake Biology 102.

(2.33)[I wokeupandbrushedmy teeth.I wentdownstairsandatebreakfast,andthenI wentto

work.] That’s all I did today.

Earlystudiesof thisphenomenarelateit to anotheruseof demonstrativesshown in (2.34),where

theantecedentis not in thediscourseat all, but ratherin thespatio-temporalsituation.This useis

called“deictic”, aGreektermmeaning‘pointing’ or ‘indicating’.

(2.34)“Aw, that’s nice,Billy!”, youexclaim,whenyour two-yearold kissesyou.

In [Lyo77]’s view, discoursedeixisachieveshigher-orderreference,wherefirst-orderreference

is definedas referenceto NPs, and higher-order referenceis definedas referenceto larger con-

stituentsinterpretedas events,propositionsand concepts. [Web91] distinguishesfive discourse

deixis interpretations,shown in Table2.9, andexemplifiedin thesecondcolumn,wherefor illus-

trative purposesthediscoursedeicticshouldbeassumedto refer to an interpretationof theclause

“Johntalksloudly”.

Demonstrativesaremostcommonlyemployedin Englishfor discoursedeixispurposes.Corpus

studies,however, have shown thezero-pronounusedin Italian [DiE89] andGerman[Eck98], and

occasionallyin Englishspeech.

[Sch85] studiesroughly 2000 tokensof it and that, andfinds that it is much lessfrequently

25

Page 43: Thesis about discourse

Table2.9: [Web91]’s Classificationof DiscourseDeictic Reference

Inter pretation Examplehlinespeechact that’s a lieproposition that’s trueevent that happenedyesterdaypuretextual repeatthatdescription that’s a gooddescription

usedthanthat asa discoursedeictic,andthatwhenusesof discoursedeictic it do occur, they are

frequentlyusedaftera discoursedeicticuseof that, in whatSchiffmancallsa “PronounChain”. A

similarobservationis madeby [Web88]. [GHZ93] notemoregenerallythetendency for it to prefer

referenceto focuseditems,while demonstrative pronounspreferreferenceto activateditems. For

example,in (2.35),Both usesof it refer to “becominga streetperson”;by the secondreference,

thispropertyis focused.thatprefersreferringto “becomingastreetpersonwouldhurthismother”,

which is not yet focused,andis highly dispreferredasthereferentfor thesecondit.

(2.35)Johnthoughtaboutbecomingastreetperson.It would hurt his motherandit/that would

make his fatherfurious.

Theoft-citedexamplein (2.36)shows what [GC00] and[Byr00] relatedlyclaim, thatpersonal

pronounstendto refer to entitiesdenotedby nounphrases,while demonstrativestendto referdis-

coursedeictically. In (2.36),thereferentof it is clearly“x”, while thereferentof that is clearlythe

resultof “add x to y”.

(2.36)Add x to y andthenaddit/that to z.

Thepreferenceof it to referto entitiesdenotedby nounphrasesandto referto abstractobjects

only afterthey arereferredto by ademonstrativesuggeststhatnounsaremoresalientthanverbsand

clausesasentities.[Byr00] however, notesthat thesalienceeffectson personalpronounresolution

canbe affectedby what shecalls “SemanticEnhancement”:with enoughpredicateinformation

gearedtowardahigherorderreferent,personalpronounscanmadeto preferhigherorderreferents,

26

Page 44: Thesis about discourse

asshown in (2.38c).

(2.37)Therewasasnake on my desk.

(2.38a)It scaredme.

(2.38b)Thatscaredme.

(2.38c)I never thoughtit wouldhappento me.(Sem.Enh)

[Eck98] notesa furtherdifferencebetweenthe resolutionof demonstrativesandpersonalpro-

nounsasdiscoursedeixis,which may indicatethat topicsaremoresalientthanverbsandclauses

asentities. In (2.39), that prefersreferenceto thespecificstory describedby Speaker A, while it

prefersreferenceto thetopic of child-carein general7. In fact,[ES99]doesnot considerthisuseof

it adiscoursedeicticuseatall; they treatit asa “vaguepronoun”.

(2.39)

Speaker A: Shehasa privatebaby-sitter. And, uh, thebabyjust screams.I

mean,thebabyis like seventeenmonthsandshejust screams.

Evenif sheknows thatthey’re gettingreadyto go out. They

haven’t evenleft yet...

Speaker B: Yeah,it/that’s hard.

[Lad66] andothersnotesubtlesaliencedifferencesbetweenthe discoursedeictic usesof this

andthat, relatedto theirspatio-temporaldifferences:this is usedwhenthereferentis close, andthat

is usedwhenthereferentis far.

2.3.5 Retrieving Antecedentsof DiscourseDeixis fr om the Tree

Many researchersfind thatdiscoursedeicticreferenceis dependenton discoursestructure.[Pas91]

uses(2.40)to show thattheclausalreferentof adiscoursedeicticis only availableif it immediately

precedesthedeictic. In (d), that cannotreferto sentence(a) unless(b) and(c) areremoved.

(2.40a)Carolinsistsonsewing herdressesform all naturalmaterials

(2.40b)andshewon’t evenconsidersyntheticlining.

(2.40c)Sheshouldtry thenew rayonchallis.

(2.40d)*That’s becauseshe’s allergic to synthetics.

7[GC00] alsoclaim thatprosodyplaysa role in resolvingdiscoursedeicticthat morethanit.

27

Page 45: Thesis about discourse

[Web91] arguesmore formally that though deictic referenceis often ambiguous(or under-

specified[Pas91]), thereferentis restrictedto theright frontier of thegrowing discoursetree. She

exemplifiesthisusing(2.41)-(2.42):

(2.41a)It’s alwaysbeenpresumedthat

(2.41b)whentheglaciersreceded

(2.41c)theareagot very hot.

(2.41d)TheFolsummencouldn’t adapt,and

(2.41e)they diedout.

(2.42)That’s what’s supposedto have happened.It’s thetextbookdogma.But it’s wrong.

Thediscoursedeictic referencein (2.42) is ambiguous;it canrefer to any of thenodeson the

right frontier of thediscourse:(thenodesassociatedwith) clause(2.41e),clauses(2.41d)-(2.41e),

clauses(2.41c)-(2.41e),clauses(2.41a)-(2.41e).

Discoursedeicticambiguityextendsto within theclauseaswell [Sch85], [Sto94]. For example,

in (2.43),thereferentof that couldbeany of thebracketedelements:

(2.43a) [ It talksabout[ how to [ go about[ interviewing ]]]

(2.43b)andthat’s goingto beimportant.

As notedby [DH95], the standardview on anaphoricprocessingis that we “pick up” the in-

terpretationof theantecedent,andthat in thenormalcase,thereis a coreferencerelationbetween

the antecedentandthe anaphor. The coreferencerelationis oneof identity, andthe antecedentis

“there”, waiting to be“pickedup”. Thus,in (2.44),mygrandfatheris saidto becoreferentto he:

(2.44)My grandfatherwasnota religiousperson.Heevenclaimedtherewasno god.

However, thefact that theinterpretationsof discoursedeixisarenot grammaticalizedasnouns

prior to discoursedeictic reference,andthefact that therearestructuralrestrictionson their refer-

ence,leadssomeresearchersto arguethat they arenot presentasentitiesin the discoursemodel

prior to discoursedeictic reference.Accordingto theseresearchers,their entity readingis coerced

andaddedto thediscoursemodelvia discoursedeicticreference.

28

Page 46: Thesis about discourse

Typecoercionis atermtakenfrom computerscience,whereit definesanoperationby whichan

expressionwhichis normallyof onelogical typeis re-interpretedasanother(e.g.whenanintegeris

understoodasa Booleanvalue).Typecoercionis usedto explain a rangeof linguistic phenomena,

suchaswhenanexpressionwhich is indeterminateasto logical typeis ’coerced’into oneparticular

interpretationandthusacquiresa fixedtype.Modelsof how coercionis achievedvary.

[Web91] arguesthat deictic useis an ostensive act, that distinguisheswhat is pointedto and

whatis referredto, whichmaybethesame,but neednotbe.Thisostensiveact,functionsto reify, or

bringinto thesetof entities,somepartof theinterpretationsof clauseswhichwerenotpresentin the

setof entitiesprior to theostensive act.Sheusesreferringfunctions8 to modelhow thereificationis

achieved,becausethey allow thedomainof what is pointedto (demonstratum)to bedistinguished

from therangeof whatis referredto (referent):

f : D K R , whereD is comprisedof focusedregionsof thediscourse,andR is asetof possible

interpretations.

In (2.41), the domainof the referring functionsare the elementsat the right frontier of the

discourse,and function applicationyields a rangeof event tokens(things that can happen). By

virtue of thereferringactionof thefunction,thesenew ‘entities’ (eventtokens)areaddedto E.

[Sto94] takesWebber’s modelonestepfurther, arguing that a discoursedeictic pronounwill

take its referentfrom therightmostsibling of theclausein which it is contained, onceits clauseis

attachedor adjoinedto the tree. That referentcannotbe found in a nodethatdominatesthenode

containingthediscoursedeicticis easyto see,becausethatwould make thedeicticself-referential,

asin (2.45),wheretheindice + indicatesthediscoursesegmentwhoseinterpretationis thereferent

of thediscoursedeictic. As theexamplemakesclear, a discoursedeicticcannotalmostnever refer

to a segmentin which it is contained.Theonly exceptionis textual deixis,asin (2.46),wherethe

demonstrative canreferto the text in which it is contained.

(2.45)*[ wxVy+#( � is aneatidea.]�(2.46)[ wxVy+#( � is a truesentence.]�

8Referringfunctionshave beenusedby [Nun79] to modelhow nounsin generalachieve their reference.

29

Page 47: Thesis about discourse

To arguethat thereferentwill not befoundin a nodethat is dominatedby thenodecontaining

thediscoursedeictic,[Sto94] first evokestheuseof discourserelations,arguing that if a discourse

deicticrefersto a segment,it will alsobein a discourserelationwith thatsegment.He thenargues

thatwhile discoursedeicticreferenceto embeddedclausesmight beviewedasanexceptionto this

generality, this exceptioncanbe avoidedby replacingWebber’s useof referringfunctionswith a

possibleworld semanticsin which thesemanticinterpretationof theelementsat the right frontier

of the discoursemake a variety of ’entity’ interpretations,or “information states”(see[Kra89]),

availableto thediscoursedeictic. For example,he arguesthatmodality in (2.47)makesavailable

assertionsaboutat leasttwo informationstates:(1) Mary left, and(2) Johnthoughtthe context

assertedof (1). The discoursedeictic in (2.48a)referencesthe first informationstate,andthat in

(2.48b)referencesthesecondinformationstate.

(2.47)JohnthoughtMary left.

(2.48a)He thoughtthishappenedyesterday.

(2.48b)Thiswaswrong.

[DH95] take a view similar to [Web91], exceptthey arguethat typecoercionis just oneof the

possiblereferent-creatingoperationsevokedby theuseof adiscoursedeictic.They arguethateach

timeananaphoris used,thedegreeto whichits antecedentis “there” will vary, andtheeffort needed

to “pick it up” will vary. In theirview traditional“coreference”asthemosttrivial case:theresultof

applyingtheidentity relationto theantecedent’s extension.They proposethatat leastthefollowing

operationsareneededto explainhow thereferentof a discoursedeicticis created:H Summationandcomplex creation:

Theseoperationsassemblesets.A setcanbeassembledby logicalconjunction,asin (2.49),or

by otherdiscourserelations,asin (2.50)(bracketsindicatethediscoursewheretheoperation

createstheantecedent):

(2.49)[Interestratesrose.Therecessionmayreduceinflation. Capitaltaxationis lower.] This

meansbrightertimesfor thosewhohave money to save.

(2.50)[If a white persondrivesthis carit’s a “classic”. If I, a Mexican-American,drivesit, it

30

Page 48: Thesis about discourse

is a “low-rider”. ] Thathurtsmy pride.H Typecoercion:

This operationis asabove, when the semanticsof an elementin the clausecontainingthe

deicticcausesanexpressionto becoercedinto oneparticularinterpretation.For example,the

verbcancoerceaninterpretation,asin (2.51)where“happen”coercesaneventinterpretation,

or thepredicatenominalcancoerceaninterpretation,asin (2.52).

(2.51)Mary wasfired. Thathappenedlastweek.

(2.52)I turnedleft. Thiswasawisedecision.H AbstractionandSubstitution:

The abstractionoperationabstractaway from specificevents,as in (2.53), wherethe an-

tecedentis “beatingone’s wife” not “Smith’s beatinghis wife”, while thesubstitutionopera-

tion substitutesoneelementof theantecedentfor another, asin (2.54),wheretheantecedent

is “X beatshis wife”:

(2.53)Smithbeatshis wife althoughthiswasforbidden50yearsago.

(2.54)Smithbeatshis wife andJohndoesit too.

Regardlessof whetherwe assumethatclausesalreadymake availablea setof semanticvalues,

or whetherwe usea referringfunctionor oneof any numberof operationsto representhow these

valuesaremadeavailable,discoursedeixisusedoesn’t determinewhich entity interpretation(s)is

(are)chosenasthe referent. Within the domainof the right frontier, the semanticsof the clause

containingthediscoursedeicticwill determinewhichof theavailableobjectsareselected.

In particular, as[Ash93] notes,thesub-categorizationframeof theverbshouldrestrictthepos-

sible referents.Sowhile theembeddedclausein (2.55)canbe interpretedasa variety of abstract

objects,thinkssub-categorizesfor a propositioninterpretationof “Mary is a genius”,asdoesthe

complex form becertainof. Similarly, happensub-categorizesfor aneventinterpretation,surprise

sub-categorizesfor a factinterpretation.

(2.55)Johnthinksthat[Mary is agenius].Johnis certainof it.

31

Page 49: Thesis about discourse

2.3.6 Summary

In this section,we overviewed the three-tieredmodelof discourseof GroszandSidner, in which

threeinteractingcomponentsaccountfor the structuringof the text into segments,eachof which

servesa purposeandcreatesa salientspacecontainingthe informationrelevant to that segment.

Their focuson intentionsasthe“relations” linking discoursesegmentsreflectsamoregeneralprag-

maticapproachto discourse.In sometreatmentsdiscourseis viewedasa plan,structuredinto sub-

goalswhosefulfillment achievesanoverall goal(see[LA90] for references).Othertreatmentscon-

cerntherole of discoursesegmentsin argumentunderstanding[Coh84]. RelevanceTheory[SW86]

is anotherpragmaticapproachproposedasa modelof a hearer’s interpretationprocess,in which

therelevanceof every segmentto thecontext is determinedbasedon a numberof interactingcon-

straints.Currentwork in dialogue(c.f. [SIG02]) investigatesintentionsin termsof “dialogueacts”.

BecauseGroszandSidnerview thepossibleintentionsthatcanunderliea discourseasinfinite,

structuralrelationsbetweensegmentsplay an importantrole in discoursecoherence.By distin-

guishinga focusspacefor eachsegmentthey modelsub-clausalcoherenceacrosssegments,and

enableCenteringTheoryto modelsub-clausalcoherencewithin segments9. Webber’s tree-based

modelsimplifiesthe GroszandSidnermodelby combiningthe componentsof text structureand

attentionalstateinto asinglestructure,while keepingthemconceptuallydistinctvia thenotionof a

right frontier, which is shown capableof modelingconstraintson discoursedeicticreferenceto the

interpretationof discoursesegments.

Groszand Sidner’s model (and Webber’s revision) provides a detailedaccountof the high-

level structuringof text in termsof attentionandintention; relationsbetweenpropositionsplay a

subsidiaryrole, servingalong with cuephrasesand utterancelevel intentionsto help the reader

recognizethe writer’s intentionsfor eachsegment. The detailsof how propositionalrelationsare

computedrecursively to build discoursesegmentsandhow theresultsarerepresentedarenot pro-

vided. In thenext section,wewill discussa tree-basedmodelthatadoptsasimilarly simpleview of

propositionalrelations,but definestheir role in theconstructionof thetreeprecisely, andclaimsto

modelsub-clausalanaphoraresolution.

9Someresearchershave alsousedCenteringasa modelof anaphoraresolution(see[BWFP87])

32

Page 50: Thesis about discourse

2.4 A TreeStructure with a Syntax-SemanticInterface

[Pol96] presentsa context free grammar(LDM) for incrementaldiscourseparsing[SP88]similar

to [Web91], but combinedwith grammarrulesincorporatingpropositionalrelationsanda dynamic

logic framework (DQL) for describingthe structuredsemanticcomponentthat resultsfrom the

parsingprocess[PSvdB94]. The resultingmodelprovidesanaccountof how anaphoraresolution

[vdB96, PvdB99, Lag98] andtemporalinterpretation[PvdB96] work acrossstretchesof discourse.

2.4.1 Constituentsand TreeConstruction

In LDM, thesurfacestructureof discourseis composedof discourseconstituentunits(DCUs)and

discourse operators (DOs). DCUs aresemanticallymotivatedstructuresthat carry propositional

information;an elementaryDCU, typically a clauseor sentence,is any minimal utteranceencod-

ing a singleeventor stateof affairs indexed for context, includingphysicalandsocialsituationof

utterance(real or modeled),genreunit, modality, polarity, andpoint of view. DOs expressnon-

propositionalinformation,suchassemanticandstructuralrelationshipsamongDCUs, andprag-

maticinformationabouttheattitudeof thespeaker andthesituationof utterance.Examplesof DOs

includelogical operators,vocatives,(dis)affirmations,certainparticles,exclamations,connectives,

andtemporalmodifiers.

Complex DCUs aredefinedrecursively, via theattachmentof elementaryDCUs to DCUs the

growing parsetreeto createoneof threetypesof structures:H coordinations, includinglistssuchastopicchainsandnarrativesH subordinations, includingelaborationsandinterruptionsH binary-attachments, includingadjacency pairs,logical relationsandrhetoricalstructures

LDM canbe viewed asa limited lookaheadparserwhich acceptselementaryDCUs asinput

andbuildssimultaneouslyastructuralandsemanticdiscourserepresentation[SP88].A discourseis

representedasanopenright discourseparsetree,composedof C (coordination)S (subordination)

33

Page 51: Thesis about discourse

andB (binary-attachment)non-terminalnodesandelementaryDCUsasterminalnodes.Eachele-

mentaryinput DCU is attachedasthe right child of anavailable, existing or newly created,node.

As in [Web91], only DCUson theright edgeof thegrowing parsetreeareavailablefor theattach-

mentof anincomingDCU; theoperationsfor attachinganincomingDCU correspondto [Web91]’s

attachmentandadjunctionoperations[Gar97b]; right attachmentis exemplifiedin Figure2.5.

Figure2.5: LDM Right-AttachmentOperation

2.4.2 The SyntaxSemanticInterface

Simultaneouswith theincrementalconstructionof discoursestructure,thesemanticrepresentation

of thediscourseis updatedwith theinterpretationof theincomingDCU. LDM is in essencea typed

unificationbasedsentencegrammaraugmentedwith a setof discoursegrammarrules. EachDCU

containssemanticinformationin theform of typedfeaturestructures,wheretypesareorderedalong

a typehierarchywhich allows for informationinheritanceandtypeunification10. For example,the

featurestructurefor thebasicDCU Johnsmiledis shown in (2.56),wherebasicrepresentsthetype

of elementaryDCUs and (/z�{q| representsthe semanticsof John smiled. The SCHEMA featureis

identicalto theSEM featurein basicDCUs.

(2.56) }~~�#� j�(�+#iSEM (qz�{/|SCHEMA (qz�{/|

�&���10Additional informationis alsocontainedin eachDCU, asmentionedabove; see[Pol96] for details.

34

Page 52: Thesis about discourse

Discoursegrammarrulesthenspecifyhow to combineDCUs into biggerDCUs,andhow the

featurestructuresof thechild constituentscombineto yield thesemanticsof theparentconstituent.

DOs signal the applicationof thesegrammarrules. For example,constructionof a coordination

structurecanbeillustratedby thegrammarrule for a list, statedasin (2.57).In [PSvdB94], coordi-

natingconjunctions� and,or,...� signalthata list is to beconstructedor extended.

(2.57)}~~� )r+�(�bSEM )r+�(�b�z4� .�_�M��%� .�_ g |SCHEMA ^q.�W�z4� .�_ M �%� .�_�gC|

�&��� K}~~� -`b#op.*.SEM � .�_�MSCHEMA ��i�VA.�_�j M

�&��� ,

}~~� -pb�op.C.SEM � .�_ gSCHEMA ��i�V;.�_�j/g

�&���Syntactically, this rule statesthat any two discoursetreescancombineto form a new treeof

type list. Semantically, this rule producesa list relationbetweenthe two trees(indicatedby the

featureSEM), andconstrainsthis relation to be betweentwo treeswhosegeneralizationis non-

trivial, wherethegeneralizationof two terms � .�_ M and � .�_�g yieldsthemostspecifictermwhich

subsumesboth ��.�_ M and ��.�_�g . TheSCHEMAfeatureindicatesthis term.For example,giventhe

list JohnsmiledandPhil cried, thevalueof theSCHEMA featurewill bethegeneralizationof the

twoclauses,roughlymanexpressedemotion, wheremanandexpressedemotionaresetof objects

sharinga commonproperty11. Notethatthis rule extendsa list relationwhenthegeneralizationof� .�_�g is ��i�V;.�_�j M .In [Lag98]’s versionof LDM, therule for constructingopV;.�b�T�o7+'i�jq) coordinations(e.g.abinary-

attachment)is very similar to the rule for lists, exceptthata (non-trivial) generalizationis not re-

quired.Rhetoricalcoordinationsaresignaledby DOssuchastherefore, so,thus,accordingly.

Theconstructionof subordinations, whichmaybesignaledby DOssuchasbecause, sincediffer

from coordinationsin thatno generalizationis calculated.TheSCHEMA andSEM featuresof the

parentDCU in subordinationstructuresaregiventhevaluesof themainclause.

11see[PSvdB94,vdB96] for theformal computationof SEMandSCHEMAin DQL.

35

Page 53: Thesis about discourse

2.4.3 Retrieving Antecedentsof Anaphora fr om the Tree

LDM claimsto modelconstraintson theantecedentsof anaphoraof incomingDCUs,asillustrated

in example(2.58),takenfrom [PvdB99].

(2.58a)Susancamehomelateyesterday.

(2.58b)Doris hadheldherupat work.

(2.58c)Shedidn’t evenhave time for dinner.

In (2.58),therelationshipbetweenDCU (a) andDCU (b) is a subordinationbecauseDCU (b)

suppliesmoredetailedinformationaboutwhy Susancamehomelate.DCU (c) continuesdescribing

thestateof affairsof Susan’s evening,andis thereforein a coordination relationwith (a). Dueto

the specificationof the semanticsat eachnodethat is provided in the LDM grammarrules for

subordinationandcoordination,only Susanis availableasa potentialreferentfor theanaphor, she,

in (c). The treefor this discourseis shown in Figure2.6, wherethe featurestructureshave been

simplifiedto show only thetypeof DCU andtheavailableantecedents.� i8T7T�o>-p+4WUj/b�+'T�WSusan �� � � � � �������� (Ck � T�o>-p+4WUj/b�+'T�WSusan �� � � ������ � jq(C+'i (a)

Susan � � � jq(C+'i (b)

Susan, Doris�� � jq(C+'i (c)

Susan �Figure2.6: LDM TreeStructurefor Example(2.58)

2.4.4 The NeedFor Upward Percolation

[Gar97b] observesthatwhile theLDM discoursegrammarrulesspecifyhow thesemanticsof child

DCUs combineto yield the semanticsof the parentDCUs, the semanticsof the restof the tree

remainsunchanged.Shearguesthattherearetwo mainproblemswith this lack of “upwardperco-

lation”. First,thesemanticsof thediscoursecannotbereadoff eithertherootor theright frontierof

36

Page 54: Thesis about discourse

thetree.That it cannotbereadoff theroot is relatively obvious; that it cannotbereadoff theright

frontier is illustratedusing(2.59). TheLDM treeassociatedwith this exampleis shown in Figure

2.7;whereonly thesemanticsof eachDCU is shown at eachnodefor simplicity.

(2.59a)Weweregoingto seeoursontonight

(2.59b)but we arenot

(2.59c)becausetheyoungeroneis cominghomefor dinner

(2.59d)becauseheis working in theneighbourhood

(2.59e)soheis cominghomefor dinner

(2.59f)sowe arenot

a but b� � � � ������j ((b becausec) sof )� � � � ������(b becausec)� � � ������ ((c becaused) soe)� � ����

(c becaused)� ���i - .f

Figure2.7: LDM TreeStructurefor Example(2.59)

If thesemanticsof thediscoursearereadoff theright frontier andconjoined,we getonly: ((a

but b) and((b becausec) sof) andf).

Thesecondmajorproblem[Gar97b] noteswith LDM’ slackof upwardpercolationisaninability

to retrieve the antecedentsof discoursedeictic reference.As [Web91] argues,the right frontier

shouldrepresenttheavailableantecedentsof discoursedeixis; obviously if the right frontier does

not containtherequiredinformation,it will notbeavailableto thediscoursedeictic.

Gardentproposesanalternative methodof treeconstructionanda specificationof thesyntax-

semanticinterfacein which upwardpercolationis incorporated,alongwith thediscoursegrammar

rules of LDM. We will discussher approachin detail in Chapter4. [Sch97]’s extensionof her

approachincorporatesasemantic-pragmaticinterface.

37

Page 55: Thesis about discourse

2.4.5 Summary

In summary, LDM providesa preciserepresentationof discoursestructure,a formal representation

of discoursesemantics,andaspecificcalculationof theinformationavailableat intermediarynodes.

Theauthorsarguethat for this reasontheLDM modelhasanadvantageover [GS86], which relies

ontheinferenceof attentionalandintentionalstates.However, theLDM modelmakesnoreference

to intentionsat all, andthe inferenceprocessis not described,althoughthey do notethat appeals

to inferenceandworld knowledgearerestrictedto specificmomentsin the interpretation,i.e. the

momentof DCU attachmentto thediscoursetree. Moreover, while LDM, like [GS86], claimsto

provideanaccountof anaphoraresolution,Gardentshowsthatthelackof upwardpercolationmakes

it unableto accountfor theresolutionof discoursedeixisanaphora.

In thenext section,we will discussa different tree-basedmodelwhich, thoughit returnsto a

descriptive approachof discoursecoherence,is widely used,becausein addition to providing an

extensive descriptionof propositionalrelations,it alsodefinesthediscoursestructuresthat canbe

producedwith them. We will then look at an alternative theory which replacessyntacticstruc-

ture with a structuredsemantics,andmodelsworld knowledgeandthe inferenceof propositional

relations.LDM arguesthis modelis lesstractable,becauseit doesnot separatethesyntaxandse-

manticsof discourse,andreferenceto world knowledgeandinferenceis lessrestricted.Thetheory

goesfurtherthanLDM however, by presentinganaccountof aseparatecomponentof intentions.

2.5 A DescriptiveTheory of DiscourseStructure

2.5.1 Analyzing Text Structure

RhetoricalStructureTheory(RST)[MT88] is oneof thesimplestmodelsof discourse,in thatit is a

purelydescriptivetheoryof text organization.RSTdescribestext structurefrom thepoint of view

of a text analyst,who hasaccessto thetext, knowledgeof context andthecultural conventionsof

thewriter, but doesnot have accessto thewriter. Therefore,theanalyst’s job is to judgethemost

plausiblerelationsthat thewriter intendedto convey. This judgmentdoesnot rely on morpholog-

ical or syntacticsignals;theauthorsclaim to have foundno reliableunambiguoussignalsfor any

38

Page 56: Thesis about discourse

relations.Recognizingintendedrelationsis to reston functionalandsemanticjudgmentsalone.

The analystchoosesfrom the list of RST relationsin Table2.10. In this table, relationsare

groupedaccordingto their similarity in definition; theauthorsacknowledgethatalternative group-

ingsarepossible,oneof whichis thedistinctionbetween“subjectmatter”relations,whoseintended

effect is that the readerrecognizesthe relation,and“presentational”relations,whoseintendedef-

fect is to increasesomeinclination in thereader, suchasthedesireto actor believe someassertion.

“presentational”relationsareitalicized.

Table2.10:Organizationsof RSTRelationDefinitions

EvidenceandJustify AntithesisandConcession RestatementandSummaryEvidence Antithesis RestatementJustify Concession Summary

Relationsof Cause ConditionandOtherwise BackgroundVolitional Cause Condition EnablementandMotivationNon-Volitional Cause Otherwise EnablementVolitional Result InterpretationandEvaluation MotivationNon-Volitional Result Interpretation OtherRelationsPurpose Evaluation Sequence

Circumstance Elaboration ContrastSolutionhood

Nuclearityis assumedto bea centralorganizingprincipleof text; for themajority of relations,

thepiecesof text beingrelatedcanbedistinguishedinto a WckAi�)2.�k�( andsatellite, with thenucleus

representingthewriter’s maincommunication,andthesatellitesproviding subsidiaryinformation.

Thepredictionis thatif thenucleusis removed,thesignificanceof theinformationin thesatellite(s)

will not be apparentand thereforethe text will be incoherent,but if the satelliteis removed, the

resultingtext will still be coherentand resemblethe original in the form of a “synopsis”. RST

relationsarethusdefinedin termsof nuclearityandthewriter’s intent.As anexample,thedefinition

of the Evidencerelation is given in Table2.11, whereR representsthe reader, W representsthe

writer, N representsthenucleus,andSrepresentsthesatellite.

Thetext analystusesRSTrelationsto relatetext spans. Atomic text spansaregenerallyclauses,

exceptthatclausalsubjects,complementsandrestrictedrelative clausesarenot treatedin [MT88]

39

Page 57: Thesis about discourse

asindependenttext spans.

Table2.11:Evidence:RSTRelationDefinition

relationname EVIDENCEconstraintson N R might notbelieve N to adegreesatisfactoryto Wconstraintson S R believesS or will find it credibleconstraintson N+S R’scomprehendingS increasesR’sbelief of Ntheeffect R’sbelief of N is increased

Complex text spansarestructurescalledschemaapplications. A schemaapplicationis a setof

adjacenttext spans(atomicor complex) linkedby anRSTrelationaccordingto oneof fivestructural

arrangements,calledschemas. Eachrelationhasacorrespondingschema,exemplifiedin Figure2.8;

therelationsnot shown all usetheschemalabeledwith the“circumstance”relation.Arcs represent

the relationholding betweentext spans,which are representedby horizontallines. The nucleus

is distinguishedfrom the satelliteby the direction of the arrow, and eachvertical line descends

from the text spanbeingdecomposedby a schemaapplicationdown to thenucleusof theschema

application.

Figure2.8: RSTSchemas

Schemasdonotconstraintheorderingof nucleusor satellite,they allow arelationthatis partof

a schemato beappliedany numberof times,andin multi-relationschemas,they requireonly one

of therelationsto hold. However, anumberof constraintsmustbesatisfiedto producea valid RST

structure.An RSTstructuremustbecomplete, consistingof asetof schemaapplicationscontaining

40

Page 58: Thesis about discourse

aschemaapplicationthatconstitutestheentiretext. Every text span,exceptfor theentiretext itself,

mustbe connectedaseitheran atomicspanor a constituentof anotherschemaapplication.Each

schemaapplicationmustbeunique, consistingof adifferentsetof text spans,andin amulti-relation

schema,eachrelationmustapplyto adifferentsetof text spans.Finally, adjacencymustbesatisfied,

in thattheresultof eachschemaapplicationconstitutesonetext span.

An exampleof an RST structureconsistingof two evidencerelationsis shown in Figure2.9.

Eachtext spanis numbered;atomicnumberscorrespondto the text units in (2.60),andcomplex

numbersrepresentundecomposedunitsof thestructure.

(2.60)

(unitM ) This computertax programreally works.

(unitg ) In only a few minutesI finishedmy tax return.

(unit� ) I printedit for you to see.

Figure2.9: EvidenceRelation

[MT88] allow thatmultiplicity mayarisefrom text ambiguityandresultingdifferencesin ana-

lysts’ judgmentsabouttherelationsholdingbetweentext spans.It is assumedhoweverthatthemore

coherentthewriter hasmadethetext, theclearereachchoiceof relationis for thetext analyzer.

41

Page 59: Thesis about discourse

2.5.2 The Needfor Multiple Levelsof DiscourseStructure

[MP92] provide an influential argumentagainst[MT88]’s claim that thoughsubject-matterand

presentationalrelationscanbe distinguished,in general,therewill be a singlepreferredrelation

holding betweenconsecutive text spans.They arguethat this distinctionis in fact a conflationof

the two levels of discourseinterpretationidentified by [GS86]: the informational level, and the

intentional level, respectively, andthata completemodelof discoursecannotdependon analyses

in which theselevels are in competition. They supportthis argumentfirst by showing both that

informationcanflow betweentheselevelsto producetherelationsbetweentext spansin adiscourse,

asillustratedwith (2.61)-(2.62):

(2.61)GeorgeBushsupportsbig business.

(2.62)He’s sureto vetoHouseBill 1711.

[MP92] arguethat thepresentationalrelationEVIDENCE is a plausibleRSTrelationbetween

thenucleus(2.62)andthesatellite(2.61)(Table2.11). Equallyplausible,however, is thesubject-

matterVOLITIONAL-CAUSErelation,where(2.62)is thenucleusand(2.61)thesatellite.[MT88]’s

definitionof aVolitional-Causerelationis givenin Table2.12.

Table2.12:Volitional-Cause:RSTRelationDefinition

relationname VOLITIONAL-CAUSEconstraintsonN N presentsa volitional actionor a

situationthatcouldhave arisenfrom avolitional action.

constraintsonS noneconstraintson theN+Scombination S presentsasituationthatcould

have causedtheagentof thevolitional actionin N to performthataction;without thepresentationof S,R might not regardtheactionasmotivatedor know theparticularmotivation;N is morecentralto W’spurposesin puttingforth theN-ScombinationthanS is.

theeffect R’s recognizesthesituationpresentedin Sasacausefor actionof N

42

Page 60: Thesis about discourse

[MP92] argue that if the readerknows (andknows that the writer knows) that the bill places

stringentcontrolson manufacturing,thens/hecanconcludethat(2.61)is evidencefor (2.62),thus

reasoningfrom informationto intention. Alternatively, if the readerdoesn’t know anything about

thebill, but expectsthewriter to supporttheclaimthatBushwill vetoit, thens/hecanconcludethat

(2.61)is acauseof (2.62),thusreasoningfrom intentionto information.

[MP92] also show that spanscan be relatedsimultaneouslyon both levels, as in (2.63). At

the informationallevel, a plausibleRST analysisis that (2.63c) is the nucleus,the writer’s main

informationalcommunication,(2.63 a) is a CONDITION on (2.63 b), and (2.63 a) and(2.63 b)

togetherarea CONDITION on (2.63c). This RSTstructureis shown first in Figure2.10.Suppose

howeverthatthewriter is planningasurprisepartyfor thereader. Thenattheintentionallevel, (2.63

a) is thenucleus,theactionthatthewriter wishesthehearerto perform,(2.63c) MOTIVATES(2.63

b), andtogetherthey MOTIVATE (2.63a). ThisRSTstructureis shown secondin Figure2.10.

(2.63a) Comehomeby 5:00.

(2.63b) Thenwe cango to thehardwarestorebeforeit closes.

(2.63c) Thatwaywe canfinish thebookshelvestonight.

Figure2.10:RSTConditionandMotivationRelations

Becausetheintentionalandinformationalstructuresfor this discoursearenot isomorphic,they

cannotbeproducedsimultaneously.

43

Page 61: Thesis about discourse

2.5.3 “Elaboration” asReference

[KOOM01] show that the object-attribute elaboration relationover- andunder- generatesRST

structures.[MT88] definetheelaboration relation(of which theobject-attribute typeis subset5)

asshown in Table2.13.

Table2.13:Elaboration:RSTRelationDefinition

relationname ELABORATIONconstraintson N noneconstraintson S noneconstraintson theN+Scombination Spresentsadditionaldetail

aboutthesituationor someelementof thesubjectmatterwhich is presentedin N orinferentiallyaccessiblein N inoneor moreof thewayslistedbelow. In thelist, if N presentsthefirst memberof any pair,thenS includesthesecond.1. set: member2. abstract: instance3. whole: part4. process: step5. object: attribute6. generalization: specific

theeffect R’s recognizesthesituationpresentedin Sasprovidingadditionaldetailfor N.R identifiestheelementofsubjectmatterfor whichdetail is provided.

In words,[MT88] definetheobject-attribute elaboration relationto hold betweenN andS if

N ‘presents’an object (e.g. containsa mentionof it), andS subsequentlypresentan attribute of

this object. [KOOM01] have found this relationto be widely usedin museumguidebooks.They

illustratethefactthatit under-generateswith (2.64a)-(2.64d):

(2.64a) In thewomen’s quartersthebusinessof runningthehouseholdtookplace.

(2.64b) Muchof thefurniturewasmadeupof chestsarrangedvertically in matchingpairs.(...)

44

Page 62: Thesis about discourse

(2.64 c) Femaleguestswere entertainedin theserooms,which often hadbeautifully crafted

woodentoilet boxeswith fold-away mirrors andsewing boxes,andfolding screens,paintedwith

birdsandflowers.

(2.64d) Chestswereusedfor thestorageof clothes.. .

In (2.64b), theobjectchestsarementioned.Discussionof this objectis takenup againin (2.64

d). Thetext is clearlycoherent,but thedesiredRSTanalysiswhere(2.64d) is asatelliteof (2.64b)

usingobject-attribute elaboration is not possible,because(2.64b) and(2.64c) arebothalready

satellitesof (2.64a) underanelaboration relation.

[KOOM01] illustratethefactthattheobject-attribute elaboration relationunder-generatesby

comparing(2.65a)-(2.65b) with (2.66a)-(2.66c) :

(2.65a) Arts-and-Craftsjewelstendto beelaborate.

(2.65b) However, this jewel hasasimpleform.

Thediscoursein (2.65a)-(2.65b) displaysa concessionrelationbetweenthesatellite,(2.65b),

andthenucleus,(2.65a). However, whenanobject-attribute elaboration relationintervenes,asin

(2.66a)-(2.66c) thediscourseis incoherent,althoughRSTallows thestructure,with (2.66c) now

beingthesatellitefor thecomplex unit (2.66a)-(2.66b), with (2.66a)asthenucleus.

(2.66a) Arts-and-Craftsjewelstendto beelaborate.

(2.66b) They areoftenmass-produced.

(2.66c) However, this jewel hasasimpleform.

If compositionality, continuousconstituency, and a tree structurearemaintainedas RST as-

sumptions,thenRSTundergenerates.[Sib92] hasarguedfor relaxingthecontinuousconstituency

constraint(i.e. that S be adjacentto N, or adjacentto a satelliteof N). [KKR91] hasarguedfor

relaxingthe treestructureconstraint(i.e. that eachtext span,exceptfor the spanconstitutingthe

entiretext, beinvolvedin exactlyoneschemaapplication,with noover-lappingspans,andnospans

not linkedto otherspans).

[KOOM01] take anothertack, arguing that the over- andunder-generationproblemsarisebe-

causeobject-attribute elaboration is notadirectrelationbetweenpropositions,but rather, adirect

45

Page 63: Thesis about discourse

(e.g. identity) relation betweenobjects,and a spuriousassociationbetweenpropositions. They

arguethat local andglobal focusmechanisms[WJP81, GS86]make theuseof this relationredun-

dant. They proposea revision of RSTin which theobject-attribute elaboration is removed,and

high-level text spansarerelatednot by RSTrelations,but by mechanismsof global focussuchas

informationallyredundantutterances(IRUs)[Wal93], nominalization,anddiscoursedeixis.

In [KOOM01]’s model,a coherenttext is a sequenceof focusspaces,calledentity-chains, that

succeedeachotherin a legalmanner, asexemplifiedin Figure2.11.

Figure2.11: [KOOM01]’sDiscourseModel

In thefigure,eachentity-chain,labeledEC1,etc,hasasits globalfocusanentityE,andconsists

of asequenceof RSTtrees,eitheratomictext spans,shown assmallboxes,or complex structuresof

RSTrelations(minustheobject-attribute elaboration relation),shown astriangles,in which the

root nucleusof eachtreeis aboutE. In a legal sequenceof entity-chainsthefocusedentity in each

chainhasbeenmentionedsomewherein the W previous propositions,wherethevalueof W is still

anopenquestion.They call this a resumptionrelation, indicatedby directedarcsin thefigure,and

its felicitoususeis claimedto bea functionof its lineardistancefrom any previousmentionof it,

ratherthanbeinga functionof its relationshipto theright frontier of a discoursestructuretree.

2.5.4 Summary

RST claims to be able to describethe majority of naturally occurringtext, and hasfound wide

applicationin the literature. For example,[Fox87] hasdemonstratedhow humanexplanationsof

thechoicebetweenpronounsandfull NPsin expositorytext canbederived from RSTstructures,

[SdS90] defineasetof heuristicsfor recognizingRSTrelations,[Mar97] hasbuilt anRSTannotation

46

Page 64: Thesis about discourse

tool and manualin which usersare instructedto selectfrom a hierarchicalversionof the RST

relations(e.g. they chooseelaboration only if no other relationfits) and[Mar00] shows that an

automatedmethodof labelingRST-typerelationscanpartially replicatehumanannotation,if “cue

phrases”andpunctuationareusedassignalsof thepresenceof particularRSTrelations.RST-type

relationshave alsobeenusedin combinationwith otherintentionalrelationsasplanningoperators

in naturallanguagegenerationsystems(see[Hov93, MP93]).

However, anumberof importantlimitationsinherentin theRSTapproachhavealsobeenraised.

First, MooreandPollackshow that RSTconflatesintentionswith treestructure.This hasserious

consequences.For example,anRST-basedsystemsuchas[Mar99] enablesoneto automaticallyde-

riveandenumerateall possibleRSTinterpretationsof a text, but doesnotprovide amechanismfor

choosingbetweenthem. Similarly, [Hov93]’s RST-basedsystemallows for multi-relationdefini-

tionsthatassigntwo labelsto consecutive discourseelements.Neithersystem,however, accommo-

datesconcurrent,non-isomorphicinterpretations.This is a problemfor all tree-basedapproaches;

if both levels are treatedstructurallythenmultiple structureswill alwayshave to be considered.

Second,[KOOM01] showed that the object-attribute elaboration relationboth over- andunder-

generatesthe spaceof discoursestructuresin an RST theory. Ratherthanrevise the assumptions

inherentin atreestructure,they proposethattreestructuresshouldholdonly within “entity chains”,

whichcorrespondroughlyto discoursesegmentsassociatedwith focusspacesin GroszandSidner’s

theory. Thedifferenceis thatKnott et. al donotconstrainfocusspacesashighly asastackor a tree

structuredoes,ratherthey proposea linearconstrainton thenumberof intervening“entity chains”

betweenreferenceto entitieswithin prior entity chains.

More generally, RST is simply not a completemodelof discoursestructure,in that it relies

wholly on reader’s intuitions,saysnothingaboutsub-clausalcoherence,definesno formal mecha-

nismfor computingtherelationsbetweentext spans,andprovidesnoobjectivemethodof justifying

their choiceof relationsover any other. Becausethey make useof so many morerelationsthan

any of theotherrelationswe have seen,suchjustificationis important;we needto proposesome

mechanismby which theserelationsareproduced,especiallyif they intuitively seemcapableof

describingthemajority of texts.

47

Page 65: Thesis about discourse

In thefinal sectionof this chapter, suchamechanismis proposed.In thenext section,however,

we presentanalternative theoryof discoursecoherence,onewhichmodelsevery majorcharacteri-

zationof discoursecoherencethatwehavesofarseenin thischapterwithoutreferenceto asyntactic

treestructure.

2.6 A SemanticTheory of DiscourseCoherence

SDRT [Ash93, LA93] is a dynamicsemanticsapproachto discoursecoherence.It hastwo main

modules:a formal languagefor representingdiscoursecontext [Ash93], anda theoryof discourse

inference(DICE) [LA93] which is usedto computediscourserelations.[LA99] furtherproposea

separatemodulefor limited reasoningaboutintentionsandcognitive states.SDRT is alsoshown

to provide anaccountof anaphoraresolution;in particular, thesemanticsit proposesincorporates

thesemanticobjectsreferredto by discoursedeixis. We now returnto our discussionof discourse

deixis,in termsof thesesemanticobjects;thisdiscussionwhichwill provide thebackgroundfor our

studiesin Chapter3.

2.6.1 Abstract Objects

Sincetheearliestwork in logic andlinguistics,propositionshavebeenviewedasthesemanticinter-

pretationof sentences(c.f. [Mon74, CQ52]);studiesof adverbialmodificationandtensein formal

linguisticshave alsomadereferenceto eventualityinterpretations(c.f. [Dav67, MS88]). Proposi-

tions,aswell asstates-of-affairs, properties,facts,causes,andeffects, have no spatio-temporallo-

cation.For someeventualitiestoo, it maybedifficult to pinpointprecisespatio-temporalor sensory

coordinates:thefall of theRomanempire, for example[Ven67]. Theseinterpretationsareabstract.

Early work in naturallanguagephilosophy(c.f. [Ven67, Aus61, Str59]) providesdiscussionof the

propertiesof theseinterpretations.

The precisenatureof theseabstractionshasneverthelessproven difficult to pin down, partly

becausethey canbedifficult to distinguish.For example,thesentenceJohnwentto thestorecanbe

interpretedasanevent,it canbeattributeda truth value,it canbeviewedasa surprisingfact,or it

canbeinterpretedasaresultof John’sneedingclothes.Wecanview theseinterpretationsasobjects,

48

Page 66: Thesis about discourse

asevidencedby thefactthatwereferto themusingnounphrases,e.g.thefall of theRomanEmpire.

Thattheseobjectscanbedistinguishedis evidencedby thefactthatwecanexplicitly indicatewhich

interpretationis beingreferredto, e.g. thefact that theRomanempire fell, theeventof theRoman

empire falling, etc.

Vendlerarguesthat we implicitly distinguishtheseinterpretationby the way we nominalize

them. Nominalizationtransformsa sentenceinto a noun phrase. The “essentialingredient” of

a nominalizationis a verb derivative. Vendlerarguesthat nominalizationsfall into perfectand

imperfectclasses,asin Table2.14.

Table2.14: [Ven67]’s ImperfectandPerfectNominalizations

imperfectnominals (2.67)It is fortunatethat Johnhasarrived unexpectedly.(2.68)John’s havingarrivedunexpectedlysurprisedme.

perfectnominals (2.69)Thebeautifulsingingof theMarseillestookall afternoon.(2.70)Theunexpectedarrival wentunobserved.

As shown in (2.67),nounclausenominalizationsallow theverb to take tense,auxiliary verbs

andadverbs,but they cannotbe modifiedby articlesor adjectives (e.g. *the beautiful that John

arrived). As shown in (2.68),-ing formsof verbsin nominalizationscantakeeither tense,auxiliary

verbsandadverbs,or articlesandadjectives,asshown in (2.69),but not both at once(e.g. *the

beautifulsinging*unexpectedly). As shown in (2.70),nounsderivedfrom verbsin nominalizations

cannottake tense,auxiliaryverbsor adverbs(e.g.the*unexpectedly*havingarrival); they canonly

take articlesor adjectives.

Sincetenses,auxiliariesandadverbscharacterizeverbs,andaregenerallyincompatiblewith

articlesandadjectives in nominalizations,andarenot permittedat all in perfectnominalizations,

Venderconcludesthat in perfectnominalizations,the verb is deadasa verb, while in imperfect

nominalizations,theverbis still aliveasaverb.

Vendlerfurtherdistinguishesloosecontainers andnarrow containers (i.e. the restof thesen-

tence),as in Table 2.15. Narrow containersdescribespatio-temporalqualitiesof events. They

permitonly perfectnominalizations,in which, if thereis a verb, it is dead.Narrow containersalso

49

Page 67: Thesis about discourse

take nounswhichcanbehave like perfectnominals,e.g.firesandblizzards, unlike tablesandcows,

canoccur, begin, end,canbesuddenor prolonged;they canbereadasevents.

Table2.15: [Ven67]’s LooseandNarrow Containers

loosesubjectcontainers is probable/certain/unlikely/surprising/a factlooseobjectcontainers hedenied/mentioned/rememberednarrow subjectcontainers is slow/gradual/anevent/aprocess/anaction,occurs/beginsnarrow objectcontainers I watched,I heard,I felt, I observednarrow PPcontainers until/before/after/since

Loosecontainersascribepropertiesto facts.They permitboth imperfectandperfectnominal-

izations. Whenthey take an imperfectnominalization,theverb is alive; whenthey take a perfect

nominalization,however, analiveverbis attributedto it, asin (2.71).Whenthey takesimplenouns,

they arereadassuppressedimperfectnominals,asin ((2.72). Moreover, if a sentenceis not nomi-

nalized,it will berelativizedby a relative clausein theform of a loosecontainer(2.73).

(2.71)John’s singingof theMarseillessurprisedme.(read:thathesangtheMarseillaise)

(2.72)Theabominablesnowmanis a fact. (read:theexistenceof thesnowman)

(2.73)Johndied,whichsurprisedme.vs. *Johndied,whichwasslow.

Vendlerusesthesedistinctionsto investigatethenatureof otherabstract objects, suchaseffects

and results. Examplessuchas(2.74) - (2.75) indicatethat effectsdescribeperfectnominals,e.g.

events,changes,or processes.An effect canreacha large area,canbe felt, measured,registered,

canbeviolentanddangerous.Theseexamplesalsoshow thateffectsattributeeventsto otherevents,

alsodenotedby a perfectnominal.Similarly, othermembersof theeffect family of terms,product,

work,creation,upshot,issue, outcomearepredicatedof events.

(2.74)Themoon’s positionhasaneffect on themovementof theoceans.

(2.75)*That themoonhasits positionhasaneffecton theocean’s having movement.

Results, on theotherhand,like causes,consequences,reasons,motives,andexplanations, can

bestated,told, believed,probableor improbable,andsometimesfortunateor unfortunate,expected

or unexpected,sad,disastrous,or horrible,i.e.,describefactinterpretations(2.76).

50

Page 68: Thesis about discourse

(2.76)Thattheoceanshave movementis a resultof themoonhaving its position.

Theseinterpretationsaredifficult to distinguish;whatemergesmostclearlyfrom thisdiscussion

is thatmany differentabstractinterpretationscanbeconveyedastheinterpretationsof sentences.Of

course,nominalizationsarenot thewayto convey thesesentences,noraresinglesentencestheonly

carriersof theseinterpretations.Discoursedeixis, aswe have seen,providesanothermechanism

for conveying them. And discoursedeixis canrefer to not only to the interpretationsof sentences

but to awidevarietyof syntacticconstituents,includingverbalpredicates,asin (2.31,sequencesof

sentences,asin (2.33),anduntensedclauses,asin (2.43),all of whicharerepeatedbelow.

(2.31)John[smiled]. He doesthat often.

(2.33)[I wokeupandbrushedmy teeth.I wentdownstairsandatebreakfast,andthenI wentto

work.] That’s all I did today.

(2.43)[ It talksabout[ how to [ go about[ interviewing ]]] andthat’s goingto beimportant.

What’s more, discoursedeixis referenceindicatesthat the rangeof abstractobjectsis even

wider thanVendleraddressed.For example,discoursedeixiscanalsorefernot only to speechact

interpretationsof sentences,as [Web91] noted(Table2.9, shown below as2.77), but as [DH95]

have shown, to the discourserelation betweensentences,suchas the contrastrelation in (2.50),

repeatedbelow, andevento apresupposeddefeasiblerulearisingfrom adiscourserelationbetween

sentences(See[Kno96] andbelow for a discussionof theserules),suchasin (2.78),where“if it’s

rainingthesunisn’t shining” is presupposedanddenied.

(2.77)Speaker A: Johnspeaksloudly. Speaker B: Repeatthat.

(2.50)[If awhite persondrivesthis carit’s a “classic”. If I, aMexican-American,drive it, it’s a

“low-rider”.] Thathurtsmy pride.

(2.78)[The sunis shiningalthoughit’spouringrain.] That’s a ruleBermudaalwaysbreaks.

Essentially, thereare as many abstractobjectsas thereare abstractnouns,and they can be

referredto implicitly via discoursedeixis,or explicitly with a demonstrative inferableNP [Pri81],

e.g. that fact. Therefore,in this thesiswe will extendthe term“abstractobjects”12 to cover all of

12It appearsthis termwascoinedby [Ash93].

51

Page 69: Thesis about discourse

thesepossibilities.

[Ash93] classifiestherangeof “saturated”abstractobjectsasshown in Figure2.12. Also con-

tainedwithin theseclassificationnodes,but subjectto a slightly morecomplex linguistic analysis,

are “unsaturatedabstractobjects”. This differenceis akin to [DH95]’s “abstractionoperation”:

somememberof theantecedentclausemustbeabstracted,replacedor otherwisealteredto achieve

the referenceto an unsaturatedabstractobject. In (2.79), for example,we mustalter the referent

to theform “Mary shouldgo out with anentity coindexed by thespeaker”, andin (2.80)we must

changethevoiceof theantecedentto “take thegarbageout”:

(2.79[JohnsaidthatMary shouldgo outwith him], andBill saidthat too.

(2.80)[The garbagehadto betakenout], so that’s whatBill did.

Figure2.12: [Ash93]’s Classificationof AbstractObjects

Asher’s classificationorganizesabstractobjectsalonga scaleof “concreteness”;eventualities

are the mostconcreteandproposition-like objectsare the leastconcreteabstractobjects. Remi-

niscentof Vendler, Asherarguesthateventualitiesbehave mostlike concreteNP entities;they are

locatedin spaceandtimeandcanalsobecausal(2.81);in contrast,fact-like objectsarenot located

in spaceandtime but canbe causal(2.82); finally, proposition-like objectsareneitherlocatedin

spaceor time,nor canthey becausal(2.83). In theseexamples,we usethedemonstrative inferable

NP form to make cleartheAO interpretationof thefirst clause.

(2.81)TheJetsscored.ThateventcausedthePatriots’ coachto throw down his headphone.

(2.82)TheJetsscored.ThatfactcausedthePatriots’ coachto throw down his headphone.

52

Page 70: Thesis about discourse

(2.83)TheJetsscored.*That propositioncausedthePatriots’ coachto throw down his head-

phone.

The differencebetweenreferenceto propositionsand facts is slight but interpretable:refer-

enceto propositionspredicateson a truth value,which appearsto betheonly propertythatcanbe

attributedto aproposition.In contrast,referenceto a facttakesits truth valuefor granted[Eck98].

2.6.2 A Formal Languagefor Discourse

In Section2.3.5we reviewed [Web91]’s useof referring functions,[Sto94]’s useof information

states,and[DH95]’ssetof operations,all of whichweremodelsof how abstractobjectsarecreated.

[Ash93]’s model usesa formal languageof discoursethat is an extensionof DRT [KR93].

DRT is a dynamicsemantics,analyzingmeaningin two steps.First, the DRS constructionalgo-

rithm providesa setof rulesfor theincremental(sentence-by-sentence) constructionof a semantic

representationof a discourse,calleda discourserepresentationstructure(DRS).Second,theDRT

correctnessdefinitionprovidesinstructionsfor homomorphicallyembeddinga DRSin a modelto

producethetruthconditionsfor adiscourse.Notethatwhile bottom-upDRSconstructionis largely

compositional,thereis noMontaguenotionof compositionalsemanticsatthediscourselevel,where

rulesfor semanticinterpretationcorrespondto syntacticrulesof construction.

Essentially, DRSconstructiontranslatesterminalnodesof thesyntactictreefor eachsentence

in a discourseinto a DRS,whoseunionthenforms theDRSof thediscourse,accordingto the in-

structionsfor correctness.A DRShastwo parts:auniverse,containingtherelevantdiscourseentity

references,anda conditionset,containingn-placeDRSpredicateswith discourseentity references

asarguments.In DRS(a)shown in Figure(2.13),x andy arediscourseentity references;thecondi-

tion setsaysthatx is aboy, y is Fred,andskicksy. To gettheDRS(b) for all of thediscourse,DRS

(a) servesascontext; thediscoursereferentsintroducedby thesecondsentenceareenteredinto the

conditionsetanduniversecreatedin (a). As shown DRS(b) containsoneincompletecondition(z =

?);anaphoraresolutionis required,to replace? by adiscoursereferentotherthanz. Accessibilityof

adiscoursereferentx to adiscoursereferenty is aconstraintwhichsaysroughlythatx is accessible

to y if it is in the universeof a DRS to the left, super-ordinateto, or the sameasthe universein

53

Page 71: Thesis about discourse

which y is declared.After identifying ? with y, thecorrectnessdefinitionsaysthat (b) is a proper

embeddingin a modelM iff M containsFredandMax suchthatMax kicked FredandFredcried.

As shown in (c), sententialoperatorssuchas if...thencreatesubDRSs,as do many determiners,

negation,attitudes,belief operators,nominalizations,gerunds,complementclauses,etc.

Figure2.13:SampleDRSs

Like [Sto94], Asherdoesnotusecoercionto createabstractobjectinterpretations;like [Sto94],

heassumestheseinterpretationsarealreadypresent,in this case,asadditionalvariablesin DRSs.

To achieve referenceto eventualities,Ashersimply addsto aneventvariable . (or a statevariable( ) to theDRStranslationof verbs,whichcanbeidentifiedby anaphoraresolutionwith thevariable

introducedby thediscoursedeicticanaphor. Eventsummationis thendefinedto permit reference

to complex eventuality interpretations(e.g. 2.33). To achieve referenceto propositionand fact

interpretations,Ashersimply allows the variableintroducedby a discoursedeictic anaphorto be

identifieddirectly with an accessibleDRS (roughly, if the variablesin the universeof a DRS are

accessibleto adiscoursereferentx, sois theDRSitself).

As defined,however, DRT constructionprovidesno accountof discourserelationsor discourse

segmentation(or theability of a discoursedeictic to referto a discoursesegment).To build a DRS

for adiscourseasawhole,onesimplyaddstheDRSconstructedfor eachincomingsentenceto the

DRS onealreadyhad. Asherpostulatesan additional level of discourseinterpretationto provide

a semantictheory of discoursestructure;the the resultingstructureis called a segmentedDRS

54

Page 72: Thesis about discourse

(SDRS).Essentially, anSDRSis a recursive structureof labeled(S)DRSs,with discourserelations

betweenthe labels,anda partial orderingon both the (S)DRSsandthe discourserelations. The

relationshave a truth-conditionalcontentconcerningthe relationof constituent(S)DRSsto each

other, but not all relationshave a truth-conditionalimpacton the contentof what is said,andthe

discoursedominancehierarchyrepresentedin SDRSalsohasno truth-conditionalimpacton the

contentof whatis said.

Note that while DRSsareamongthe basicconstituentsof an SDRS,andDRSsaretypically

viewedascorrespondingto sentences,Asherleavesopenthequestionof whetherconstituentDRSs

cancorrespondto a clauseor severalsentences,etc,notingthatpurposeplaysanimportantrole in

determiningindividualsegments[GS86]. EachSDRSalsocontainsadistinguishedDRS,adiscourse

topic, which summarizesthecontentof a constituentin anSDRSandbearsa particularstructural

relationto thatconstituent.

In an SDRS,every new constituent(S)DRSmustbe attachedto an antecedentlyconstructed

constituent.Only openor d-freeconstituents,however, areavailablecandidatesfor attachment.The

theorydistinguishestopic-updatingandSDRSupdatingwith an incomingDRS,via distinguished

discourserelations. Continuationand Elaborationrelationsare distinguishedas topic-based;to

attachto a constituentthey requireonly that it beopen. Thecurrent constituentis theconstituent

DRS containingthe information from the previous sentence,and is always open. Also openis

the SDRSin which the currentconstituentoccurs,andto any SDRSthat enclosesthat, etc. This

correspondsroughlyto theright frontier of thediscoursesyntactictree,but it is moregeneral,akin

to [KOOM01]’s notion of “resumptive links”, becausesubordinateconstituentsarenot treatedas

partof theattachment.

Non-topicbasedrelationsrequiretheconstituent� to which theincomingconstituentattached

to be both openandd-free, whered-freemeansroughly that � is not containedwithin an SDRS

whosetopic subsumes� .

An exampleof a simpleSDRSK is shown in Figure2.14,consistingof a (sub)SDRSK1 and

a DRS k L specifyingthe discoursetopic (indicatedby the arrow) which summarizestheseDRSs

andis in anElaborationrelationwith K1. K1 consistof DRSsk1-3 andthehierarchicaldiscourse

55

Page 73: Thesis about discourse

relationsthatholdbetweenthem.

Figure2.14:SampleSDRS

Morecomplex SDRSsresultfrom relationsbetweennon-adjacentconstituents.For example,in

(2.84),k4 is relatedby thetopicbasedContinuationrelationto k1.

(2.84)

(k1) I atea lovely dinner.

(k2) I hadsalmon.

(k3) I hadtiramisu.

(k4) ThenI wentfor awalk.

In this case,the SDRSK containsa discoursetopic k0, a subSDRSK1, andan Elaboration

relationbetweenthem. K1 containsa subSDRSK2, its a discoursetopic k1, andan Elaboration

relationbetweenthem.k1 is alsoin a Continuationrelationwith k4. K2 containsk2 andk3 anda

Continuationrelationbetweenthem. As this exampleshows, multiple discourserelationsbetween

theconstituentsof anSDRSarepossible,so long asthey areaccessible.Theprocessof inferring

discourserelationsis describedbelow.

56

Page 74: Thesis about discourse

2.6.3 Retrieving Antecedentsof Anaphora fr om the DiscourseStructure

In general,theeffect of updatinganSDRSis thatwhethera discoursereferent� maybeanaphor-

ically linked to anotherdiscoursereferentor discoursestructure� dependson whether � occurs

in a constituentthat bearsa discourserelation to � or the constituentin which � occurs. Other

constituentswill not yield potentialantecedents.

For example,in Figure2.14, this availability constraintpredictsthat � canbe resolved to j ,becausek2 is in a Continuationrelationwith k1, and i canbe resolved to � , becausek3 is in a

Continuationrelationwith k2, and k canresolve to j , becausek0 is in anElaborationrelationwith

K1. But if the text correspondingto k1 were“Kathleenis taller thanthepeoplesheworkswith”,

and the text correspondingto k3 were “She teacheswith them”, this constraintwould correctly

predictthat themcouldnot beresolvedto thepeoplesheworkswith. However, basedon examples

like (2.85),Asherrelaxesthisconstraintfor ParallelandContrastrelations.

(2.85)Johndoesnotbelieve that[Mary is treatinghim fairly]. But Fredis certainof it.

In suchcases,anaphoricreferenceis successfullymadeto anembeddedDRS,(thecomplement

of believe); thusAsherallows discoursereferentsor structuresembeddedin a constituent� that is

relatedby theserelationsto anotherconstituent� to betheantecedentsof discoursereferentsin � .

2.6.4 A Systemfor Inferring DiscourseRelations

In [HSAM93]’s approachto inferring discourserelations,discussedin Section2.2, discoursere-

lations are “proven” throughlogical abduction,wherethe total cost of a proof of eachpossible

discourserelationis determinedby thesumof thecostof abducingits premises,andthatthecheap-

estproof wins. [Lag98] notesthreeproblemswith this method. First, only onerelationwill be

chosen.Second,thecostsof differentdiscourserelationswill becomparedregardlessof possible

inconsistencies.Third, it is notclearhow thesecostsshouldbedetermined.While [HSAM93] sug-

geststhatpsycholinguisticexperimentscandeterminerelative costs,in reality theexperimentation

thatwouldberequiredto establishthesevaluesseemsno smallfeat.

DICE [LA93] is an alternative discourseinferencesystem,usedfor computingthe discourse

57

Page 75: Thesis about discourse

relationsin SDRSs.Like [Keh95], discourserelationsaredefinedin termsof constraints,or rules,

thatmustbesatisfied,andlike [HSAM93] asystemof ruleapplicationis definedUnlike [HSAM93]

however, DICE is drivenby theprocessof determininginconsistencies.Multiple discourserelations

mayhold,solongasthey areconsistentwith eachother. Moreover, establishingdiscourserelations

is representedasa consequenceof linguistic, world, andlexical knowledge,whoseinteractionis

explicitly defined,ratherthanvia theassociationof coststo premises13.

Thedefinitionsof thediscourserelationsoriginallydiscussedin [LA93] areshown in Table2.16;

their definitionsaregiven in [Lag98]’s simplified format (Narrationis roughly the Continuation

relationdefinedin theprevioussection).

Table2.16:DICE: discourserelationdefinitions

Narration �E������� ��� t Narration(��� � )Elaboration �E������� ��� J Subtype(��� � ) t Elaboration(��� � )Result �E������� ��� J cause(e  , e¡ ) t Result(��� � )Explanation �E������� ��� J cause(e¡ , e  ) t Explanation(��� � )Background �E������� ��� J overlap(e  , e¡ ) t Background(��� � )

Theserelationsaredefinedasdefeasiblerules,where t representsadefeasibleimplication(e.g.

normally, if...,then...) governinghow eachrelationis inferred,anddescribingthe knowledgethat

is neededto infer it. In eachcasethis knowledgeis definedwith respectto an updatefunction�E������� �B� : � representsa clausein thediscourse� to which�

is related,suchthat � is updatedwith�via adiscourserelationbetween� and

�.

Thedefeasiblerule of Narrationis theleastdemanding:every clause�

maybeconnectedwith

a clause� in � via Narration. The �Bk � b��CP�. predicatein theElaborationrelationrequiresthat the

eventuality(state,eventor process)in � , e  , beasubtypeof thatin�

, e¡ , suchthattheinformation

in�

extendsthat in � . In both ExplanationandResultrelations,theseeventualitiesmustbe in a

causalrelation; in the former, e¡ is the cause,and in the latter, e  is the cause. In Background

relations,theseeventualitiesmustdisplaypartialoverlapin temporalorder, andat leastonewill be

astate.13It canbearguedthatrepresentingtheseknowledgebasessuffersfrom thesamecomplexity asdeterminingtheplau-

sibility of assumptions.See[Lag98] for discussion.

58

Page 76: Thesis about discourse

In DICE, the knowledgeneededto defeasiblyassumea discourserelationmust interactwith

certain +'WU-/.C¢c.Cjq(C+ � )2. axioms,shown in Table2.17. Only thoserelationswhich areconsistentwith

theseaxiomswill be inferred. Eachaxiom expressescausalandtemporalpropertiesof eventual-

ities. £ representsa temporalorderingon eventualities;the left argumentmustprecedethe right

argument,except N e ¤£ e¡ is understoodto mean N (e ¥£ e¡ ), i.e. the left argumentmust WUT�bprecedetheright argument.NotethatCausesprecedeEffectsis anaxiomonlexical knowledge,not

discourserelations.

Table2.17:DICE: Indefeasibleaxioms

Axiom on Narration Narration(��� � ) K e ¦£ e¡Axiom on Elaboration Elaboration(��� � ) K§N e ¨£ e¡Axiom on Result Result(��� � ) K e  £ e¡Axiom on Explanation Explanation(��� � ) K§N e ¨£ e¡Axiom on Background Background(��� � ) K overlap(e  , e¡ )CausesprecedeEffects cause(e  , e¡ ) K§N e ¨£ e¡

DICE representstheworld andlexical knowledgeneededto infer discourserelationsin terms

of defeasiblerules(calledlaws). Someexamplesareshown in Table2.18.

Table2.18:DICE: Defeasiblelaws on world knowledge

PushCausalLaw �E������� �B� J fall(x, e  ) J push(y, x, e¡ ) t cause(e¡ , e  )Revolt Law �E������� �B� J revolt(x, e  ) J pacified(x,e¡ ) t©N overlap(e  , e¡ )

Theselaws neednot be storedin the lexicon; they may be derived anew at the momenttwo

clausesareutteredandneedto beassociatedwith eachother, i.e. when � is updatedwith� 14. The

ideais that if “x revolted” is uttered,followedby “x waspacified”, thedefeasibleresult,basedon

lexical knowledgeof revolt andpacify, is thatthecorrespondingeventsdid notoverlapin time. This

resultis thenavailableto therulesandaxiomsthatproducediscourserelations.

In DICE therearealsodefeasiblelaws on discourseprocesses,which specify interactionsof

temporal,causal,andlexical phenomena,andareconsideredpartof areader’s linguisticknowledge.

Someexamplesareshown in Table2.19.14See[LA93] for referencesconcerninghow thesederivationsaremodeled.

59

Page 77: Thesis about discourse

In StatesOverlap, the effect of a statein a discourseis described:if � and/or�

expressesa

state,thentheeventualitiesthey expresswill overlapin time. MaintainCausalTrajectoryconcerns

successionsof relations:if � is updatedwith ª relatedto�

, andit is known from context thate causede¡ , thene« can’t causee¡ . No Causerestrictstheuseof thediscourseconnective whenwith

respectto thedirectionof causalrelations.Read“ � when�

”, the law preventse  from beingthe

causeof e¡ .Table2.19:DICE: Defeasiblelaws on discourseprocesses

StatesOverlap �E������� �B� J (state(e  ) ¬�¬ state(e¡ )) t overlap(e  , e¡ )MaintainCausalTrajectory �E������� �B� J R(� ,

�) J cause(e  , e¡ ) t©N cause(e« , e¡ )

No Cause dxVA.�W�zE��� � |�t­N cause(e  , e¡ )DICE deductionprinciplesgoverntheinteractionbetweenthelaws andaxiomsgivenin Tables

2.16- 2.19. Theseprinciplesareshown in Table2.20. ¬¯® defineswhat canbe derived from the

defeasiblelaws andindefeasibleaxioms.Theconditions(e.g. (A1)...(A3)) mustall besatisfiedto

derive a result.

Table2.20:DICE: Deductionrules

DefeasibleModusPonens (A1) G°t²±(A2) G ¬¯®©±(A3) (not: NB± )

Complex PenguinPrinciple (B1) G¦K³±(B2) ±´t¶µ(B3) G°t­· ¬¯®¸·(B4) µaK³¹(B5) ·ºK»N�¹(B6) G

Nixon Diamond (C1) ±´t´N�µ(C2) G°t²µ Not: ¬¯®©µ (or NBµ )(C3) ±(C4) G

DefeasibleModusPonensstatesessentiallythata readermayapplyany of thelaws. For exam-

ple, if a readerreadsMax fell. Johnpushedhim, andif s/hederivesthePushCausalLaw, thenthe

60

Page 78: Thesis about discourse

conditionof that law (in this example �E������� �B� J fall( _�j`� , e  ) J push({pT>VyW , _�j`� , e¡ )) is G , and

the resultof that law (cause(e¡ , e  )) is ± , andso long asthe readerdoesnot know N G (thatMax

did not fall), s/hemaydefeasiblyderive ± (thatthereis a causalrelationbetweenthetwo clauses).

Deriving thePushCausalLaw requiresinformationfrom thelexicon to supportthecausalrelation

between“to push” and“to fall”. Furthermore,theaxiom CausesprecedeEffectsaxiom mustnot

be violatedby knowledge,for example,that the falling precededthepushing.Once ± is derived,

DefeasibleModusPonenscanbeusedagainto derive a Explanationrelationfrom thedefinitionof

Explanationrelationsgivenin Table2.16.

DefeasibleModus Ponenswill also allow the derivation of a Narrationrelation; in fact, an

Explanationrelationlogically entailsa Narrationrelation,for the conditionsfor a Narrationrela-

tion ( �E������� �B� ) area subsetof theconditionsfor anExplanationrelationsaresatisfied( �E������� �B� Jcause(e¡ , e  ) ).

However, theAxiom on NarrationrequiresMax’s falling to precedeJohn’s pushinghim, while

the Axiom on Elaboration(and the CausesprecedeEffectsaxiom) requiretheseeventsto occur

in theoppositeorder. Complex PenguinPrincipleresolvestheseinconsistencies,by statingthat if

thereareconditionsof a law or discourserelationthatlogically entailtheconditionsof anotherlaw

or discourserelation,but otherlaws of discourserelationsmake thetwo relationsinconsistent,the

rule with themostspecificconditionswins: in this case,Explanation,because(B1) is satisfiedby

theconditionson Explanationentailingtheconditionson Narration,(B2) and(B3) aresatisfiedby

theconditionson Narrationdefeasiblyimplying a Narrationrelation,andtheconditionson Expla-

nationdefeasiblyimplying an Explanationrelation,respectively. (B4) and(B5) aresatisfiedby a

theAxiom on Narrationindefeasiblyimplying oneeventordering,andtheAxiom on Explanation

indefeasiblyimplying thenegationof thateventordering. (B6) is satisfiedbecausetheconditions

for anExplanationhold,andthereforeanExplanationrelation,but notaNarrationrelation,holds.

Nixon Diamondrepresentsincoherentdiscourses,by forbiddingtheassumptionof a law or dis-

courserelationif it resultsin adirectcontradictionwith anotherassumedlaw or discourserelation.

For example,if a Resultrelationis inferredbetweenthefirst two clausesin (2.86),thentheMain-

tain CausalTrajectorylaw (C1) on discourseprocessesassumesthethird clausecannotalsobethe

61

Page 79: Thesis about discourse

causeof thesecondclause,so if an Explanationrelation(C2) is inferredbetweenthe secondand

third clauses,incoherenceresults.Assumingtheconditionsof bothrulesaresatisfied(e.g.thetrip-

ping, falling, andpushingoccurred),Nixon Diamondrequiresthatcausality(or lack of it) between

Max falling andJohnpushinghim cannotbeestablished.

(2.86)JohntrippedMax. Max fell. Johnhadpushedhim.

2.6.5 Extending the Theory to CognitiveStates

DICE consistsof amodalpropositionallanguageaugmentedwith t , defeasibleimplication,which

computesdiscourserelationsin anSDRSfrom thecompositionalsemanticsof its clauses.However,

DICE doesnothavefull accessto theformallanguageof SDRSs,only to theform of theinformation

it contains.Theconditions(topic-hoodanddiscourserelations)of anSDRSK   aretranslatedinto

predicatesof thepropositionvariable � ,where� labelsK   .

[LA99] notethat if DICE did have full accessto the languageof SDRSs,discourseinterpreta-

tion would be undecidable,becauseit computesdiscourserelationsusingreasoningaboutpartial

informationin a modalpropositionallanguage,while thelanguageof SDRSsis first order. To per-

form consistency checksoverafirst orderlanguage,DICE wouldalsohave to befirst order, andthis

wouldmake discourseinterpretationgo beyondwhatis recursively enumerable.

Thesamereasoningleadsthemto claimthatthelogic usedfor computingbeliefsandintentions

mustalsobe“shallower” thanthe logic thatmodelstheir content;becauseonelacksdirectaccess

anotherperson’s cognitive state,default reasoningis necessaryand thus consistency checksare

needed.

Like in DICE, [LA99] usea modalpropositionallanguageaugmentswith t to computecog-

nitive states. The propositionalvariablesin this languageare indexed to cognitive statesand to

discoursecontent;In DICE, the conditionsof an SDRSK   are translatedinto predicatesof the

propositionvariable � ,where � labelsK   , andin the languagefor computingcognitive states,the

propositionalvariable P   is indexed to � . Thenan interpretationin the languagefor computing

cognitive statesis j/-`_¼+#(�(�+ � )2. only if the worlds assignedto PA  arealso the onesthat make ½¦ true. Modal operatorsB(believes)andI(intends)operateover P   . B ¾C¿�À correspondsto S believing

62

Page 80: Thesis about discourse

thecontentrepresentedin theSDRS½Á  . Theauthorsassumethatif S intendsthiscontent,hedoes

notalreadybelieve it is true,i.e. I ¾ PA ÂK»N B ¾ PA  .

Defeasibleprinciplescan then accountfor links betweencognitive statesand the contentof

clausesin dialogue.For example,a principleof “Cooperativity” statesthatH will normallyadopt

S’sgoals;if notH will normallyindicatethis to S, i.e.:

I ¾ �¥t I ÃÄ�(I ¾ ��JÅN I ÃÆ� ) t I ÃÈÇ ¾ N�É�ÃÄ�As [GS86] note,thepossiblegoalsa speaker mayhave areinfinite. [LA93] notehowever that

certaintypesof speechactshave goalswhich onecancompute;for example,thegoalof a speaker

saying(representedwith :) a question(representedwith ?� ) is to know ananswer(representedas�), i.e.:

S:?�Êt ( N B ¾ � J I ¾yǺ¾ � |Similarly, Grice’s maxims15 suchas“Be Sincere”,andactionsto achieve goals,suchasPrac-

tical Syllogism(i.e. If S intends� andbelievesthat�

normally implies � thenS intends�

), can

be representedas defeasibleprinciples,enablingthe reasoningbetweencognitive statesand the

informationalcontentof clauses.

2.6.6 Summary

In this section,we presentedan alternative theoryof discoursecoherence,onewhich modelsdis-

coursecoherencewithout invoking the notion of a discoursetreestructure. The theory replaces

syntacticstructureof discoursewith a structuredsemantics.Vendler’s characterizationof abstract

objectreferencewaspresentedandextended,andits majordivisionswereshown capableof being

modeledin SDRT asreferenceto substructuresor eventvariables.Thetheoryfurthermodelsworld

knowledgeandthe inferenceof propositionalrelationsin termsof indefeasibleaxiomsanddefea-

sible rules. Interactionof the structuredsemanticsandinferencesystemis regulatedby relation-

specificformal rulesof construction,and the availability of anaphoricreferenceis dependenton

15We will discussGrice’s maximsin detail in Chapter5.

63

Page 81: Thesis about discourse

relationsandthestructuresthey create.Thetheoryfurtherextendsitself to theinferenceof cogni-

tivestates,therebymodelingthelinks betweeneverymajorcharacterizationof discoursecoherence

thatwe have sofarseenin thischapter.

2.7 Discussion

Sofarin thischapter, wehaveoverviewedhow avarietyof discoursemodelscharacterizeall or some

part of discoursecoherence,andhave summarizedwhat they do anddon’t do in relationto other

models.In this section,we discussunresolved issuespertainingto theapproachof thesemodelsto

discourseanaphoraanddiscourserelations,andpresentanew modelof a rich intermediatelevel of

discourse,which providesa meansof resolvingthe issuesposed,while leaving a numberof open

questions.In subsequentchapterswe undertake the processof answeringsomeof the remaining

openquestions.

2.7.1 Proliferation of DiscourseRelations

As [Keh95] notes,in 1748,[Hum48] discernedthreebasicconnectionsthatcanexist betweenideas:

Resemblance, Causeor Effect, and Contiguity (in time or place). Sincethen, as we have seen

in the prior sections,many alternative waysof categorizing the atomicsetof discourserelations

have beenproposed.[Lon83] and [MT88], for example,eachprovide a uniquecharacterization

of the “deep” semanticrelationsbetweenpropositionsthat underliethe surfacestructureof text.

[MT88]’s characterization,whichis alsointendedto cover presentational,or intentional,relations,

is flexible aboutthe numberandtype of possiblerelationsthat exist. In both of thesetheoriesit

is arguedthatmorpho-syntacticcuesdo not reliably indicatethesediscourserelations.[HH76], in

contrast,deriveswhatheclaimsto beacomplete,thoughdifferent,setof semanticrelationsbetween

propositionsfrom the“surfacecues”availablein a text. [Mar92] combinesthetwo approachesto

derive yet anothersetof discourserelations,by claiminga “deep” relationexists at a placein the

text if an explicit surfacecuecanbe insertedthere. Theapproachtaken by theseauthorstowards

defininga usefulsetof discourserelationsis different thanthe approachtaken by the remaining

authorssurveyed in this chapter, in that theformeraim to describethepossiblediscourserelations

64

Page 82: Thesis about discourse

thatcouldexist in atext in termsof theirown intuition. Clearly, their intuition displaysasignificant

degreeof variation.

On theonehand,as[Keh95] notes,many of thesevariationscanbeviewed asterminological

variantsof eachother. And as[Kno96] notes,someflexibility is desirable,at leastuntil aparticular

set of relationsis proven useful and possiblyeven to allow for languageevolution. Moreover,

certainrelationsmayprovemoreusefulin somedomainsthanothers,andeachsetis not in all cases

claimedto cover exactly thesamerangeof discoursecoherence,or applyto thesameconstituents,

astheothers.[Kno96] providesa usefulcomparisonof a numberof differentsets,includingthose

overviewedin thischapter.

On the other hand,as ([Keh95]) notes,the variation illustratesa commonobjection to this

“laundry-list” approachto understandingdiscoursecoherence:without an explanatorybasisfor

producingand constrainingthe productionof a particularset of atomic discourserelations,and

a characterizationof how more complex relationscan be derived from them, it is impossibleto

objectively selectoneparticularsetover any other. For example,in suchan approach,it would

beentirelypossibleto claim theexistenceof a relationdefinedexplicitly for (2.87). As ([Kno96])

definesit, sucha relationcouldbecalled: inform accidentand mention fruit .

(2.87)Johnbroke his arm. I like plums.([Kno96, 35])

Thus, while the “laundry list” approachis useful for perceiving the range of ways in which

constituentsarerelatedin coherentdiscourse,anapproachthatdescribesthemechanismsinvolved

in deriving, constraining,andcombiningdiscourserelationspotentiallyhasthe additionalbenefit

of yielding a distinctionbetweenthe differentkinds of discourserelations,andconstituents,that

shouldbeassociatedwith eachmechanism.

This is theapproachto modelingdiscoursecoherencetakenby theremainingauthorsdiscussed

in this chapter;their associationof discourserelationswith mechanismslikely explainswhy these

modelsmakeuseof asignificantlysmallernumberof discourserelationswhendescribingthemech-

anismsassociatedwith themthando thepurelydescriptive models.Nevertheless,therecontinues

to beconsiderablevariationin thenumberandtypeof discourserelationsthesediscoursemodels

defineaswell. As we saw in prior sections,[Hob90] associatesfour propositionalrelationswith

65

Page 83: Thesis about discourse

a processof discourseinference,[Keh95] reducesthis set to three,anddefinesthemin termsof

constraintswhich [HSAM93]’s inferencesystemdeterminessatisfiedor not. [SSN93]focuson the

cognitive resourcesunderlyingtheproductionof intentionalandpropositionalrelations,identifying

themwith four underlyingfeatures.[GS86] allows for an infinite numberof intentionalrelations

betweendiscoursesegments,while distinguishingonly two structuralrelations,andtwo proposi-

tional relationsbetweeninformationconveyed by segments.[Pol96] distinguishesthreestructural

relationsbetweendiscourseunits, andan unspecifiednumberof propositionalrelations. [LA93]

associatesfivepropositionalrelationswith arule-basedlogic for discourseinference,andaninfinite

numberof intentionswith arule-basedlogic for inferringcognitive states.Moregenerally, [Hov90]

surveys over350differentdiscourserelationsthathave beenproposedin theliterature.

2.7.2 Useof Linguistic CuesasSignals

The argumentagainstusingexplicit “cue phrases”assignalsof discourserelationsbetweendis-

courseunits (generallydefinedasclausesandsequencesof clauses),hastwo parts.First, a single

cuephrasemaynot unambiguouslysignala singlediscourserelation. For example,thecoordinat-

ing conjunction,and, notoriouslyplacesvery few constraintson thediscourserelationthatcanbe

derivedbetweentheclausesit connects.Second,thepresenceof thesecuephrasesis notobligatory;

discourserelationscanbederived in anappropriatecontext whentheassociatedcuephraseis not

present.[Mar92]’s insertiontestprovidesa resolutionfor thesecondargument,but not theformer.

Nevertheless,many cuephrasesaresoclearlyassociatedwith discourserelations(e.g.“as a result”K Resultrelation)that,aswe saw, mostof the mechanism-basedmodelsmake useof or at least

acknowledgethis association.For example,even thoughRST’s authorsargueagainstthe useof

cuephrasesassignalsof RSTrelations,extensionsof RST[Mar00,Mar97, SdS90]haveshown that

linguistic cuescanbeusedto manuallyandautomaticallylabelRSTrelations.

Thereis in facta wealthof literatureconcerningthedifferentwayscuephrasescanbeusedto

signalboth propositionalandintentionalrelations. [Coh84], for example,arguesthat cuephrases

can function to reducethe processingload on the hearerand facilitate recognitionof the argu-

mentstructureof a discourse.[EM90] identifiesandassociatespragmaticfeatureswith a variety

66

Page 84: Thesis about discourse

of cuephrasesandshows thatthesefeaturescanbeusedto implementtheuseof cuephrasesin an

argumentation-basedtext generationsystem.[Hov95] arguesthatcuephrasescansignalin parallel

semantic,goal-oriented,attentionalandrhetoricaldiscourserelations,all of which canpotentially

yield distinct structuralanalyses;the speaker mustselectthosewhich minimize the overall struc-

tural ambiguityfor thehearer. [Bla87] arguesthatcuephrasesindicatehow therelevance[SW86]

of onepropositionis dependenton the interpretationof another. In [Sch87]’s multi-dimensional

discoursemodel, cue phrasesindex eachutteranceto the speaker and heareralong a variety of

pragmaticandsemanticplanesin the surroundingdiscourse.[Fra88, Red90, vD79] characterize

how a varietyof cuephrasessignalpragmaticandpropositionalrelations.[MS88, LO93] provide

a detailedinvestigationof the formal semanticsof temporalcuephrases,and[Lag98] providesa

detailedinvestigationof theformal semanticsof causalcuephrases.

Building on thesesemanticinvestigations,[Kno96] developsa theoryof coherencerelations

which providesbotha solutionto theproblemthatcuephrasesaren’t reliablebecausethey canbe

ambiguous,anda solution to the problemof the proliferationof discourserelations. Knott first

definesan intuitive test to isolatea corpusof cuephrases.Humanannotatorsusethe test,which

isolatesevery phrasethatmodifiesa clauseor sentencein a naturally-occurring text, togetherwith

its host clause. If the humansjudge that the isolatedunit cannotbe interpretedwithout further

context, but canbe interpretedif the selectedphraseis removed, then the phraseis a cuephrase

(discourseconnective). Knott’s corpuscontainscoordinatingandsubordinatingconjunctions,anda

widevarietyof prepositionalandadverbialphrases,in additionto asmallnumberof relative clause

markersandotherphrasesthattake sententialcomplements.

Drawing on [Mar92]’s insertiontest,Knott thenusesa substitutiontestto organizethecorpus

into ahierarchicaltaxonomy. Thetestselectscontexts in whichacuenaturallyoccurs,andareader

decidesif thatcuecouldbereplacedwith anothercuein thosecontexts withoutchangingthemean-

ing of the discourse.By testinga variety of cuephrasesin a variety of contexts, a hierarchical

taxonomyis formed,in whichany cue � canbecharacterizedassynonymouswith, a hypernym of,

ahyponym of, exclusive with, or contingentlyinter-substitutablewith any othercue � .By investigatingtheinter-substitutability contextsanddrawing [SSN93]’scognitiveapproachto

67

Page 85: Thesis about discourse

discourserelationsandthesemanticsof cuephrasesmentionedabove, Knott defineseightbinary-

valuedcognitive featuresthatcharacterizethepossiblediscourserelationsandassociatesthelexical

semanticsof cuephraseswith valuesfor thesefeatures,by interpretingthetaxonomye.g.asfollows:

if twocuesaresynonymous,thenthey sharethesamevaluesof all features;if twocuesareexclusive,

thenfor at leastonefeaturethey take oppositevalues;if a cue � is a hypernym of � , then � shares

the valueof � in all featuresfor which � is defined,but additionallyhasa value in a featurefor

which � is notdefined,etc.Featurevaluesfor threeconnectivesareillustratedin Table2.21.

Table2.21: [Kno96]’s Featuresof DiscourseConnectives

Feature DiscourseConnectiveasa result unfortunately furthermore then

Sourceof Coherence semantic semantic pragmatic –Anchor of Relation cause resultPattern of Instantiation bilateral unilateral unilateralFocusof Polarity counterpartPolarity of Relation positive negative positive positivePresuppositionality non non nonModal Status actual actual actual –Rule Type causal causal ind?

An undefinedfeaturevalueis representedas“–”; anuncertainvalueis representedwith ablank

cell, anda“?” indicatesthatthevalueis likely but requiresfurtherstudy. The“sourceof coherence”

featurerepresentswhetherthe semanticsof the connective assertsthat the readeris intendedto

believe that the relation holds (semantic)or world knowledgealreadyindicatesthat the relation

holds(pragmatic).Theremainingfeaturesrelateto Knott’s argumentthatevery discourserelation

betweentwo discoursespansA andB correspondsto thefilling outof oneof two typesof defeasible

rules,causalor inductive (i.e. generalizations),specifiedby the“rule type” feature.Theserulesare

of the form: P1 J ... J Pn K Q. The remainingfeaturesspecifyhow theserulesarefilled out,

i.e. which span(A or B) yields P andhow, andwhich spanyields Q andhow. For example,the

“anchor” featuresrepresentsthefact thatA correspondseitherto someP (cause)andis known, or

Q (result)andis desired.The“patternof instantiation”,“focus of polarity” and“polarity” features

roughly indicatewhetherC or its negationis on the sameor oppositesideof the rule asA or its

68

Page 86: Thesis about discourse

negation. The “presuppositionality” featureindicateswhetherprior context is alsoinvolved in the

relationassomeP, andthe“modal status”featureindicateswhetherthis context is known (actual)

or not (hypothetical)by thereader.

2.7.3 Structural and Anaphoric CuePhrases

DLTAG [FMP�

01, CFM�

02, WJSK03,WKJ99, WJSK99,WJ98] proposesadiscoursemodelthat

will provide thefoundationfor thestudiesin theremainderof this thesis.TheDLTAG modelincor-

poratesmany of the insightsof themodelsandinvestigationsdiscussedabove, but it alsodisplays

somesignificantnovel insightsthat enablethe incorporationinto the modelof viable solutionsto

boththeproliferationof discourserelationsandtheambiguityof cuephrases,andtheincorporation

of solutionsto otherproblemsnotpreviously addressedby othermodels.

At thecoreof DLTAG is the insight thatdiscourseconnectivescanbemodeledaspredicates,

akin to verbsat the clauselevel, except that they can take clausesas their arguments. DLTAG

[FMP�

01] currentlymodelsthis syntaxusingthestructuresandstructure-building operationsof a

lexicalizedtree-adjoininggrammar(LTAG )[JVS99], which itself is anextensionof TAG [Jos87],

andis widely usedto modelthesyntaxof sentences.Wewill presenttheLTAG andDLTAG models

in detailin Chapter4, andasyntax-semanticinterfacefor DLTAG will bediscussed,drawing onthe

discourse-level interfacepresentedin [Gar97b] andthesentencelevel interfacepresentedin [KJ99].

Briefly, in a lexicalizedTAG, therearetwo typesof elementarytrees:initial treesthatencode

basicpredicate-argumentrelations,andauxiliary treesthatencoderecursion.Eachelementarytree

hasat leastoneanchor: the lexical item(s)with which it is associated.A lexicalizedTAG pro-

videstwo structurebuilding operationsto createcomplex trees:substitution(indicatedby Ë ) and

adjunction(indicatedby Ì ).In DLTAG, theanchorfor anelementarytreemaybea cuephraseor a featurestructurethat is

lexically null, in whichcaseaninferredrelationmayberepresentedin termsof [Kno96]’s features.

DLTAG distinguishesthreetypesof elementarytreestructures,exemplifiedin Figure2.15with cue

phrasesasanchors.As thefigure exemplifies,subordinatingconjunctions(e.g. because), coordi-

69

Page 87: Thesis about discourse

natingconjunctions(e.g. and)16, and lexically null featurestructuresaremodeledas“structural

connectives”, i.e. predicatesthat retrieve both argumentsstructurally. The semanticsassociated

with thesecuephrasesandtheir argumentscanbecomputedcompositionally.Í Í Í ÍÎÎÎÎË because Ë � � ����Ì and Ë Ï ÏÐÐasa result Ì

Figure2.15:ElementaryDLTAG Trees

Adverbials(e.g. as a result), on the otherhand,aremodeledas“anaphoricconnectives”, i.e.

predicatesthat retrieve only oneargumentstructurally, the discourseunit they modify. The other

argumentmustbe retrieved anaphorically. The semanticsassociatedwith thesecuephrasesand

their argumentscannotbe fully computeduntil the anaphoricargumentis resolved. To attachto

the growing discoursetree, they mustadjoin to the right argumentof a structuralconnective. If

no overt structuralcue phraseis present,the structuralargumentof the adverbial is attachedto

the discoursestructurevia a lexically-emptyelementarytreestructurallyidentical to the tree for

and in Figure2.15,which conveys continuationof thedescriptionof the larger treeto which it is

attached.Althougha morespecificrelationmaybeinferredandrepresentedasfeaturesin thetree,

therelationprovidedby thesyntaxaloneis semanticallyunder-specified,analogousto thesemantics

of noun-nouncompounds.

DLTAG’sdistinctionbetweenstructuralandanaphoricconnectivesis basedonconsiderationsof

computationaleconomyandbehavioral evidencesuchasfound,for example,in thecaseof multiple

connectives([WKJ99]). In (2.88),taken (in simplified form) from [CFM�

02], becauseencodesa

causalrelationbetweentwo eventualities,Q = RAISE IRE (SALLY, FRIENDS) andR = ENJOYS

(SALLY, CHEESEBURGER), andneverthelessencodesa violatedexpectationrelationbetweenR =

ENJOYS (SALLY, CHEESEBURGER) andP = SUBSCRIBES (SALLY, VEGETARIAN TIMES).

(2.88a)Sallysubscribesto VegetarianTimes.

(2.88b)Lately, she’s raisedtheire of herveganfriends

(2.88c)becausesheneverthelessenjoys theoccasionalbaconcheeseburger.

16This view will befurtherqualifiedin Chapter4.

70

Page 88: Thesis about discourse

To modelbothpredicatesstructurallywouldcreateadirectedacyclic graph,whichgoesbeyond

thecomputationalpowerof LTAG andcreatesacompletelyunconstrainedmodelof discoursestruc-

ture [WJSK03]. However, preliminary investigationsinto the behavior of cuephrasesreveal that

while subordinatingandcoordinatingconjunctionsseemto alwaysandonly takeadjacentdiscourse

segmentsastheir arguments,adverbial cuephrasesseemto sharemany propertieswith anaphora.

For example,their left argumentmay be found intra- or inter-sententially, asshown in (2.89)and

(2.90),respectively. Moreover, their left argumentcanbeaninferredsituation,asshown in (2.91)

(thesituationin which theaddresseedoesnot wantanapple),andit canevenbederived from the

interpretationof a discourserelationbetweentwo segments,asshown in (2.92)(theresultrelation

impartedby so).

(2.89)A personwhoseeksadventuremight, for example,try skydiving. [[WJSK03]]

(2.90)Somepeopleseekadventure.For example,they might try skydiving. [[CFM�

02]]

(2.91)Do youwantanapple?Otherwise,youcanhave apear. [[WJSK03]]

(2.92)Johnjustbroke his arm.So,for example,hecan’t cycle to work now. [[WJSK03]]

DLTAG doesnot claim to provide a completemodelof discourse;it is not committedto the

view of theentirediscourseasasingletree,treestructuresmayonly bebuilt within segments(akin

perhapsto [KOOM01]), andmultiple treesmay be possible,for exampleat an informationaland

intentionallevel, as[MP92] have shown. While DLTAG is not committedto a particularview of

discoursestructure,it is very committedto the idea that a rich intermediatelevel betweenhigh-

level discoursestructureandclausestructure,namely, thesyntaxandsemanticsassociatedwith cue

phrases,mustbe specifiedand recoverablein order to interpreta discourse.Moreover, because

cuephrasesaresomeof theclearestindicatorsof discoursestructure,andtheir argumentscanbe

reliably annotated[CFM�

02], large-scaleannotationstudieswill provide information about the

rangeof possiblediscoursestructures.

2.7.4 Comparison of DLTAG and Other Models

Unlike thediscoursemodelsdiscussedin prior sections,DLTAG doesnot claim to provide a com-

pletemodelof discourse.Rather, it proposesanintermediatelevel of discoursestructureandinter-

71

Page 89: Thesis about discourse

pretationthatcanbebuilt directlyon topof theclausestructureandinterpretation.BecauseDLTAG

builds a treestructurefor discourse,it is similar in computationalpower to theothertree-baseddis-

coursegrammarsdiscussedin this chapter. But althoughDLTAG, like many othermodels,argues

that(anintermediatelevel of) discoursecanbemodeledin termsof syntaxandsemantics,DLTAG

alonedefinesthis discoursestructureandinterpretationin termsof thesamemechanismsthatare

alreadyusedat the sentencelevel. Moreover, DLTAG alonearguesthat not all relationscanbe

modeledstructurallyin a tree-baseddiscoursemodel;DLTAG decouplesanaphoricandstructural

connectionsbetweendiscoursesegments,andthusonly theDLTAG treeis ableto modelrelations

betweenadiscoursesegmentandmultiple prior segmentswithin asingletree.

Othertree-basedmodelsalsointroduceunnecessaryredundancy in termsof theadditionalmech-

anismsthey proposeto build discourse.Thesemodelsmake useof a predefinedsetof discourse

relations,andcuephrasesaretreatedas“signals” of thesepredefinedrelations. The dependency

of theserelationdefinitionson thepresenceor inferenceof a cuephraseis clearlyvisible, suchas

the “otherwise” relationproposedin RST[MT88]. Numerousotherexamplesof this arefound in

[Mar97]’s instructionsfor manuallylabelingRSTrelations.However, if, asin DLTAG, thesyntax

andsemanticsof cuephrasesare taken into account,then it is redundantto postulateadditional

relationdefinitionsandgrammaror semanticrulesto creatediscourserelations.Moreover, theRST

manualalsoillustratescaseswherediscourserelationsbetweendiscoursesegmentsareredundant

even in the absenceof a cuephrase.As illustration, in (2.93)-(2.95)(from [Mar97]), a discourse

deictic(italicized)in (b) refersto (a),andits predication(bold-faced)is synonymouswith theRST

relation(capitalized)between(a) and(b) (orderis representedby W (nucleus)and ( (satellite)).

s INTERPRETATION n

(2.93a) All evidencepointsto thefactthatKennedywasassassinatedby theCIA.

(2.93b) Thissuggeststo me thattheorganizationis untrustworthy.

n EXPLANATION-ARGUMENTATIVE s

(2.94a) Mostof thedinosaursdiedabout65,000,000yearsago.

(2.94b) Someresearchersassumethattheimpactof abig meteoritecausedthis.

72

Page 90: Thesis about discourse

n EVALUATION s

(2.95 a) Featureslike our uniquelysealedjacket andprotective hub ring make our discslast

longer. And oursoft innerliner cleanstheultra-smoothdiscsurfacewhile in use.

(2.95b) It all addsup to betterperformanceandreliability.

[KOOM01] suggestthat discoursedeixis useis not a discourserelation,but rathera form of

“resumption”. However, their dataonly containscasesconcerningthe RST elaboration object-

attrib ute relation. In thefollowing examplesfrom [Mar97], we find similar redundancy, although

thereis no discoursedeixis, but rathera nominal referencein (a) to a noun in (b). Again, the

predicationand/orreferenceis synonymouswith thelabeledRSTrelation.

n/sCAUSE-RESULT n/s17

(2.96a) Unfortunatelyfor the athlete,the anaerobicmetabolismof carbohydratescanyield a

buildup of lactic acid,whichaccumulatesin themuscleswithin two minutes.

(2.96b) Lacticacid andassociatedhydrogenionscauseburningmusclepain.

n INTERPRETATION s

(2.97a) Steepdeclinesin capitalspendingcommitmentsandbuilding permitspushedthelead-

ing compositedown for thefifth time thismonth.

(2.97b) Such a declineis unusualat this stagein anexpansion.

n EVALUATION s

(2.98a) Policy makershave four options...

(2.98b) Thelastof theseis ultimately the only sustainableoption.

ELAB n PROC-STEPs

(2.99a) A usershouldinvoke theprogramwith thenameof thefile andthenameof thefile to

becreated.17[Mar97, 24] statesthatthewriter’s intentionsareunclearasto whetherthiscauseor this resultis thenucleus.

73

Page 91: Thesis about discourse

(2.99b) Theprocessopenstheexistingfile, createsthenew file, andcreatesachild process.

WhattheRSTrelationsin theseexamplesdo is simply restatetheclause-level semantics,rather

thandescribenew informationaboutthelink betweentheclausesthat is not alreadyrepresentedin

thesemanticsof theclauseitself. Becausemostof themodelsdiscussedabove do not incorporate

clauselevel semanticsinto their discourseinterpretation,they are likely to postulatewhat is in

facta redundant“discourserelation” in suchcases.But becauseDLTAG builds discoursestructure

directly on top of clausestructure,this semanticsis available to the discourseinterpretation;in

DLTAG the “relations” between(a) and(b) in theabove caseswould be representedby an empty

connective, signaling“continuation”of thediscourse.By retainingthesemanticinterpretationsof

theclausesandtheinformationabouthow thereferentiallinks betweenthemareresolved,theneed

to supplyanadditionalRST-typerelationis removed.

Themodelsof inferencediscussedearlierin this chapteralreadymake someuseof theseman-

tic contributionsof somecuephrases.[HSAM93], for example,mentionusingthe propositional

contentof because, and[GS86]mentionusingthepropositionalcontentof but. SDRT [LA93] also

incorporatesthesemanticcontributionsof somecuephrasesinto their structuredsemanticsandin-

ferencesystem,anddoesallow multiple relationsbetweendiscoursesegments,but its useof only

structuralattachmentsandpredefineddiscourserelationsagainproducesredundancy. For example,

theassociationof a structural“result relation”with anadverbialcuephrasesuchasasa resultwill

be shown in Chapter3 to be redundant.In fact, aswill becomeapparentafter the discussionin

Chapter3, SDRT canalreadyhandlethesemanticsof adverbial cuephrasesby extendingto them

thesemanticsthey employ for discoursedeixis.Moregenerally, anincorporationof thesemanticsof

all discourseconnectiveswould likely reducethecomplexity of inferencesystems;if thesemantics

of a connective canasserta relationbetweenclauses,this relationno longerneedsto beinferred.

[Kno96]’s proposalto decomposediscourserelationsinto featuresthat are attributed to the

semanticsof cuephrasesis model-independentandthuscanbeincorporatedinto DLTAG.However,

thereare a numberof problemsmainly resultingfrom his useof an intuitive test to isolatecue

phraseswhich must first be resolved. For example,Knott’s list includeslexical items from all

five syntacticcategoriesoriginally notedby [Qui72] ascontainingcuephrases:coordinatingand

74

Page 92: Thesis about discourse

subordinatingconjunctions,adverb and prepositionalphrases,and phraseswhich take sentential

complements(e.g. it follows that, all told). If the interpretationof verbsand small clausesare

availableto thediscoursemodel,asis thecasein DLTAG, thenattributing themadditionalfeatures

of adiscourserelationbecomesredundant;any presuppositionsof theseverbs(e.g.thatif x follows,

then � follows something) will alreadybe available at clauselevel. Similarly, Knott’s intuitive

testoverlooksthe fact that the semanticsof complex cuephrasescanbe treatedcompositionally.

For example,his list includescomplex relative clausecomplementizers(e.g. especiallywhich), as

well ascomplex subordinatingconjunctions(e.g. especiallyif/when,even if/when,only if/when,

etc). Adverbssuchasespecially, even,only canattachto many cuephrases;listing eachresulting

combinationindividually is unnecessaryif thecompositionalsemanticsof cuephrasesis takeninto

account.Moreover, like verbsandsmall clauses,if thesemanticsof complementizersis available

to thediscoursemodel,thenattributing themfeaturesof adiscourserelationis redundant.

Moregenerally, Knott’s intuitive testenablesothermechanismsof discoursecoherence,suchas

inference,implicature,andintonation,to be conflatedwith the semanticsof cuephrases,causing

errorsof comissionand omission. For example,Knott erroneouslyincludesunfortunately, and

surprisingly, but not unhappilyor not surprisingly. Moreover, investigationof naturallyoccurring

casesof unfortunatelyrevealsthatsomefeaturevaluesareincorrect.Inclusionof unfortunatelyas

a cuephrasewould requireit to beundefinedfor themajority of features,akin to and. While and

is clearlyuninterpretablein isolationwith its hostclause,unfortunatelyis not. In addition,Knott’s

featurescannotfully describethe idiosyncraticmeaningof eachconnective. [JR98], for example,

arguethatthefeaturesincorrectlydescribetheuseof donc(similar to therefore) in French,because

furthersemanticpropertiesof theadverbmustbetakeninto account.Finally, it is notclearwhether

Knott intendsto associateinferreddefeasibleruleswith all cuephrases;althoughthevalueof the

‘rule type’ featureis left blank for adverbial temporalconnectives(e.g. then,next), they neednot

bedefinedin relationto featurevaluesthatfill out thesemanticsof defeasiblerules.

In essence,DLTAG arguesthat it is thesemanticsof cuephraseswhich drivestheconstruction

of a discoursemeaning. Associatingcuephraseswith semanticssolves both the proliferationof

discourserelationsandtheambiguityof cuephrasesas“signals” of theserelations.It is no longer

75

Page 93: Thesis about discourse

necessaryto “map” the useof cuephrasesto separatelydefineddiscourserelations;instead,cue

phrasessupplyrelationsin their semantics,andlike all predicates,themeaningof therelationthey

supply canbe vagueor preciserelative to other predicates,and it can also be over-loaded. For

example,the verb go at the clauselevel hasa variety of meanings,mostof which canbe stated

morepreciselyby otheractionverbs.And becauseDLTAG views inferenceasanothermechanism

separatefrom compositionalsemantics,additionalrelationscanbe inferredwhetheror not a cue

phraseis used. Incorporatingthe intermediatemoduleDLTAG proposesinto othermoduleswill

yield amorecomputationallyeconomicalandmoreobservationallyvalidcompletediscoursemodel.

2.7.5 RemainingQuestions

Becausethesemanticsof cuephrasesdrivestheconstructionof discoursemeaningin DLTAG,each

cuephrasemustbeassociatedwith asemantics.While, asSection2.7.2indicates,thesemanticsof

subordinatingandcoordinatingconjunctionshasbeeninvestigatedin greatdetailbothat theclause

anddiscourselevel, thesemanticsof mostadverbial cuephrases,consistinglargely of adverband

prepositionalcuephrases,have beenmainly ignoredat the clauselevel, andhave received much

lessattentionat thediscourselevel. Wewill seeclearevidenceof this in Chapter3.

In fact,partof this “attentioninequality” is simply dueto thefactthatthesetsof subordinating

andcoordinatingconjunctionsarerelatively small,while thesetof adverbialsis a largeset.In fact,

as[Kno96] notes,the setof adverbialsis compositional,andthereforeinfinite. For example,the

adverbgenerally canbemodifiedby innumerablymany instancesof very (e.gverygenerally, very

verygenerally,...), eachtime producinga uniquememberof this set. It is not surprising,therefore,

thatwhile DLTAG proposesthat certainadverbialsfunctionasdiscourseconnectives,they do not

isolatethis subsetfrom thesetof all adverbials. Becauseit is not possibleto answerthequestion

of which adverbialswhich functionasdiscourseconnectiveswith a list, onemustaskinsteadwhat

mechanismscauseanadverbial to functionasadiscourseconnective.

In Chapter3 we will uselinguistic theory to investigatethe semanticmechanismsthat cause

an adverbial to function asa discourseconnective. This investigationwill alsoshedlight on the

questionof whatdiscourseunitsanadverbialdiscourseconnective relates.For while thediscourse

76

Page 94: Thesis about discourse

modelsoverviewed in this chaptermake frequentuseof eventualityinterpretationsof clauses,and

frequentuseof theterm“discoursesegment”,theseinterpretationsandconstituentsarein mostcases

not well definedor distinguishedfrom otherinterpretationsor constituents.We will usediscourse

deixis researchto betterclarify both the semanticnatureof the argumentsof adverbial discourse

connectives andthe syntacticconstituentsfrom which they canbe drawn. Our investigationwill

alsoshedlight on thespaceof discourserelationsimpartedby awidevarietyof adverbialdiscourse

connectives,andenableaprecisesemanticrepresentationof theirbehavioral anaphoricity. In Chap-

ter 4 we theninvestigatehow this semanticscanbe incorporatedinto a syntax-semanticinterface

for the DLTAG model. In Chapter5, we investigateotherwaysin which the interpretationof an

adverbialcancontributeto discoursecoherence.

2.8 Conclusion

Thediscoursemodelspresentedin this chaptergenerallyagreethatdiscoursehasa recursive struc-

ture andthat this structureaffects the interpretationof discourse.We have seena numberof ef-

forts to formalizetheseinsights,including descriptive, inference-based,syntacticand/orsemantic

approachesto modelingdiscourserelationsand anaphoricconstraints. We have arguedthat the

DLTAG approachshouldbeincorporatedinto thesemodels,therebyremoving theneedto selecta

singlesetof “primiti ve” relationsunderlyingall coherenttext spans,anda singlemechanismfor

producingthem. In DLTAG’s view, thesyntaxandsemanticsof cuephrasesprovidesoneway of

producingcoherencebetweendiscourseunits,anddiscourseinferenceprovidesanother. By under-

standinghow differentmodulesinteractwith eachotherandwith othercharacteristicsof discourse

to producecoherence,we canthenbegin to understandhow completeandcoherentdiscourseinter-

pretationsareproduced.

77

Page 95: Thesis about discourse

Chapter 3

SemanticMechanismsin Adverbials

3.1 Intr oduction

In Chapter2, we describedsimilaritiesanddifferencesbetweena variety of modelsof discourse

coherence,which, takentogether, distinguishdifferentmodulesrequiredto build a completeinter-

pretationof discourse.We introducedDLTAG ([FMP�

01, CFM�

02,WJSK03,WKJ99,WJSK99,

WJ98]) asa theorythatbridgesthegapbetweenclauseanddiscourseinterpretations,by usingthe

samesyntacticandsemanticmechanismsthatbuild theclauseinterpretationto build anintermediate

level of discourseinterpretationon topof theclauseinterpretation.

In DLTAG, cuephrases, or discourseconnectives, arepredicates,like verbs,except they can

take interpretationsof clausesas arguments. For coordinatingand subordinatingconjunctions,

both argumentscomestructurally. For adverbial cuephrases,which aremainly adverb (ADVP)

andprepositional(PP)phrases,only oneargumentcomesstructurally. Basedon considerationof

computationaleconomyandbehavioral evidence,DLTAG arguesthat theotherargumentof these

adverbialsmustbe resolved anaphorically. However, while DLTAG proposesthat certainadver-

bialsfunctionasdiscourseconnectives,it doesnot isolatethis subsetfrom thesetof all adverbials.

Becausethesetof adverbialsis compositional,andthereforeinfinite ([Kno96]), it is not possible

list theadverbialsthatfunctionasdiscourseconnectives.In thischapter, wepresentacorpus-based

investigationof the semanticmechanismsthat causecertainadverbials,which we call discourse

78

Page 96: Thesis about discourse

adverbials,to functionasdiscourseconnectives,while otheradverbials,which we call clausalad-

verbials,do not functionasdiscourseconnectives.

In Section3.2we review thefunctionandstructureof adverbialsanddescribehow theseprop-

ertieswereusedto extracta datasetfrom a parsedcorpora.In Sections3.3-3.4we overview major

syntacticandsemanticissuesthathave beenaddressedin clause-level analysesof adverbials,and

introduceanextensionto theseanalysesthat incorporatesthediscoursedeixis researchintroduced

in Chapter2 anddefinesthe propertiesof semanticobjectswhoseanaphoricitycancauseadver-

bials to function as discourseconnectives. Following DLTAG, in this extension,anaphoricityis

definedastheuseof “discourseconnecting”devicessuchasanaphoricreferenceandpresupposi-

tion by adverb andprepositionalforms to retrieve objectsat the discourselevel, just as they are

employedfor theretrieval of objectsat theclauselevel. In Sections3.5-3.6,we incorporateinto this

extensiondiscourse-level analysesthathave alreadybeenproposedfor a smallvarietyof adverbial

discourseconnectivesandpresenta rangeof semanticobjectsthatcomposeadverbialsanda range

of anaphoricdevices that determineif andhow theseobjectsrelateto the surroundingdiscourse.

Weconcludein Section3.7.

3.2 Linguistic Background and Data Collection

This sectionprovidesthe linguistic backgrounduponwhich thestudiesin subsequentsectionsare

built. We first discusswaysin which adverbial functiondiffers from otherfunctionsof ADVP and

PPandreview thegrammaticalstructureof ADVP andPPadverbials.We thendescribehow these

propertieswereusedto extractthedatasetstudiedin this thesisfrom acorpusof naturallanguage.

3.2.1 Function of Adverbials

Adverbialsareadjuncts, elementswhosepresencein a clauseis not obligatoryfor clauseinterpre-

tation. They functionasmodifiers, elementsthat supplyadditionalinformationabouttheelement

they modify. Adverbialsfrequentlymodify verbphrases(VP) or sentences(S).

The term adverbial denotesa functional ratherthan a syntacticcategory. As exemplified in

(3.1)-(3.8),avarietyof syntacticcategoriescanfunctionasadverbials:in additionto ADVP andPP

79

Page 97: Thesis about discourse

(3.1)-(3.2),which arethefocusof this thesis,finite clauses(3.3),non-finiteclauses(e.g. infinitives

(3.4),-ing participles(3.5),-edparticiples(3.6)),verb-lessclauses(3.7),andnounphrases(3.8)can

alsofunctionasadverbials([Ale97]).

(3.1)Hopefully, therewill never beanotherworld war.

(3.2) In mostsituations, Johnremainscalm.

(3.3)Johnworkedlatealthoughhewasverytired.

(3.4)Johnplaysto win.

(3.5)Reading, Johnrelaxes.

(3.6)Urgedbyhis mother, Johndid thedishes.

(3.7)Johnraninto thestreet,unaware of thedanger.

(3.8)JohncameTuesday.

ADVP andPPdo not alwaysfunctionasadverbials;for example,ditransitive verbs,asshown

in (3.9), aswell asverbsof behavior, movement,andsituation,asshown in (3.10) - (3.12),may

lexically sub-categorizefor anADVP or PP([MG82]).

(3.9)He gavea carto Mary.

(3.10)He behavedawfully/inan unexpectedway.

(3.11)He residesnearby/atmyhouse.

(3.12)He dresseswell/in slacks.

Moreover, asshown in (3.13)-(3.18),someADVP andPP(italics)canbeusedto modify phrasal

categories(bold-face)otherthanVP (3.13)andS (3.14), includingnounphrases(NP) (3.15),ad-

jective phrases(AP) (3.16),andotherPPs(3.17)andADVPs(3.18)([ODA93]).

(3.13)They workedquickly/in a frenzy.

(3.14)Probably/Inall likelihood, he will survive.

(3.15)Evendogsin captivityeatbones.

(3.16)Sheis completelycrazy abouthim.

(3.17)He sentflowersright/over to his enemies.

(3.18)At leastonce, hefell veryseriously in love.

80

Page 98: Thesis about discourse

It is alsoimportantto notethatADVP andPPdon’t alwaysadjointo theelementbeingmodified.

Althougheveryadverbialmaynotbelicensedin everyposition,four positionswithin theclauseare

available to S- and VP-modifying ADVP and PP adverbials ([Ale97]), as shown in (3.19). As

[Ale97] discussesin moredetail,eachof thesepositionsmayalsohave aparentheticalcounterpart,

in which theadverbialis setoff by commas(or pauses)from therestof theclause.Again,however,

not all adverbialsarelicensedin every parentheticalposition([Ale97]). As discussedbelow andin

subsequentchapters,our dataset includesboth S-initial adverbialsthat arecomma-delimitedand

S-initial adverbialsthatarenotcomma-delimited,andwe do notdistinguishthemsemantically1.

(3.19)

S-initial: Of courseJohnwashurt.

S-medial(beforeauxiliary): Johndefinitelywashurt.

S-medial(afterauxiliary, beforemainverb): Johnwasprobablyhurt.

S-final(aftermainverb): Johnwashurtslightly.

3.2.2 Structure of PP and ADVP

Generallyin linguistics,thegrammaticalstructureof aphraseis representedwith aphrasestructure

rule2. Sucha rule is shown in (3.20),whereXP representsa phrasalcategory, andX representsthe

minimalelement,or head, aroundwhich therestof thephraseis built. Thearrow readsas“consists

of (in theordershown)”, andparenthesesindicateoptionalelements:SPECabbreviatesspecifier,

which is definedasan additionalphraseusedto make themeaningof theheadmoreprecise,and

COMP abbreviatescomplement, (or internal argument), which is definedasan additionalphrase

usedto supplyinformationthatis alreadyimpliedby themeaningof thehead([ODA93]).

(3.20)XP K (SPEC)X (COMP)

Prepositions(P) in English,which correspondthe headof a PP, area closedclassof lexical

items.Someexamplesof prepositionsareshown in (3.21).

(3.21)about,after, as,at,by, before,down, for, from, in, of, on,over, since,until,...

1See[FMPÑ 01] for furtherdiscussionof DLTAG’sextractionof S-internaladverbialdiscourseconnectives.2Seee.g.[ODA93] for furtherdetails

81

Page 99: Thesis about discourse

Adverbs(ADV), which correspondtheheadof a ADVP, aregenerallyclassifiedinto two mor-

phologicaltypes:non-derived,or unsuffixed,andderived, includingsuffixedandcompoundforms

([Ale97]). In English,the-ly suffix predominates3; it canbeaffixedto mostadjectivesandto some

nouns(in the lattercasethe resultingform canoftenbeusedasbothanadverb andanadjective).

Thereareotheradverb suffixesaswell, including -wiseand-ally (the latter is usedon adjectives

endingin -ic). In addition,theprefix a-, which maybehistoricallyderivedfrom theprepositionon

([Suz97]), canform adjectivesandsomeadverbsfrom nouns.Examplesareshown in Table3.1.

Table3.1: Non-DerivedandDerivedAdverbs

Mor phological Type Examples

non-derived often,well, todaycompound-derived therefore,however, neverthelessadjective+ -ly briefly, fortunately, accusingly, swimmingly, *atomiclynoun+ -ly yearly, monthly, purposely, partly, kingly, *arrowlyadjective+ -wise likewise,otherwise-ic adjective+ -ally specifically, atomicallya- + noun ahead,apart

Perhapsbecausemany ADVP arecomposedof asingleadverb,thestructureof ADVP is notas

well studiedin linguisticsasPP([ODA93]). [CL93] consideradverbsa ‘minor’ lexical category,

wherethe four major lexical categoriesaredefinedby the featuresystemin (3.22). It is assumed

thatsubsidiaryfeatureswill distinguishadverbsandadjectives([Ale97]).

(3.22)[+N, -V] = noun [+N, +V] = adjective

[-N, +V] = verb [-N, -V] = preposition

Nevertheless,specifiersarecommonlyfoundin bothPPandADVP. Thesespecifiersaregener-

ally adverbs,asshown (italicized)in (3.23)-(3.25).In (3.25),theADVP specifieris obligatory.

(3.23)Themailmanis [ ÒcÒ almost/barely at thedoor].

(3.24)[ Ó$Ô Õ�Ò Very/Quitefrequently],I go to themovies.

(3.25)[ Ó$Ô Õ�Ò Longago],theearthwasformed.

3[Ale97] further notesthat in Greek-a and-ospredominate,in German-weisepredominates,in French-mentpre-dominates,andin Italian -mentepredominates.

82

Page 100: Thesis about discourse

Frequently, prepositionstake a complement.Nounphrase(NP) complementsarecommon,as

italicizedin (3.26). PPandS complements(thelattercreatesa subordinatingconjunction)arealso

found,shown in (3.27)-(3.28),but in somecontexts a prepositionmaytake no complement,shown

in (3.29).A few adverbsalsotake complements,asshown in (3.30)([Ale97]).

(3.26)Mary hikes[ ÒcÒ in theGreenMountains].

(3.27)Thekidsare[ ÒcÒ down in thecellar].

(3.28)Herparentsarrived[ ÒcÒ aftersheleft].

(3.29)Herparentscame[ ÒcÒ over G ].

(3.30)Johnsucceeded[ Ó$Ô Õ�Ò independentlyfromour efforts].

While phrasestructurerulesrepresenttheinternalcompositionof anXP, tree-basedgrammars,

suchas[JVS99, Gro99], alsorepresentanXP’sexternalarguments,e.g.thephrasesto whichanXP

canadjoin,alongwith thestructurethatresults.XTAG treesfor anS-adjoiningPPwith aninternal

NPargumentandanS-adjoiningADVP with no internalargumentareshown respectively in Figure

3.1,where( Ì ) indicatesadjunction.

S� � ����PPÖ Ö××

P

As

NP

a result

SØ S� � ����ADVP

ADV

Consequently

SØFigure3.1: S-AdjoiningPPandADVP

3.2.3 Data Collection

Differencesbetweendiscourseadverbialsandclausaladverbialscannotbe attributedto their syn-

tacticstructure;asshown in Figure3.2,bothcanadjointo anSnode,andtheresultis anS.

Nevertheless,we canusetheir commonsyntaxto extract thedatafor study. Becausethesetof

adverbialsinfinite, it is not possibleto studythemall; thegoalof this datacollectionwasto gather

a largeandrepresentative setof thedifferentADVP andPPadverbialsthatwill commonlyappear

83

Page 101: Thesis about discourse

in any Englishcorpora,with theexpectationthattheanalysiscanbeextrapolatedto novel tokens.

S� � � � ������PPÖ Ö××

P

As

NP

a result

SÍ Í ÍÎÎÎNP

people

VP

areself-centered

S� � � �����PPÙ ÙÚÚ

P

In

NP

spring

S� � ����NP

thelilacs

VP

bloom

Figure3.2: S-AdjoinedDiscourseandClausalAdverbials

Thedatastudiedin this thesisconsistsof the(correctlyannotated)S-initial, S-adjoinedADVP

andPPthat appearat leastoncein the PennTreebankI versionsof the parsedWSJandBrown

corpora[PT]. Thesecorporawerechosenbecausethey representa wide varietyof texts, including

news articles,essays,fiction, etc. In thePennTreebankI POS-taggingandbracketing system,S-

adjoinedADVP andPParebracketedassiblingsof theS they modify. In Figure3.3, for example,

thenis thetargetedADVP, andin fact is thetargetedPP.

Figure3.3: S-AdjoinedADVP andPPAdverbialsin PennTreebankI

Althoughasdiscussedabove,mostadverbialscanbefoundin avarietyof positions,S-adjoined

ADVP andPPwerechosenfor studyundertheassumptionthatthis is the“default” syntacticposi-

tion for mostadverbialdiscourseconnectivesandthat themajority of adverbial cuephrasesfound

84

Page 102: Thesis about discourse

in otherpositionswill alsobe found S-adjoined4. The positionalvariability of adverbialsmeans

however that thecountsof extractedS-adjoinedadverbialsdoesnot reflectthecountsof thesead-

verbialsappearingin otherpositions.Moreover, becausetheunderlyingpositionof left-dislocated

adverbialsis not representedin PennTreebankI parses,thecountsalsoincludesomeleft-dislocated

nonS-modifiers.

The datawascollectedusing tgrep5, which is the UNIX commandgrep modifiedfor useon

syntacticparses.With tgrep,theuserspecifiesapatternusingnodenamesandrelationshipsbetween

nodes. The patternis thenmatchedagainsta corpusof syntacticparsesand thoseparseswhich

matchareextracted.Thepatternsusedto extractbothADVP andPPareshown in thefirst column

(topandbottomsections)of Table3.2; they areidenticalexceptfor theextractedelement.

Table3.2: tgrep Resultsfor S-AdjoinedADVP andPPin WSJandBrown Corpora

TGREP Pattern WSJ Tokens Brown Tokens

TOP Û (S Û (ADVP $. S)) 71 1604TOP Û (S Û (ADVP $. (/ Ü , / $. S))) 460 1388TOP Û (S Û (ADVP $. (/ Ü “/ $. S))) 0 1TOP Û (S Û (ADVP $. (/ Ü , / $. (/ Ü “/ $. S)))) 2 2TotalADVP Tokens 533 2995

TGREP Pattern WSJ Tokens Brown Tokens

TOP Û (S Û (PP$. S)) 372 1801TOP Û (S Û (PP$. (/ Ü , / $. S))) 4970 3135TOP Û (S Û (PP$. (/ Ü “/ $. S))) 0 1TOP Û (S Û (PP$. (/ Ü , / $. (/ Ü “/ $. S)))) 6 10TotalPPTokens 5348 4947

TherelevantrelationshipsbetweennodesareA Û B, meaningA immediatelydominatesB, and

A $. B, meaningA andB aresiblingsandA immediatelyprecedesB. TheTOPnoderestrictsthe

searchto mainclauses.Regularexpressionsareindicatedby surroundingthenodenamein slashes

(/). The caret(Ü ) anchorsthe regular expressionat the beginning of a word. Regular expressions

wereusedto extractcaseswherepunctuationintervenedbetweentheadverbialandthesisterS.

4Adverbialcuephrasesin otherpositionsmayserve information-structuringpurposes[FMPÑ 01]. Positionalvariationwill bediscussedfurtherbelow.

5See[PT] for documentationof tgrep, thePennTreebankI systemof POS-taggingandbracketing,andtheWSJandBrown corpora.

85

Page 103: Thesis about discourse

The secondandthird columnsin Table3.2 shows the countsof adverbialsretrieved by each

patternin theWSJandBrown corpora.As shown, a quotationmarkrarely intervenedbetweenan

adverbialandtheir siblingS,but acommafrequentlyintervened.

Total countsof S-adjoinedadverbialstokensareshown in the first columnof Table3.3. Perl

scriptswerewritten to obtaintotal countsof eachadverbial typethatoccurredoverall, asshown in

thesecondcolumnof Table3.3.

Table3.3: TotalS-AdjoinedAdverbialsin WSJandBrown Corpora

Adverbial Category Total Tokens Total Types

S-adjoinedADVP 3528 849S-adjoinedPP 10295 7424S-adjoinedADVP andPP 13823 8273

Becauseit was taggedby humanannotators,however, the PennTreebankI-taggedWSJand

Brown corporacontainerrors. For thepurposesof this thesis,themoresignificanttypesof errors

werethefollowing6:H Incorrecttagging. For example,numerousADVP areincorrectlytaggedasPPandviceversa.H Incompletebracketing. For example,numerousADVP andPPareincorrectlytaggedasRB

(adverb)andIN (preposition),ratherthanasfull phrasalcategories.H Incorrect bracketing. For example,numerousADVP andPParebracketedasimmediately

dominatingtheadjacentS (insteadof their correctsibling relationship).

Incorrecttaggingerrorswerecorrectedby hand.While in principleit wouldhavebeenpossible

to incorporateincompleteandincorrectbracketingerrorsinto thetgrepextraction,doingsowould

have introduceda large amountof extraneousmaterial. As the goal of the datacollectionwasto

obtaina largeandrepresentative set,sucherrorswerenot incorporated.

6JuliaHockenmaier(personalcommunication)hascorrectedmany annotationerrorsin PennTreebankfor thepurposeof building aCCGgrammar([Ste96]);however dueto thepropertiesof CCGit wasnotnecessaryfor herto correcterrorsin theannotationof S-adjuncts.

86

Page 104: Thesis about discourse

3.2.4 Summary

In thissectionwehaveoverviewedthestructureandfunctionof adverbials,andexplainedtheuseof

thesepropertiesto extractadatasetof ADVP andPPS-modifiersfor studyin subsequentsections.

3.3 Adverbial Modification Types

With theexceptionof the lexical semanticsproposedin [Kno96] thatwasdiscussedin Chapter2,

thesemanticsof mostdiscourseadverbialshasnotbeenwell studied.Mostpost-generative investi-

gationsof adverbials(c.f. [Ver97, Jac90, Gaw86,Ern84, Jac72]), treatsyntacticandsemanticissues

thatariseat theclauselevel, someof whichwill bediscussedbelow; whendiscourseadverbialsare

mentionedat all, they arecalledconjunctiveadverbialsor discourseconnectivesandarespecified

asthedomainof discourseresearch.

Nevertheless,we can usetheseclause-level investigationsas a guide when investigatingthe

semanticmechanismsthat causecertainadverbialsto function asdiscourseconnectives, because

all adverbialscanbe classifiedalongtwo semanticdimensions:(1) the type of modificationthey

perform; and (2) the semanticobject(s)they apply to. In this section,we investigatehow prior

analysesof modificationtypecanbeextendedto includediscourseadverbials.

3.3.1 Clause-Level Analysesof Modification Type

Becausethesetof adverbialsis so large,adverbialsareoftenclassifiedin theliteraturein termsof

thetypeof modificationthey perform.[Ale97] summarizesavarietyof modificationtypesthathave

beenproposedin the literature,asshown in Table3.4 with a correspondingexampleshown in the

secondcolumn7. Thoughthis classificationhasbeenappliedonly to ADVP, it canbeappliedto PP

aswell, asshown in thethird columnof thetable.

Modificationtypesvaryasto whetherthey areattributedto Sor VP modification.For example,

thefirst setshown in Table3.4 is generallyattributedto S-modification,while thethird setis gen-

erally attributedto VP modification;thesecondsethave beenvariouslyanalyzedasbothS andVP

7Negation’s adverbialpropertiesarevariouslytreated(c.f. [Ale97]) andarenotdiscussedin this thesis.

87

Page 105: Thesis about discourse

modification(see[Ale97] for discussion).

Table3.4: [Ale97]’sModificationTypes

Modification Type ExampleADVP Example PP

Conjunctive consequently asaconsequenceEvaluative unfortunately to my disfortuneSpeechAct/Speaker-Oriented frankly to speakfranklySubject-Oriented courageously in acourageouswayModal probably in all likelihoodDomain legally in legal terms

Time today on thisdayFrequency frequently at mosttimesLocation here at thisplace

Manner correctly in acorrectwayCompletion/Resultative completely in acompletefashionDegree/Aspectual/Quantificational very/always to a largeextent/atall times

Modification typesalsovary somewhat dependingon the researcher. For example,Table3.5

shows [Ern84]’s classificationof themodificationtypesof S-modifiers.Again, thoughErnststud-

ies only ADVP, thesetypescanbe appliedto PP, as shown in the third column of the table. A

comparisonof theS-modificationtypesin two tablesrevealsthatwhile somedifferences,suchas

“modal” versus“epistemic”, are nominal, othersare categorical; Ernst, for example,subdivides

“subject-oriented”into two classes,“agent-oriented”and“mental-attitude”.

[KP02] definea yet anothersetof modificationtypesfor both S andVP modification,shown

in Table3.6alongwith ADVP andPPexamples;not all ADVP have a clearexample,however, as

indicatedby “?” in thetable.They usethis setto annotateall of theadverbialsthatappearin Penn

Treebank.This annotationis partof a largerprojectinvolving theadditionof semanticinformation

to parsedcorpora. As the primary focus of this project concernsthe similarity in the semantic

rolesplayedby verbalarguments,not adjuncts,acrossa variety of syntacticstructures,the setof

modificationtypesthey defineis understandablymoregeneral.For example,astheADVP andPP

examplesindicate,“temporal” is usedto labelboth“time” and“frequency” adverbials,and“other”

representsa remainderclass,e.g. for thoseadverbialswhosesemanticinterpretationdoesnot fall

88

Page 106: Thesis about discourse

into any of theothertypes.

Table3.5: [Ern84]’sModificationTypes

Modification Type ExampleADVP ExamplePP

Conjunctive therefore asa resultEvaluative surprisingly to my surprisePragmatic frankly to speakfranklyAgent-Oriented wisely in awisewayMentalAttitude willingly in awilling wayEpistemic probably in all likelihoodDomain linguistically in linguistic terms

Table3.6: [KP02]’s ModificationTypes

Modification Type ExampleADVP Example PP

Temporal usually in themorningLocative here at thebarnDirectional back backto workManner quickly in ahurriedfashionPurpose ? in orderto getaheadDiscourse however in additionCause therefore becauseof thisOther probably exceptfor July4

Building on [Gaw86], [Ver97] takesanevenmoregeneralview of modificationtypes,although

heranalysisincorporates,to someextent,onthemorespecificmodificationtypesalreadydiscussed.

Following [Kas93, PS87],shedistinguishesonly threedifferenttypesof modifiers,basedontheway

in whichthey incorporatethesemanticcontentof themodifiedelement.Althoughthesedistinctions

coverbothADVP andPPadverbials,it is notclearwhetherall S-modifiershavebeenconsidered;the

focusof heranalysisis mainly on semanticdifferencesbetweenverbaladjunctsandcomplements.

In hercategorization,restrictiveadjunctsspecifythevalueof an index associatedwith a semantic

objectthatwaspreviously under-specified,suchasthetime or locationof anevent,asexemplified

by thetwo italicizedtemporalandlocative adverbialphrasesin (3.31a).

(3.31a) Johnjoggedyesterdayin thepark.

89

Page 107: Thesis about discourse

In contrast,operator adjunctspredicateon thesemanticobjectthey modify andtherebybuild a

morecomplex semanticrepresentationof theobject,asexemplifiedin (3.32a)by thepredicationof

aneventby thetwo italicizeddurative andfrequentive adverbialphrases.

(3.32a) Johnjoggedtwicea dayfor twentyyears.

The third classof modifiersshecalls thematicadjuncts. Thesearestatedto be a group that

doesnot fit into eitherof the other two typesandaredescribedasadjunctswhich add thematic

informationaboutthe semanticobject,suchasthe causeof the event, or the meansby which the

eventoccurs,asexemplifiedby thetwo italicizedadverbialphrasesin (3.33).

(3.33a) Johnopenedhis doorwith a credit card becauseof therobbery.

Thepurposeof thesedistinctionsis to explain often-observed orderingrestrictionson different

adjuncts.In termsof thiscategorization,therelativeorderingof restrictiveadjunctsdoesnotusually

changethe interpretationof a sentence,asa comparisonof (3.31b) with (3.31a) shows. Because

restrictive adjunctssimply specify(or restrict) indicesof events,the orderin which theseindices

arespecifiedis irrelevant.

(3.31b) Johnjoggedin theparkyesterday.

Changingtheorderingof operatoradjunctscanchangetheinterpretationhowever, asacompar-

isonof (3.32b) with (3.32a) shows. Thefirst operatoradjunctassertsthedurationof the jogging

event,andthesecondoperatoradjunctassertsthefrequency of thejoggingeventwith its specified

duration. (3.32b) is in fact ungrammaticalbecauseit is temporallyimpossibleto repeatan event

thatlastsfor twentyyearstwice aday.

(3.32b) *Johnjoggedfor twentyyears twicea day.

Althoughnot discussedin [Ver97], thematicadjunctscanalsodisplayanorderingpreference;

(3.33b) seemsharderto processthan(3.33a).

(3.33b) ?Johnopenedhis doorbecauseof therobberywith a credit card.

In Section3.4we discussotheranalysesof relative orderingrestrictionsobservedin adjuncts.

90

Page 108: Thesis about discourse

3.3.2 Problemswith CategoricalApproaches

As the varioussetsof modificationtypesindicate,it canbe difficult to achieve agreementon the

propernumberandtype of modificationsthat adverbialsperform. Dependinguponwhich setof

modificationtypesis selected,thereareadditionaldifficultiesaswell.

On theonehand,at a finer grainof analysis,suchasthatemployed in [Ale97] or [Ern84], it is

oftendifficult to decideinto which particularmodificationtypea givenadverbial shouldbeclassi-

fied. For example,thoughuseof theadverbquickly suchasis shown in (3.34)is oftenclassifiedas

conveying “manner”modification,at thesametimeit alsoconveys temporalinformation.Similarly,

thoughuseof adverbslike generally andhistorically] suchasshown in (3.35)areoften classified

asconveying “frequency” modification,they canalsobeviewedasconveying “domain” modifica-

tion, aswell asan expectationaboutprobability, akin to “epistemic” adverbs. And a greatmany

modifiersof a giventypecansimultaneouslybeviewedas“evaluative”, particularlythosethatare

“epistemic”(e.g.probably) and“agent-oriented”(e.g.wisely).

(3.34)Johnranquickly to thestore.

(3.35)Generally/Historically, Johnarrivesatwork on time.

Moreover, many adverbialshave different “readings”,eachof which may be classifiedinto a

differentmodificationtype. For example,briefly canbe classifiedasconveying a “manner” (and

“temporal”) modificationin (3.36a),but in (3.36b) it canbeclassifiedasconveying a “speech-act

oriented”(and“temporal”)modification.And many adverbsthatcanbeclassifiedas“manner”in a

sentencesuchas(3.37a)canbeclassifiedas“domain” in asentencesuchas(3.37b).

(3.36a) Johnsaidhewill stopby briefly.

(3.36b) Briefly, Johnsaidhewouldstopby.

(3.37a) Johnis growing emotionally.

(3.37b) Emotionally, Johnis growing.

Casessuchasclearly andobviouslycombinetheseambiguities.In additionto having multiple

“readings”which canbeclassifiedasconveying eithera “manner”or “evaluative” (and“spatial”)

modification,asshown respectively in (3.38a)-(3.38b) (examplesfrom [Ern84]), they simultane-

91

Page 109: Thesis about discourse

ously convey propertiesthat areepistemic,in that the truth of the modifiedclauseis conveyed as

apparent.

(3.38a) Johnburpedclearly/obviously.

(3.38b) Clearly/Obviously, Johnburped.

And evenatsuchafinegrainof analysis,notall clausaladverbialsarerepresented.For example,

althoughthesetypeswerenot intendedto cover thewide varietyof PPadverbials(e.g. “exceptfor

this”) foundin ourcorpus(aswe will seein Section3.5), it is not clearwhereanADVP adverbials

suchas“regardlessof rightsandwrongs”would fall, which is alsofoundin our corpus(aswe will

seein Section3.6).

On theotherhand,even if a moregeneralsetof modificationtypes,suchasthatemployed in

[KP79] or [Ver97], is used,thesamedifficulty in decidingamongcategoriesfor a givenadverbial

canarise,in additionto theproblemthatpotentiallyvaluabledistinctionsarelost,anda“container”

classis necessaryto gatherremaindersthatdon’t fit in otherclasses.

3.3.3 Modification TypesasSemanticFeatures

On the surface, it might appearthat the modificationtypescould at leastbe usedto distinguish

discourseadverbials. After all, in Tables3.4 and3.5, discourseadverbialsareclassifiedas“con-

junctive”, and in Table3.6 they areclassifiedas“discourse”. But thesemodificationtypeshave

not beenableto distinguishclausalanddiscourseadverbials. For example,while clause-level re-

searcherstreat surprisingly, unfortunately, clearly as clausaladverbials, [Kno96] treatsthem as

discourseconnectives, as discussedin Chapter2. And in fact, all of the “conjunctive” or “dis-

course”adverbialscanbeclassifiedinto someothermodificationtype, for in contrastto theother

modificationtypes,whosepurposeis to isolatea particularpropertyconveyed by a setof adver-

bials,the“conjunctive” or “discourse”typeisolatesaparticularsyntacticstructureunderlyingaset

of adverbials.For example,discourseadverbialssuchasthen,first, finally, alreadycanbe(andare

in [KP79]) classifiedas“temporal”,becausethey convey temporalinformationabouttherelationof

theelementthey modify to thesurroundingdiscourse(or spatio-temporalcontext). Discoursead-

verbialssuchasasa result,consequentlycanbeclassifiedas“evaluative” if thecausalconnection

92

Page 110: Thesis about discourse

betweentheelementthey modify andthesurroundingdiscourse(or spatio-temporalcontext) is not

commonknowledge8. Above we discussedproblemsof multiple typing andunclearclassification

asthey applyto clausaladverbials.Herewe seethattheseproblemsextendto discourseadverbials

aswell.

One solution to theseproblemslies in adoptinga non-categorical approachto modification

type. If modificationtypeswereviewedin termsof semanticfeatures,for example,thenadverbials

couldberepresentedassupplyingmultiple(compatible)features.Moreover, thesefeaturescouldbe

usedto representbothclausalanddiscourseadverbials. In suchterms,probablymight beused,as

exemplifiedin (3.39),to supplya degreeof likelihoodfeature9, while clearly andobviouslymight

beused,asexemplifiedin (3.40),to supplyevaluative,epistemic,andspatio-temporalfeatures.

(3.39)ProbablyJohnwoke upat 5 a.m.

(3.40)ObviouslyJohnwoke up at5 a.m.

Similarly, asexemplifiedin (3.41)-(3.42),both thenandon March 14, 1946might be usedto

supply temporalfeatures;only then suppliesthesefeaturesin termsof a relation with the prior

discourse.

(3.41)On March 14,1946, my fatherwasborn.

(3.42)Then, my fatherwasborn.

While suchanapproachrequiresfurtherstudy, suchfeaturescouldbesupplementedwith [Kno96]’s

featuresfor cuephrases,discussedin Chapter2. As statedabove, this approachwould overcome

theproblemsof multiple typingandunclearclassificationfoundin categoricalapproachesto modi-

ficationtypeby allowing adverbialsto supplyfeaturesfrom avarietyof modification“types”.

Whatmodificationfeaturescannotdistinguish,however, is thefact thatwhile thefeaturessup-

pliedby clausaladverbialsgenerallyapplyto entitiesand/orpropertieswithin themodifiedelement,

thefeaturessuppliedby discourseadverbialsapplyto themodifiedelementitself andthesurround-

ing discourseor context. Accountingfor thisdifferencerequiresdiscussionof thesemanticobjects

8[Kno96]’s“sourceof coherence”featuredistinguishesthisuseas“semantic”,asopposedto “pragmatic”,asdiscussedin Chapter2.

9[Ern84] makesa referenceto sucha feature,but hisanalysisappliesto categoricalmodificationtypes

93

Page 111: Thesis about discourse

that theseadverbialssupplyfeaturesto. Considerationof theseobjectscanalsoclarify problems

with multiple readingsin bothclausalanddiscourseadverbials,therebydecreasingtheamountof

variationthatmodificationfeaturesarerequiredto cover. And it works theotherway too; under-

standingthemodificationfeaturesof adverbialscanclarify thepossiblesemanticobjectsthey can

apply to; togethertheseclassificationscanbeusedto build a formal semanticrepresentationof all

adverbials.It is thisdiscussionwhich is thesubjectof thenext few sections.

3.3.4 Summary

In thissection,wehaveshown thatthemodificationtypealoneis notsufficient to distinguishclausal

anddiscourseadverbials. We have suggestedthat someproblemsof multiple typing andunclear

classificationcanhoweverbeovercomeif modificationtypeis viewedin termsof semanticfeatures.

3.4 Adverbial SemanticAr guments

A numberof issuesconcerningthe semanticinterpretationof the externalsyntacticargumentof

adverbialshave beenaddressedat theclauselevel. In thissectionwe presenttheseissuesandshow

how they extend to the discourselevel. We show how the discoursedeixis researchintroduced

in Chapter2 exposesthe semanticobjectsthat adverbialsapply to anddistinguishesclausaland

discourseadverbialsin termsof thethenumberandtypeof thesesemanticobjects,therebylaying

thefoundationfor understandingthesemanticmechanismscausingthe“discourseconnectivity” of

discourseadverbials.

3.4.1 (Optional) Ar gumentsor Adjuncts?

As discussedin Section3.2,ADVP andPPcanfunctionbothasadjunctsandVP arguments;am-

biguity betweenthesefunctionshasmainly beenaddressedin relationto PP, but the analysesare

applicablealsoto ADVP.

A standardsyntactictest for VP argumentstructureis the “do so” test ([Ver97, KP79]). As

shown in (3.43a-b),wherebracketsindicatetheboundariesof theverbandits internalarguments

94

Page 112: Thesis about discourse

anditalics indicatethereferentof “do so”, “do so” mustreplacetheentireVP, e.g. theverbandall

its arguments;thePPargumentcannotbeleft out.

(3.43a) *Mik e [gavea recommendationto Phyllis] andMary did soto Liz.

(3.43b) Mike [gavea recommendationto Phyllis] andMary did sotoo.

As shown in (3.43c-d), if thePPis anadjunct,however, it neednotbereplacedby the“do so”.

[SBDP00,Har99] explain thedifferenceasfollows: if thePPsuppliesanargument,thereis noway

to determinewhatsemanticfunction it serveswhenit modifies“do so”, because“do so” canrefer

to theinterpretationof modifiedandunmodifiedverbphrases10.

(3.43c) Mike [readthetextbook] in thebedroomandMary did soin theclassroom.

(3.43d) Mike [readthetextbook] in thebedroomandMary did sotoo.

[PS87]alsodiscussthe“iterability” test,which distinguishesVP adjunctsfrom argumentsbe-

causeonly adjunctscanbe“iterated”,asshown in (3.44).

(3.44a) JohnmetSusanin Chicagoin theHyatt hotel in thelobby.

(3.44b) *Johngaveabookto Debbieto Paul.

As [Ver97] pointsout,however, failing theiterability test,asin (3.44c), doesnotalwaysimply

argumentstatus;adjunctsarein generalnot iterableif their semanticcontribution is contradictory.

(3.44c) *JohnmetSusanin Chicagoin Boston.

[PS87]alsonotethat in English,argumentstendto precedeadjuncts,asshown in (3.45a). In

(3.45b), changingtheorderof thePPschangestheinterpretation;thebook,ratherthanthe“giving”

is readasbeinglocatedin thelibrary.

(3.45a) Johngave abookto Debbiein thelibrary.

(3.45b) Johngave abookin thelibrary to Debbie.

Furthermore,[PS87]note,many adjunctscauseextractionislands,asshown in (3.46a), while

unboundeddependency into argumentsis generallypossible,asshown in (3.46b). As [SBDP00]

10Suchtestsaren’t perfect,however; asBonnieWebbernotes(personalcommunication),(3.43a) soundsfine if wereplacegiveto with providefor: Mike provideda recommendationfor PhyllisandMary did sofor Liz.

95

Page 113: Thesis about discourse

explain, in LTAG,extractionis modeledarelationamongelementarytreesin atreefamily thathave

essentiallythe samemeaninganddiffer only in syntax. In a treewherean elementis extracted,

its original positioncanbe locatedin a differenttreein thesamefamily, andits semanticscanbe

computed.Failedextractionindicatethat the modifier from which the elementis extractedis not

presentin any elementarytreein thetreefamily.

(3.46a) *Wheredid Johngive abookto Debbiein ?

(3.46a) Whomdid Johngive abookto in thelibrary?

Semantictestsfor distinguishingverbalargumentsandadjunctshave alsobeenproposed.As

[PS87]note,thesemanticcontribution of anargumentis dependentonthemeaningof thehead.For

example,in “Johntold astoryto Mary”, the“telling” musthavebotha “thing told” anda recipient;

“to Mary” fills the latter role. More formally, thesemanticcontribution of anargumentis entailed

by the sentencecontainingthe verb11. For example,the PPsin (3.47 a) areentailed( ¬ruv ) by the

sentencein (3.47b), thoughnot instantiated,but thePPsin (3.48a)arenotentailedby thesentence

in (3.48b) (examplesfrom [Ver97]). Theseentailmentpatternsindicatewhetheror not semantic

informationsuppliedby a PP(or ADVP) is directly relevantto themeaningof theverb([Ver97]).

(3.47a) Johncomplainedto Mary abouttheheat.

(3.47b) Johncomplained.¬ v hq�����Á¬ Johncomplainedto � about�(3.48a) Johnsangto Mary abouthis homeland.

(3.48b) Johnsang. ¬ruv h��Ŭ Johnsangto � , ¬ruv hq�Á¬ Johnsangabout�Another semantictest to distinguishverbal argumentsand adjunctsis the “presupposition”

test12. As definedin [Sae96], the presuppositiontest is appliedby comparingthe interpretation

of a sentencewith a modifier, asshown in (3.49a), to a correspondingsentencewithout themod-

ifier, wherethe informationsuppliedby themodifier is availablein theprior context, asshown in

(3.49b) (examplefrom [SBDP00]).If thetwo sentencescanhave thesameinterpretation,thenthe

modifierexpressesapresupposedsemanticargument.

11Theentailmentrelationin linguisticsis onein whichthetruthof onesentencenecessarilyimpliesthetruthof another([ODA93]).

12Generally, a presuppositionis an assumptionor belief implied by the useof a particularphrase([ODA93]), but acompleteaccountof presuppositionis still anopenquestion(c.f. Chapter5).

96

Page 114: Thesis about discourse

(3.49a) Find thepowercable.Disconnectit from thepower adapter.

(3.49b) Thepowercableis attachedto thepower adapter. Disconnectit.

While noneof thesetestsarefoolproof ([SBDP00]),andtheargumentstructureof many verbs

is still undecided,clearcasesserve to show thatverbalsemanticargumentsaresometimessyntacti-

cally optional, althoughtheir semanticcontribution is still interpretedusingeithercontext or world

knowledge. Suchsyntacticallyoptionalarguments,alsocalled“hidden” or “implicit”, have been

widely discussedin linguistics,psycholinguistics,andcomputationallinguistics (c.f. [MTC95]).

Implicit argumentsaregenerallyclassifiedinto two types:definiteandindefinite. Definitehidden

argumentsareanaphoricto somesaliententity in the discourseor spatio-temporalcontext; their

interpretationis context-dependent. For example,in (3.50a)-(3.51a)optionalargumentsareinstan-

tiatedasPPswhoseanaphoriccomplementresolvesto thebold-facedelementin theprior clause.

(3.50a) Theduedatefor the grant haspassed.Mary didn’t apply for it.

(3.51a) Bill nearlyforgot aboutgoing to the bank. Johnremindedhim aboutit.

As shown in (3.50b) and(3.51b), theseargumentscanbeimplicit, but only if theinformation

necessaryto resolve themis suppliedby theprior context. (3.50c) and(3.51c) areinfelicituous,

becausethis informationis not retrievable.

(3.50b) Theduedatefor thegranthaspassed.Mary didn’t apply G .

(3.51b) Bill nearlyforgot aboutgoingthebank.Johnremindedhim G .

(3.50c) *Mary didn’t apply G .

(3.51c) *JohnremindedBill G .

Indefinitehiddenargumentsarenot anaphoricwith anything; their interpretationis indepen-

dentof context. [Mit82] discussesthe VP eat in detail, in constructionssuchas(3.52),which is

grammaticalwith or without theexplicit argumentevenwithout prior context. As shown in (3.53),

relationalNPs,suchasmother, winner, etc.,canalsocontainhiddenindefiniteNP arguments.

(3.52)Mary ate(something).

(3.53)Mary talkedto amother(of someone)today.

97

Page 115: Thesis about discourse

Example(3.54)(from [MTC95]) shows that thedistinctionis not alwaysstrict. Whenimplicit,

thesyntacticallyoptionalargumentof theverbdonatecanbeanaphoricto informationin theprior

context, asin (3.54b), or canbeinterpretedasunspecifiedandthusindependentof any information

in theprior context, asin (3.54c).

(3.54a) TheUnitedWayaskedJohnfor acontribution. Johndonatedfivedollarsto them.

(3.54b) TheUnitedWayaskedJohnfor acontribution. Johndonatedfivedollars G .

(3.54c) Hardlyanyoneknows thatJohndonatesthousandsof dollarseachyear G .

3.4.2 External Ar gumentAttachment Ambiguity

EvenwhenanADVP or PPis identifiedasanadjunct,additionalambiguitiesmayexist. In partic-

ular, thepositionalvariability of ADVP andPPdiscussedin Section3.2 producesambiguityasto

thephrasalunit beingattachedstructurallyandmodifiedsemantically. A commonexampleof this

ambiguityis shown in (3.55),wherethePPwith a telescopeis ambiguousbetweenadjectival (NP

modifying)or adverbial(VP or Smodifying)attachment,yieldingambiguityasto whois in posses-

sionof thetelescope.Therearea numberof approachesto structuraldisambiguationof adjectival

versusadverbialPPattachment,includingprobabilisticones(c.f. [Bik00, CW00,McL01]).

(3.55)Johnsaw themanwith thetelescope.

A centralconcernfor many clause-level ADVP researchersis the ambiguitybetweenVP and

S modification;mostof theseresearchersseekto associateVP andS modificationwith particular

modificationtypes. (Schreiber1971),for exampletakesa “deepsyntactic”approach,arguing that

evaluative andmodaladverbsareunderlyinglyadjectivesthattake a sententialsubjectnominal.As

evidenceof thishecitesthesynonymy of thetwosentencesin (3.56a). Hearguesthatthissynonymy

distinguishesthemfrom VP modifiers,which, asshown in (3.56 b), cannotbe paraphrasedwith

sententialsubjectnominalsevenwhenthey areS-initial.

(3.56a) ThatJohnwasat faultwasobvious. K Obviously, Johnwasat fault.

(3.56b) *That Johnspoke to his friendwaswarm. K Warmly, Johnspoke to his friend.

98

Page 116: Thesis about discourse

In [Swa88]’s terms13, the essenceof [Sch71]’s argumentis that a sentencecontainingan S-

modifyingadverbexpressesat leasttwo setsof propositionalcontent:theunmodifiedandthemod-

ified proposition. Swan take a differentview of the modifiedproposition,however, andincludes

subject-andspeaker-orientedadverbs.In Swan’s view, sentencesmodifiedby speaker-orientedad-

verbsarerepresentedasin (3.57). This representationindicatesthat, in additionto conveying the

unmodifiedproposition(S),speaker-orientedadverbsalsoconvey areducedpropositionconcerning

thespeaker (Speaker says(S)),which is modifiedby theADJ derivative of theadverb.

(3.57)(S), (ADJ(Speaker says(S)))e.g.frankly, honestly

Sentencesmodified by evaluative, modal and subject-orientedadverbsare representedas in

(3.58).This representationindicatesthat,in additionto conveying theunmodifiedpropositionitself

(S), theseadverbsalsoconvey areducedpropositionconcerningthespeaker in which thespeaker is

evaluatingasADJ theinformationin themodifiedproposition.

(3.58)(S), (Speaker says(ADJ(S)))e.g. fortunately, probably, courageously

[Gre69, AC74] take a “surfacesyntactic”approachto distinguishingVP andS modification,

arguingthatS-modifyingadverbscannotbethefocusof negationor interrogation.Thesetestsare

exemplifiedin Table3.7with theADVP italicized;asshown thesetestsalsoapplyto PPadverbials.

Table3.7: [Gre69]’sSyntacticTestsfor DistinguishingVP andS Modification

Syntactic Test ADVP Examples PP Examples

Adverbialis the He didn’t walk slowly. He didn’t walk at a slowrate.focusof negation *He didn’t walk probably. *He didn’t walk in all likelihood.

Adverbialis the Did hewalk slowly? Did hewalk at a slowrate?focusof interrogation *Did hewalk probably? *Did hewalk in all likelihood?

However, noneof theseapproachesaddressthe multiple readingsof adverbssuchasclearly,

obviously, strangelydiscussedabove; they passthetestsin Table3.7,asexamplifiedin (3.59a)-(3.59

b), becausethey displaymultiple “readings”;asVP-modifiers,they aresomevarietyof “manner”

13Swan’smaingoalis to provideacorpus-basedanalysisof English-ly Smodifyingadverbsashaving developedfromVP modifying intensifierandmanneradverbs,throughOld andMiddle English,by meansof syntactic/pragmaticshifts.

99

Page 117: Thesis about discourse

adverbs,while asS-modifiersthey areoftenclassifiedas“evaluative”.

(3.59a) Johndidn’t burp clearly/obviously/strangely.

(3.59b) Did Johnburp clearly/obviously/strangely?

[Jac72] and[MG82] argue that suchcasesshouldbe treatedashomonyms, andgiven multi-

ple lexical entries. [Ern84] simplifiesthis analysisby giving adverbs,whenever possible,a single

semanticinterpretation,while leaving their syntacticargumentunderspecifiedin the lexical entry

whenmultiple interpretationsarepossible.For example,becausemodaladverbssuchasprobably

donotchangetheirmeaningnomattertheirpositionin theclause,asaclasstheirsyntacticargument

canbelexically specifiedin a singlelexical entryasS.But whenmultiple readingsarepossible,as

for obviouslyandstrangely, Ernsttreatsthesyntacticargumentaslexically underspecified,drawing

its syntacticargumentfrom theclosestdominatingnode14. Obviously, for example,alwaysconveys

that theargumentis obvious, but theargumentcanbea VP, yielding a mannerinterpretation,or an

S,yielding anevaluative interpretation.

Ernstalsoappliesthisargumentunderspecificationanalysisto adverbsthathavebeenpreviously

classifiedwith asinglemodificationtypebut thatalsodisplaysubtlydifferent“readings”.Examples

commonlyclassifiedas“agent-oriented”areshown in (3.60).

(3.60a) JohnapproachedtheDuchesstactlessly.

(3.60b) Tactlessly, JohnapproachedtheDuchess.

In Ernst’s analysis,tactlesslyin (3.60a)describesa judgmentaboutthe“approaching”,andthe

adverb takesa VP argument,while tactlesslyin (3.60b) describesa judgmentaboutthesituation

involving John’s actionof “approaching”,andtheadverbtakesanSargument.

In both cases,“John” is interpretedas “tactless”. However, Ernst arguesthat casessuchas

(3.61)-(3.62)show thatthis is only aninference.For example,wewouldnormallydraw from (3.60

a) the inferencethat Johnis tactless.But we canusethe S-modifierwith the opposingmeaning

(tactfully) in (3.61),alongwith theadditionalcontext providedby theVP adverbial(knowing...), to

blocktheinferredjudgmentthatJohnis tactless,while still assertingthatJohn’sapproachis tactless.

14In thenormalcase,that is. Many readerswill beableto getmultiple readingsin any position;arguablytheclosestdominatingnodecorrespondsto thefirst, or easiest,reading.

100

Page 118: Thesis about discourse

(3.61)Tactfully, Johnapproachedthe Duchesstactlessly, knowing that apparentdisregardfor

authoritywashighly prizedin thisculture.

(3.62)Stupidly, Alice hadanweredthequestionswiselyandblown hercoverasaninmateof the

insaneasylum.

While S-readingscannotbefaked,asexemplifiedin (3.63),wherebothadverbsareS-modifiers

becausetheclosestdominatingnodeof bothis S,Ernstarguesthatthis lackof a felicituous“f aked-

reading”is accountedfor by thefactthattheS-modifiersassertopposingjudgmentsaboutthesame

semanticobject;theonly way to block oneof thejudgmentsaboutAlice is to explicitly claim that

thesurroundingcircumstanceswereabnormal,asin (3.64).

(3.63)*Stupidly, Alice wiselyhadanweredthequestions.

(3.64)Althoughit normallywouldhavebeenstupidfor Alice to blow hercoverasamime,Alice

wiselyhadansweredthequestionsbecauseshehadalreadynoticedthatherbackuphadarrived.

Ernst’s useof variationsin theexternalsemanticobjectinutitively feelscorrect. Investigating

theuseof adverbs(or PPs)with gradientmeanings(insteadof just opposingmeanings)would in-

dicateif morethantwo argumentattachmentsitesfor interpretingsemanticobjectswererequired;

adjectives, from which many adverbsarederived, aregenerallynon-gradient,however [WN98].

Nevertheless,Ernst’sanalysisis largelyunformalized,with respectto boththesyntacticandseman-

tic representationof adverbsandtheir arguments;moreover, it doesnot determinehow to decideif

asinglemodificationtype(suchas“temporal”) shouldbeviewedasSor VP.

[Ale97] basesthisdecisiononobservedorderingrestrictionsonadverbs,while alsoproviding an

analysisof thesyntacticrepresentationof adverbattachment.As discussedin Section3.3, [Ver97]

explainscertainorderingrestrictionsin termsof modificationtype; [Ale97], in contrast,explainsa

widervarietyof theserestrictionsin termsof bothmodificationtypeandunderlyingadjunctionsite.

[Ale97] notesthatstrict sequencingandscopehierarchiesattestedacrosslanguagesarecorrelated

with modificationtype. For example,in English,asexemplifiedusingmodaladverbsin (3.65a)-

(3.65b),S-modifyingadverbsmustappearhigherin thesyntactictreethanmanneradverbs,andas

shown in (3.66a)- (3.66b),evaluative adverbsmustappearhigherthanagent-orientedadverbs.

101

Page 119: Thesis about discourse

(3.65a) ProbablyJohncleanedtheroomcarefully.

(3.65b) *Carefully Johnprobablycleanedtheroom.

(3.66a) Fortunately, Johncleverly climbedto thetopcarefully.

(3.66b) *Cleverly, Johnfortunatelyclimbedto thetopcarefully.

Workingwithin theMinimalist program,Alexiadouprovidesacross-linguisticsyntacticaccount

for suchrestrictionson ADVP accordingto their modificationtype. In his analysis,“manner”ad-

verbsarebase-generatedbelow theverb,andtheirappearancein otherpositionsinvolvesmovement

to aspecifierpositionof a functionalprojectionaslicensedby Minimalist principles.Otheradverbs

arebase-generatedasspecifiersof functionalprojectionsto the left of the verb. He distinguishes

a varietyof functionalprojections,distributesalongthem(astheir underlyingplaceof attachment)

ADVP that exemplify the variousmodificationtypes,andin eachcasedemonstratesthe rangeof

felicituoussurfacepositionsthatresultfrom movementof themodifiedphrasalcategories.

Of course,given how large the set of ADVP adverbials is, and the inability of modification

typesto fully categorizeall adverbs,it’s not clear if [Ale97] accountsfor all possiblesequences

andpositionsfor all ADVP adverbials. In particular, theapparentfreedomof comma-(or pause-)

delimitedadverbialssuchasshown in (3.65c) - (3.66c) is largelyunaccountedfor usingMinimalist

principles,althoughto be surethe felicitousnessof thesecasesmay vary to a greatextent on a

speaker-to-speaker basis,andthecasesthemselvesmaybemuchmorecommonin speechthanin

text, thusmakingthemanalyzableasself-corrections.

(3.65c) CarefullyJohn,probably, cleanedtheroom.

(3.66c) Cleverly, John,fortunately, climbedto thetopcarefully.

[Bie01] cites[McC88]’s treatmentof comma-delimitedmodifiersasS-adjuncts,but notesthat

this analysisallows infelicituoussentencessuchasshown in (3.67),whereotherthanMary should

beinterpretedasmodifyingeveryone, despiteits comma-delimitedS-medialposition.

(3.67)*What food,otherthanMary, repelseveryone?

In suchananalysis,the relationshipbetweeneveryoneandtheother thanMary phrasewould

have to be resolved anaphorically. [Bie01] proposesinsteada fully structuralanalysisin CCG

102

Page 120: Thesis about discourse

[Ste00c] which usestype-raisingto cover the felicituouspositionsof alternativephrases.We will

discusshis semanticanalysisof thesephrasesin Section3.5.

3.4.3 SemanticRepresentationof External Ar gument

As discussedin Chapter2, muchof the linguisticsliteratureaddressesonly the propositioninter-

pretationof anS; thesemanticrepresentationof S is usuallya truth value: trueor false.Theabove

discussionindicatesthatS makesavailableotherinterpretationsaswell. In [Ver97, KP79], for ex-

ample,andacrosstheadverbialliteraturein general,referenceis madeto themodificationof events

andsituations. [Ern84] distinguishesaneven wider varietyof S interpretations,includingactions

(or events),situations,statesof affairs,andmentalstates. However, in mostcasesthisclassification

is describedrathervaguely;while theseinterpretationsmaybeusedto distinguishvariousproperties

of adverbials,thepropertiesdistinguishingtheseinterpretationsarenot well-defined.For example,

Table3.8providesthesemanticinterpretationsof someof Ernst’s modificationtypes15.

Table3.8: SemanticInterpretationsof [Ern84]’sModificationTypes

Example Modification Semantic Semantic DerivedADVP Type Inter pretation Object ( � ) ADJ

slowly Manner � is ADJ. action slowwisely Agent-Oriented agentjudgedADJ dueto � situation wiselegally Domain � relevant in ADJ domain situation legal

possibly Epistemic � is ADJ. situation possiblefortunately Evaluative � is ADJ. stateof affairs fortunate

angrily MentalAttitude ADJ manifestsagent� mentalstate angry

Thefirst columnof the tablecontainsanexampleadverb,andthesecondcolumncontainsthe

modificationtypeinto which it is classified.Thethird columncontainsthesemanticinterpretation

of themodification,anddefinesthe involvementof thesemanticinterpretation,� , that is equated

with theexternalsyntacticargument.Thespecificobjectequatedwith � for eachinterpretationis

shown in the fourth column. Note that Ernstmakesuseof the adjective (ADJ) derivative of the

ADVP; thefifth columncontainstheADJ derivative for eachadverb.

15Certainof Ernst’smodificationtypesaresubdividedwith respectto theseinterpretations.

103

Page 121: Thesis about discourse

[Moo93], in contrast,providesa formal representationof thepossiblesemanticinterpretations

of ADVP adverbials’externalarguments.His formalizationis basedon thesamedataaddressedin

[Ern84], in particularthetwo differentinterpretationsexemplifiedin (3.68)-(3.69)thatareproduced

by a singleadverb locatedin differentpositions. Moore notesthat in (3.68),strangely intuitively

modifiestheeventof Johnsinging. In (3.69),however, thesingingeventmaybequiteordinary;it

is thefactthatJohnsangthatis strange.

(3.68)Johnsangstrangely.

(3.69)Strangely, Johnsang.

Following [Dav67], Moorearguesthateventsentences16 asserttheexistenceof anevent in the

domainof entities. It is this event, theseresearchersargue,thatstrangely modifiesin (3.68). This

is representedin predicatelogic asshown in (3.68’), where � representsa hiddenargumentto the

verb17 thatrangesoverevents,andstrangely is representedasadditionalpredicationof � . In words,

theformulaassertsthatthereexistsa singing-by-Johnevent,andthateventis strange.

(3.68’) h x(Sang(j,x)& Strange(x))

Moore further arguesthat true propositionsassertthe existenceof a situation (or fact) in the

domainof entities. In Moore’s view, it is this situationthat strangely modifiesin (3.69). This

is representedin predicatelogic asshown in (3.69’), where“Fact” denotesa relationbetweena

situationanda proposition,� is representedasa hiddenargumentto this relationthat rangesover

situations,andstrangely is representedasadditionalpredicationof � . In words,theformulaasserts

thereexistsasituation(fact)of therebeingasinging-by-Johnevent,andthatsituationis strange.

(3.69’) h y(Fact(y, h x(Sang(j,x)))& Strange(y))

3.4.4 SemanticAr gumentsasAbstract Objects

In essence,both Moore andErnstassociateadverbial VP attachmentwith event (or action) mod-

ification. While Moore associatesadverbial S attachmentwith situation (or fact) modification,a

16sentencesthatdescribeevents,in contrastto sentencescontainingthebeverb,e.g.Mary is kind.17See[Moo93, Dav67, BP83]for reasonswhy this argumentis associatedwith theverbandnot thesentence.

104

Page 122: Thesis about discourse

situationis only oneof thepossiblesemanticinterpretationsthatErnstassociateswith adverbialS

attachment;healsomentionsthemodificationof statesof affairs andmentalstates.

In fact,thediscoursedeicticresearchdiscussedin Chapter2 providesevidencethatthepossible

semanticinterpretationsassociatedwith adverbial attachmentare more diversethan Moore and

even Ernstacknowledge. In Chapter2 we referredto thepossibleinterpretationsthatan S makes

available as abstract objects(AOs). As illustration, recall the variety of AO typesproposedby

[Ash93], shown againin Figure3.4. We demonstratedhow discoursedeixiscouldbeusedto refer

to theseobjects.

Figure3.4: [Ash93]’s Classificationof AbstractObjects

Also in Chapter2, [Ven67], [Web91], [DH95] and we ourselves observed that a variety of

additionalAOs, including descriptions,beliefs,speech acts, textual objects,andeven defeasible

rules anddiscourserelationscanbe the objectsof discoursedeixis reference.More generally, as

statedin Chapter2, thereareat leastasmany AOsasthereareabstractnouns.In many casesthey

canbereferredto via discoursedeixissimplyby insertingtheminto thesentence“That is a(n)...”.

Whetheror not theclassificationof AOs shown in Figure3.4 is complete,we arguethatAOs

provide an appropriateway of understandingadverbial semantics.We arguethat the rangeof se-

manticobjectsevokedwhenanadverbialmodifiesanScoincideswith therangeof semanticobjects

to whichadiscoursedeicticrefers.

As a preliminaryillustration,considerTable3.9. The third andfourth columncontainPPand

ADVP clausalanddiscourseadverbials.Thesecondcolumncontainscomparablediscoursedeictic

reference.For illustrative purposes,considerthesentencePeoplemademistakesastheonebeing

105

Page 123: Thesis about discourse

referredto andmodified.Thefirst columnspecifiestheresultingAO interpretation.

Table3.9: AbstractObjectInterpretations

Abstract Object DiscourseDeictic PP Adverbial ADVP Adverbial

event/situation thathappenedafterwards afterthat, afterwards,fact that’s a fact in fact, really,proposition that’s true in truth, truly,description that’s agooddescription asadescription, descriptively,belief that’s my belief in my view, personally,speechact that’s to befrank in plainEnglish frankly,textual object repeatthat asa repetition again,

As the table shows, we can interpret the sentenceas a variety of AOs. If referredto by a

discoursedeictic,it is thepredicationonthediscoursedeicticthatdeterminestheAO interpretation.

For example,“happeningafterwards”,is apropertyof events,sothereferentof thediscoursedeictic

is interpretedasan event. If the sentenceitself is modified, it is the adverbial that performsthe

predication,therebydeterminingtheAOinterpretation.Forexample,“afterwards”,is alsoaproperty

of events,sothesentenceisagaininterpretedasanevent.Moregenerally, noticethatwecanin many

casescreatea PPadverbial for an AO simply by insertingit into thePP“As a(n) ...”. In Sections

3.5-3.6we will seeawide varietyof AOsinstantiatedin thisandotherS-modifyingadverbials.

Thoughour dataconsistsof S-modifying adverbials,we have not excludedeventsor distin-

guishedthemfrom situationsin theabove table,becausefor many S-modifyingadverbials,includ-

ing temporal,frequency andspatialadverbials,eventandsituationmodificationcannoteasilybe

distinguished.And if theseadverbialsmodify events, andeventsaremadeavailableby VP, then

eithera movementanalysisor a percolationanalysisis requiredto explain how theseadverbials

canmodify eventswhile adjoiningto S. Moreover, we will seein our datathat thepossibleinter-

pretationsof VP areasimportantfor distinguishingdiscourseandclausaladverbialsasthepossible

interpretationsof S,justasthey arefor distinguishingdiscoursedeicticreferencefrom NPreference

in demonstrative use.In Chapter6 wewill returnto theissueof semanticrepresentationof AOsand

adverbials,takinginto considerationboththeanalysesproposedin discoursedeixisresearchandin

adverbial research.Until then,we retainthedistinctionbetweenthesemanticorigin of eventsand

106

Page 124: Thesis about discourse

situationsvia thephrase“AOsmadeavailableby S”.

For thepurposesof brevity in subsequentsections,we definea working terminologicaldistinc-

tion betweenconcreteobjectsandabstract objects. Theseobjectsaredistinguishedis asfollows:

concreteobjectsareentity interpretationsmadeavailable by (e.g. denotedby or inferablefrom

([Pri81])) NPs.Concreteobjectsareusually, but notalways,perceivableby thesenses[Ash93], and

includepeople, organizations,physicalobjects, etc.Abstractobjectsareentity interpretationsmade

availableby bothNPsandnon-NPconstituents,includingverbphrases,clauses,etc. As discussed

in Chapter2, abstractobjectsareusually, but not always,imperceptableto thesenses,andinclude

reasons,beliefs,trials, defenses,theories,rights, etc. In practicalapplication,the distinction is

moreof the “I will know it whenI seeit variety”, and it may not be categorical; demonstrating

it for every entity would requirean infinite corpus,which is impossible,or constructedexamples,

which areoftensuspect.Nevertheless,thetheoreticaldistinctionbetwenthesetwo typesof entites

clarifiesthe differencebetweenthe anaphoricityof discourseandclausaladverbials. However, it

doesnot accountfor AOsretrieved from thespatio-temporalcontext. Our corpusconsistsonly of

text, andthoughthereis somediscussionof this in thesubsequentsections,in our view, therole of

thespatio-temporalcontext is still anopenquestion.

3.4.5 Number of Abstract Objects

Clausaladverbials(e.g. in myview) anddiscourseadverbials(e.g. afterwards) arebothcontained

in Table3.9becausebothtake astheir externalsemanticargumentanabstractobjectinterpretation

madeavailableby S.Thus,atboththesyntacticandsemanticlevel, clausalanddiscourseadverbials

arenotdistinguishablein termsof theirexternalargument.

It is in termsof thenumberandinterpretationof their argumentsthatsignificantsemanticdif-

ferencesbetweenclauseanddiscourseadverbialsappear. Wearguethatthenumberof argumentsan

adverbial containsandthe interpretationof theseargumentsdeterminewhetheror not it functions

semanticallyasa discourseconnective andis therebyclassifiedasa clauseor discourseadverbial.

We claim thatdiscourseadverbialscontainat leastoneargumentthatdependsfor its interpretation

on somesalientAO containedin or derivable from the discoursecontext, which therebyrenders

107

Page 125: Thesis about discourse

themuninterpretablewith respectto theirmatrixclausealone, evenafterresolutionof any concrete

objectarguments.In contrast,clausaladverbialscontainno suchargument;their interpretationis

context-independent afterresolutionof any argumentsto contextual concreteobjects18.

Discourseadverbialsare thusvery similar to discoursedeixis in that both requirean AO in-

terpretationfrom theprior discourseor spatio-temporalcontext for their interpretation.[WJSK03]

providestrongbehavioral evidencefor thisview. Forexample,justasin (2.50),repeatedfrom Chap-

ter 2, thediscoursedeictic takesasits referentthediscourserelationbetweenclauses,sodoesthe

PPadverbial,asoneexample, in (3.70)derive its prior argumentfrom theresultrelation(imparted

by so) betweenthetwo clauses.

(2.50)If a white persondrivesthis carit’s a “classic”. If I, a Mexican-American,drive it, it’s a

“low-rider”. Thathurtsmy pride. [DH95]

(3.70)Johnjustbroke his arm.So,asoneexample,hecan’t cycle to work now.

More generally, noticethatjustasadiscoursedeicticcanbereplacedby its explicit demonstra-

tive+AO counterpart,e.g. that text, that speech act, or that contrast in (2.50),in thesameway, we

canconstructan adverbial from a discoursedeictic or its explicit demontrastive+AO counterpart,

therebycreatinga discourseadverbial which explicitly relatestwo abstractobjects,e.g. for that

reason,in this case, after that, or asan exampleof theconsequencesof that in (3.70). Of course,

as(3.70)indicates,explicit demonstrative referenceis not theonly mechanismby which discourse

adverbialsarecreated.In thenext two sectionswe discussthesemechanismsin detail.

Beforebeginning our corpusanalysis,however, notethat theappropriatesemanticrepresenta-

tion of almostall the mechanismsdiscussedin Sections3.5-3.6is still an active line of research.

As such,thevariousresearchwe will discusshasemployeda varietyof formalizations(at various

levelsof complexity). As our goal in thesesectionsis to discernthesemanticmechanismsunder-

lying adverbial function, we will not be advocatinga particularformalization,but will presenta

varietyasrequiredto representtheparticularpropertieswe arefocusingon in eachof thesemantic

mechanismswe discuss.18As will be discussedin Chapter5, however, thereare other ways apart from their semanticsthat both typesof

adverbialcanevoke contextual AO interpretations.

108

Page 126: Thesis about discourse

We alsoemphasizethat we have distinguishedclausalanddiscourseadverbialssemantically

basedonly onthenumberandinterpretationof theargumentsthey contain.It mayor maynotprove

useful in practicalapplicationto further distinguishclausalanddiscourseadverbialsaccordingto

the resolutionof their arguments.For, asnotedabove, andwe will seein the next two sections,

AOs aremadeavailable by both NPs and non-NPconstituents(as well as by language-external

context). Thus,if anAO is nominalizedin thediscourse,thedefinitesemanticargumentof a dis-

courseadverbialmayresolve to thatnominalization.A strongerdistinctionbetweendiscourseand

clausaladverbialswould assertthatan adverbial functionssemanticallyasa discourseconnective

if andonly if the interpretationof oneor moreof thesemanticargumentsit containsis dependent

on anAO thatis retrieved,or reifed([Web91]) from a non-NPconstituent.Sucha distinctionmight

be madeundertheassumption,for example,that NP-referenceis distinguishedfrom referenceto

non-NPconstitutentsin anaphoraresolutionalgorithms.Wewill discussthis issuein moredetailin

Chapter6.

3.4.6 Summary

In this section,we have presenteda numberof analysesproposedat the clauselevel to account

for the semanticinterpretationof adverbial modification. We have argued,andwill show in the

next section,that suchclause-level semantictools canbe extendedto accountfor commonalities

betweenclausalanddiscourseadverbials,while alsodistinguishingsemanticdifferencesbetween

them. In particular, we have arguedthat discoursedeictic researchprovidesa betterbasisfor the

classificationof thesemanticargumentsof adverbialsthanclause-level research,andhavepresented

our formaldefinitionsof thesemanticdistinctionbetweenclausaladverbials,whichdonot function

semanticallyasdiscourseconnectives,anddiscourseadverbials,whichdo functionsemanticallyas

discourseconnectives.Wehave definedthisdistinctionaccordingthenumberandinterpretationof

semanticargumentstheseadverbialsapplyto:

discourseadverbials: containat leastoneargumentthatdependsfor its interpretationonsome

salientAO containedin or derivable from the discoursecontext; their interpretationis context-

dependent.

109

Page 127: Thesis about discourse

clausaladverbials: containno suchargument;their interpretationis context-independentafter

resolutionof any semanticargumentsto contextual concreteobjects.

We view thesedefinitions,andour discussionin thenext two sectionsof thesemanticmecha-

nismsthat motivatethem,asa startingpoint, providing a foundationthat is generalenoughto be

supplementedby avarietyof semanticformalisms.In Chapter4 wewill selectaworkingformalism

to show how thesemanticsdevelopedin thischaptercanbeincorporatedinto theDLTAG model.In

Chapter5, we will seeotherwaysanadverbialcancontribute to discoursecoherence.

3.5 S-Modifying PPAdverbials

As discussedin Section3.2, thesyntacticstructureof theS-modifyingPPadverbialsin our corpus

canbe respresentedwith the treein Figure3.5, whereP representstheprepositionhead,S repre-

sentstheexternalargument,Arg�6�� representstheinternalargument,andSPECrepresentsoptional

(specifiers)modifiersof thehead19.

SÍ Í ÍÎÎÎPPÍ Í ÍÝÝÎÎÎ

Spec P Arg�6�* S

Figure3.5: SyntacticStructureof S-Modifying PPAdverbials

Both syntacticallyandsemantically, therefore,all theprepositionalphrasesin our datasetare

binary predicates.As illustration, they canall be representedsemanticallyusinglambdacalculus

(c.f. [HK98]) as in (3.71), where“[[ ]]” representsan interpretationfunction, and preposition

representstherelationsuppliedby theprepositionto its arguments,� and � , wherewe resolve � to

the interpretationof the internalargument,and � to the interpretationof theexternalS argument.

For example,if m representstheinterpretationof theNPMary, thentheinterpretationof thePPfor

19Althoughasdiscussedin Seciton3.2,internalargumentsarenotalwaysfoundin PPs,only a few PPsin ourdatadidnot take an internalargument.Thesewereeithermisparsedsubordinatingconjunctionsor topicalizedverbconstituents(e.g.“on they cameout” Þ “they cameonout”.

110

Page 128: Thesis about discourse

Mary canrepresentedasin (3.72),andif is(w,l) representsthe interpretationof theS work is life,

thenthefinal interpretationof theS-adjoinedclausaladverbialcanberepresentedasin (3.73).

(3.71)[[PP]] = ß x ß y.preposition(x, y)

(3.72)[[for Mary]] = ß y.for(m, y)

(3.73)[[For Mary, work is life]] = for(m, is(w,l))

Thus,thepredicate-argumentstructureof thePPadverbialdoesnotdistinguishclausalanddis-

courseadverbials.However, if we exchangeMary for that reason, weproducethediscourseadver-

bial for that reason. In orderto distinguishclausalanddiscoursePPadverbials,therefore,we must

investigatethesemanticmechanismsunderlyingthepredicateargumentstructureandinterpretation

of theinternalargumentsof thePPadverbialsin ourcorpus.

3.5.1 Proper Nouns,Possessives,and Pronouns

Approximatecountsin our corpusof typesandtokensof the internalargumentsdiscussedin this

sectionareshown in Table3.10.

Table3.10:ApproximateCountsof TokensandTypesof someInternalPParguments

# Tokens # Types Inter nal Ar gument

776 428 propernounsandyears307 273 singlenounsmodifiedby possessives192 68 pronouns

As shown, a numberof PPinternalargumentsin our datasetarepropernounsor years.Table

3.11providessomeexamplesalongwith their corpuscounts;themajority occurredonly once.

Table3.11:PPAdverbialswith ProperNounor YearInternalArgument

# PP Adverbial # PPAdverbial

2 after1832 8 in Tokyo2 by God 30 onFriday2 for Blanche 1 to Africa24 in August 1 until 1971

111

Page 129: Thesis about discourse

Although propernounsandyearsmay be interpretedwith respectto the discourseor spatio-

temporalcontext, they do not refer to abstractobject interpretations;rather, they denotepeople,

places,animals,etc.,which we have calledconcreteobjects.Their semanticinterpretationcanbe

representedasexemplifiedin (3.74),wherebold-facerepresentsthedenotedconcreteobject.

(3.74)[[God]] = God

Someof theinternalargumentsin our corpusaresinglenounsmodifiedby a possessive proper

noun;Table3.12providessomeexamplesalongwith their counts.

Table3.12:PPAdverbialwith Possessive ProperNounInternalArgument

# PPAdverbial # PP Adverbial

1 despiteBerger’s report 1 to Ann’s consternation2 for God’s sake 1 to Welch’schagrin1 in Blanche’s defense 1 underYakov Segal’sdirection1 in Krutch’s view 1 with Herberet’s blessing1 in Plato’s judgment 1 within Erikson’s schema

As shown, we’ve providedcasesin which theheadnoundenotesanabstractobject.For exam-

ple, views, judgements,consternation,blessingareall abstractobjects.However, themodification

by possessive propernounsdoesnot make theseabstractobjectscontext-dependent;the seman-

tic representationof theseNPsis akin to therepresentationin (3.74).Of course,theseNPsmaybe

coindexedwith abstractobjectsin theprior discourse,asshown in (3.75),whereHerberet’sblessing

canbeinterpretedastheactionof Herberetraisinghishandandpraying.

(3.75) Herberetraisedhis handsandbeganto pray. Mike andMary knelt beforehim. They

lookedinto eachother’s eyesandsmiled.With Herberet’s blessing, they wouldbemarried.

A numberof the PP internal argumentsin our dataset arepronounsor areadverbsthat are

functioningaspronouns(e.g.now, then,here). Table3.13providessomeexamplesalongwith their

corpuscounts. As shown, many of thesepronounsareanimatepronouns,by which we meanthe

first andsecondpersonpronouns,andthird personpronounsthatreferto animateentities.Although

animatepronounsdo not refer to abstractobjectinterpretations,they areanaphoricor deictic,and

112

Page 130: Thesis about discourse

mustbeinterpretedwith respectto NPspresentin (or inferablefrom [Pri81]) thediscourseor spatio-

temporalcontext. In [HK98], thesemanticinterpretationof apronounis representedasdenotingan

entity . via anindex + thatis mappedto . relative to anassignmentfunction j , wherej is determined

by acontext i . An exampleis shown in (3.76).

(3.76)[[you� ]] ��à = a3 (i), wherei might provide theassignmentfor + : j 3 =

}~~� 1 K Kim

2 K John

3 K Sandy

�&���Table3.13:PPAdverbialswith PronominalInternalArgument

# PPAdverbial # PPAdverbial # PP Adverbial

1 above me 4 for now 8 to me14 afterthat 2 from here 28 sincethen2 amongthem 2 from this 8 until recently5 beyondthat 4 in it 5 until then6 by now 4 in this 2 with that4 for me 1 like you 1 with them

Not all pronounsonly referto concreteobjects,however. As Table3.13shows, therearenumer-

ouspronominalinternalPPargumentsfoundin ourcorpusthatmaybediscoursedeicticandreferto

anabstractobjectinterpretation.Thesearecasesof demonstrative pronouns,inanimatethird person

pronounsit20, or adverbsfunctioningaspronouns.Demonstrative referenceto abstractobjectswas

discussedin Chapter2; we will discuss(discourse)deictic adverbsin Section3.6. The semantic

interpretationof thesepronominalscanberepresentedin thesameway asotherpronouns,e.g. as

denotingan entity . relative to a context-determinedassignmentfunction. The differenceis that

herethe context may determinethat + is assignedto an abstractobject interpretationof eitheran

NP or a non-NPconstituent.Thattheseinternalargumentscanreify abstractobjectsfrom non-NP

constituentsis shown in (3.77)-(3.78),whereboth that andthenrefer to theevent interpretationof

thefirst sentence.Notethepotentialfor ambiguity, however. In (3.79),thencanrefereitherto the

NP themorningor to theeventof wakingearly.

20Recallfrom Chapter2 however thatunlessthepredicationon it is sufficiently semantically-enhanced([Byr00]) (andthis is notprovidedby a preposition),anabstractobjectit refersto musthave alreadybeenreferredto with anNP.

113

Page 131: Thesis about discourse

(3.77)I wentto themovies.After that, I ranahundrederrands.

(3.78)I wentto themovies.Sincethen, I’ve runahundrederrands.

(3.79)In themorning,I woke up early. Sincethen, I’ve runahundrederrands.

Therearealsoa numberof casesin our corpuswherethe internalargumentis a singlenoun

modifiedby apossessive pronoun.Someexamplesareshown in Table3.14.Again,we’veprovided

casesin which theinterpretationof themodifiednounsis anabstractobject.For example,natures,

opinions,knowledge areabstractobjects.Moreover, possessive pronounsareanaphoricor deictic,

but they referto concreteobjects.Therefore,likepossessive propernouns,theseNPsmay, but need

not,becoindexedwith abstractobjectsin theprior discourse.

Table3.14:PPAdverbialwith Possessive Pronoun

# PP Adverbial # PP Adverbial

1 by its nature 2 in my opinion1 beforehis departure 2 in my view1 despitehis yearning 1 to his surprise3 for his part 2 to my knowledge1 in ourcase 1 underhissupervision

3.5.2 Demonstrativeand Definite Determiners

Approximatecountsin our corpusof typesandtokensof the internalargumentsdiscussedin this

sectionareshown in Table3.15.

Table3.15:ApproximateCountsof TokensandTypesof someInternalPParguments

# Tokens # Types Inter nal Ar gument

468 266 singlenounsmodifiedby definitearticle310 180 singlenounsmodifiedby demonstrative determiner

Numerousinstancesof singlenounsmodifiedby thedefinitearticle (the) arefound asthe in-

ternalargumentof the PPadverbialsin our dataset. In somecases,thesenounsdenoteconcrete

objects;Table3.16providessomeexamplesalongwith their corpuscounts. In many cases,how-

114

Page 132: Thesis about discourse

ever, thesenounsdenoteabstractobjects.Table3.17providessomeexamplesof definiteabstract

objectsalongwith their counts. Herewe seenumerousnovel examplesof abstractobjects,such

ascriticism, record, evidence. We alsoseethat somenounshave oneinterpretationasa concrete

object and anothermetaphoricalinterpretationas an AO, suchas board. We also seea deictic

noun,past(presentalsooccursin ourcorpus),which is alwaysinterpretedwith respectto eitherthe

spatio-temporalcontext or thediscoursetime.

Table3.16:PPAdverbialswith DefiniteConcreteObjectInternalArgument

# PPAdverbial # PP Adverbial

1 above thetongue 2 in thecity3 at thedoor 1 insidethecourtroom1 below thefort 1 nearthecoast1 beyondtheforest 1 sincethehurricane1 down theboulevard 1 towardthewest1 for theboy 1 within theindividual

Table3.17:PPAdverbialswith DefiniteAO InternalArgument

# PPAdverbial # PP Adverbial

1 acrosstheboard 1 for therecord2 afterthepayment 1 from theevidence2 afterthesplit 20 in theend6 alongtheway 14 in themeantime2 at theclose 18 in thepast11 at themoment 6 in theprocess17 at thetime 12 on thecontrary2 by theway 3 on thesurface1 despitethecriticism 4 on theway2 duringthetrial 7 undertheagreement4 for themoment 1 with theincrease

Semantically, definitenounsdenotea specificknown entity. However, definitenounsarenot

necessarilyanaphoricto somethingsalientin the prior discourseor spatio-temporalcontext. For

example,in (3.80),thetrial is inferablefrom thereifiedAO interpretationof thesuingevent.But in

(3.81),therecord simply refersto anabstractnotionof right andwrong.

115

Page 133: Thesis about discourse

(3.80)JohnsuedMary for all shewasworth. During thetrial , hecriedprofusely.

(3.81)For therecord, graduatestudentsdon’t getpaidenough.

The definingsemanticfeatureof definitedescriptionsis that they areusedwhenonly oneen-

tity correspondsto their description([HK98]). Thus“the king of America” is infelicitousbecause

thereis no correspondingentity. [HK98] views this failureto referasa presuppositionfailure,and

representsdefinitedeterminersaspartial functionswhosedomaincontainsonly thosenounsthat

correspondto oneentity in thesetof individuals,andwhoserangecontainsthedenotationsof those

nouns.As noted,adefinitenounmayor maynotbeanaphoric,however, in thesensethatit recovers

anentity in theprior discourse.

Therearenumerousinstancesof singlenounsmodifiedby demonstrative determinersfound

asthe internalargumentof the PPadverbialsin our dataset. In somecases,thesenounsdenote

concreteobjects;Table3.18providessomeexamplesalongwith their counts.

Table3.18:PPAdverbialswith Demonstrative ConcreteObjectInternalArgument

# PPAdverbial # PPAdverbial

1 above thesejobs 2 in thisarticle2 at thatprice 2 in thisplay1 at theseoffices 2 of thatamount1 in thesefamilies 1 on thesegenerators1 in theseorganizations 1 to thesepeople

In thegreatmajority of cases,however, thesenounsdenoteabstractobjects.Becausemany of

thesenounsappearwith avarietyof demonstratives,we have conflatedthemin Table3.19in order

to presentawider variety. Herewe see“basic” AOssuchasevents,situations, andfacts. However,

we alsoseebackdrop andcircumstances, which canbesituations,studyandservice, which canbe

events,andbasisandreasonwhichcanbefacts.

Demonstrative NPs, like demonstrative pronouns,areanaphoricor deictic andmustbe inter-

pretedwith respectto the discourseor spatio-temporalcontext. They can either be represented

semanticallyakin to otheranaphoricreference,e.g.via anassignmentfunction,or they canberep-

resentedusingpartial functions,akin to definitedescriptions.And asis alsotruefor demonstrative

116

Page 134: Thesis about discourse

pronouns,thecontext maydeterminethat they reify abstractobjectinterpretationsof VP or S. Ex-

amplesareshown in (3.82)-(3.83),wherethatserviceandthat reasonreferto theAO interpretation

of the first sentence.Again, however, the demonstrative NP’s referentmay be a previously men-

tionedabstractobjectNP, asin (3.84),or it maybeambiguous,asin (3.85),wherethat reasoncan

refer to thepreviously mentionedNP a reason, or it canrefer to theAO interpretationof thefirst

sentence.

(3.82)JohnhelpedMary washthecar. After that service, shepaidhim $40.

(3.83)Johncouldn’t sleep.For that reason, hegotoutof bed.

(3.84)Yesterdayyougavemeagoodreasonto move. For that reason, I thankyou.

(3.85)It wasn’t until yesterdaythatyou told meyour reasonfor leaving. For that reason, I am

madat you.

Table3.19:PPAdverbialswith Demonstrative AO InternalArguments

# PP Adverbial # PP Adverbial

1 afterthatservice 2 in thesecircumstances2 againstthat/thisbackdrop 5 in thisconnection1 alongtheselines 2 in thatevent21 at that/thispoint 1 in thatfunction2 at this stage 3 in this instance1 by thatlogic 4 in thismanner2 by that/thesemeasure(s) 8 in this/theserespect(s)1 by thisstandard 2 in thissense13 by that/thistime 12 in that/thisway1 despitethesechallenges 3 on this basis1 despitethesefacts 1 outsidethoselimits1 duringthisstudy 5 to that/thisend4 for thatmatter 2 underthisplan5 for that/thisreason 1 with thissituation25 in that/this/these/those case(s) 2 within that/thisframework

3.5.3 Indefinite Articles, Generic and Plural Nouns,and Optional Ar guments

Sofarwehaveexaminedhow explicit AO referencecreatesPPdiscourseadverbials.In thissection

weusesinglegeneric,pluralandindefinitesinglenounsto demonstrateothersemanticmechanisms

117

Page 135: Thesis about discourse

at work in the internalargumentsof PPdiscourseadverbials.Approximatecorpuscountsof these

PPinternalargumentsareshown in Table3.20. This analysisalsoappliesto someof the internal

argumentsdiscussedabove; we have simply ignoredthisaspectuntil now.

Table3.20:ApproximateCountsof TokensandTypesof someInternalPParguments

# Tokens # Types Inter nal Ar gument

229 68 singlenounsmodifiedby indefinitearticles1020 233 singlegenericandpluralnouns

Therearenumerousinstancesof singlenounsmodifiedby indefinitearticlesfound asthe in-

ternalargumentof the PPadverbialsin our dataset. In somecases,thesenounsdenoteconcrete

objects;Table3.21providessomeexamplesalongwith their counts.

Table3.21:PPAdverbialwith IndefiniteConcreteObjectInternalArgument

# PPAdverbial # PPAdverbial

1 aftera roundup 1 in asaucepan1 asaboy 1 to astranger3 asagroup 1 underamicroscope1 for ananthropologist 1 with abellow

In many othercases,thesenounsdenoteabstractobjects.Table3.22providessomeexamples

alongwith their counts.Againwe seenumerousnovel AOs,includingrule, quirk, sense, etc.

Table3.22:PPAdverbialwith IndefiniteAO InternalArgument

# PP Adverbial # PP Adverbial

2 asa rule 1 in a fashion1 by a quirk 4 in asense18 in astatement 4 in away10 for amoment 1 on animpulse7 for awhile 1 to adegree

The semanticsof indefinitesis the subjectof much current research(c.f. [Roo95b, vdB96,

HK98, Hei82]). In [HK98] indefinitearticlesarerepresentedastotal functions,e.g. they placeno

118

Page 136: Thesis about discourse

requirementson the nounsin their domain. More generally, theseindefinitenounscanbe repre-

sentedasunarypredicatesdenotinganunspecifiedentity; in predicatelogic this is representedasin

(3.86).

(3.86)[[an impulse]]= h x.impulse(x)

Theindefinitenessof thesenouns,whetherthey denoteconcreteobjectsor abstractobjects,does

not causethemto resolve or refer to entitiesin theprior discourse.This not to saythat certainof

theadverbialsin Table3.22arenotsometimestreatedasdiscourseconnectives;wearguehereonly

thattheir semanticinterpretationdoesnot requireanabstractobjectin theprior discourse.We will

returnto this issuein Chapter5.

Not all indefinitenounsareunarypredicates,however. Certainof the indefinitenounsin our

corpusare relational nouns,which take a syntacticallyoptional,or hidden, or implicit, argument

thatis anaphoricto somesaliententity in theprior discourseor spatio-temporalcontext. Examples

areshown alongin thefirst columnof Table3.23. In thesecondcolumntheseargumentsaremade

overt with ademonstrative.

Table3.23:PPAdverbialwith RelationalIndefiniteAO InternalArgument

# PP Adverbial Explicit Ar gument

2 asanalternative asanalternative to that2 asaconsequence asaconsequenceof that1 asa restatement asa restatementof that84 asa result asa resultof that1 for anexample for anexampleof that

We introducedsyntacticallyoptionalargumentsin Section3.4,usingexamplessuchas(3.50).

(3.50a) Theduedatefor the grant haspassed.Mary didn’t apply for it.

(3.50b) Theduedatefor thegranthaspassed.Mary didn’t apply G .

(3.50c) *Mary didn’t apply G .

Intuitively, whenoneapplies, they apply for something; in (3.50 a), this (for it) argumentis

overtandanaphoricto thegrant in theprior sentence,while in (3.50b), it is implicit but retrievable

from thecontext. In (3.50c) it is not syntacticallyoptionalbecausethecontext doesn’t supplyit.

119

Page 137: Thesis about discourse

Similarly, if somethingis a result,consequence, or restatement, etc.,it is a result,consequence,or

restatementof somespecificsomething. Thus,we canmake this semanticargumentovert, as in

(3.87a),or leave it implicit, asin (3.87b), becauseit is retrievablefrom thecontext. However, the

discourseis infelicitousin (3.87c) becausethecontext doesn’t supplyit.

(3.87a) Theduedatefor thegranthaspassed.As a resultof that, Mary didn’t applyfor it.

(3.87b) Theduedatefor thegranthaspassed.As a result G , Mary didn’t applyfor it.

(3.87c) *As a result,Mary didn’t applyfor agrant.

Thesenounsarethusbinary predicates.Othernounsarealsoviewed asbinary predicates,as

discussedin Section3.4, including part (of something)andmother(of somebody). In predicate

logic, suchbinarypredicatenounscanberepresentedasshown in (3.88).

(3.88)[[result]] = h x h y[result(y,x)]

What predicatelogic cannotrepresent,however, is whetheror not this hiddenargumenthas

to be resolved in the prior discourse.As discussedin Section3.4, implicit argumentshave been

distinguishedinto two types: definiteandindefinite. As exemplifiedabove in (3.50), the implicit

argumentof apply is definite;it mustbeanaphoricto somethingin theprior context in orderto be

interpreted.In contrast,the implicit argumentof mother is indefinite; in Section3.4, (3.53) was

usedto exemplify indefinitehiddenarguments,which arenot necessarilyanaphoricwith anything;

their interpretationcanbeindependentof context.

(3.53)Today, Mary talkedto amother.

Althoughtheverbdonatein Section3.4(example3.54)indicatesthatsomeimplicit arguments

caneitherbe definiteor indefinitedependingon context, the implicit argumentsof the indefinite

nounsin Table3.23appearto alwaysbedefinite;at leastwhenfoundasthe internalargumentsof

thesePPadverbials,theseindefinitenounscannotbeinterpretedindependentlyof context andthus

the containingPPadverbialscannever be discourse-initial.Moreover, many of thesearguments,

implicit or explicit, appearto beabstractobjects.For example,a resultis a resultof acause, which

is anabstractobject.A restatementis a restatementof astatement, which is alsoanabstractobject.

Recallhoweverthediscussionof AOsin Chapter2,wherewenotedVendler’sobservationthatsome

120

Page 138: Thesis about discourse

concreteobjectssuchasfire, blizzard canbeinterpretedascauses.

Of course,it canbedifficult to distinguishwhethera particularcasetakesa hiddenargument,

or merelyapotentialadjunct,especiallywhenthisargumentappearsto beindefinite.This is trueof

many of the indefinitenounsshown in Table3.22. For examplea sense, foundin thePPadverbial

in a sensein Table3.22, is likely interpretedasa senseof something. This somethingis clearly

indefinite,however, like thesomebodyargumentof mother, andsoit doesnot causethecontaining

PPadverbial to be uninterpretablediscourse-initially. More generally, we canusethe tests(e.g.

entailmentandpresupposition)discussedin Section3.4 that have beenusedto distinguishverbal

argumentsandadjuncts,but thesetestsarenot foolproof.

As onemight expect,thesamevariability in form we have seendisplayedby theinternalargu-

mentof ourPPadverbialscanbedisplayedby thesesyntacticallyoptionalarguments.For example,

in Table3.23,we madetheargumentexplicit with a demonstrative, but this argumentcanalsoap-

pearasa full nounphrase,andif thenounphraseis not anaphoric,asin (3.89),no prior context is

neededto interprettheadverbial.

(3.89)As aconsequenceof war, peopledie.

Furthermore,thesamevariability in resolutionwehaveseendisplayedby theinternalarguments

of our PPadverbialscanbe displayedby thesesyntacticallyoptionalarguments.For example,in

(3.90a), theinterpretationof thedemonstrative resolvesto theabstractobjectinterpretationof the

VP, e.g.giving him a book, but in (3.90b), theinterpretationof thedemonstrative is ambiguous;it

canresolve to givinghima bookor to theNP a book.

(3.90a) If Johnis bored,give him abook.As analternative to that,take him to thezoo.

(3.90b) If Johnis bored,give him abook.As analternative to that,give him amagazine.

Thereis a distinction,however, betweenthesebinarynounsthatwe have sofar ignored.While

thehiddenargumentsof result,consequence, restatementcanresolve to singleAO interpretations,

thehiddenargumentof examplemustresolve to a setof interpretations.This propertyof example

is notedin [WJSK03]. For example,in (3.91a), this argumentresolvesto thesetof consequences

of Johnbreakinghis arm,asis madeovert in (3.91b).

121

Page 139: Thesis about discourse

(3.91a) Johnjustbroke his arm.So,for anexample,hecan’t cycle to work now.

(3.91b) Johnjust broke his arm. So, for anexampleof theconsequencesof breakinganarm,

hecan’t cycle to work now.

Thisanalysisreadilyextendsitself to genericandpluralnouns,whicharequitefrequentlyfound

in our corpusasinternalargumentsof PPadverbials.Again, thesenounsmaybeconcreteobjects;

Table3.24providessomeexamples,alongwith their counts.

Table3.24:PPAdverbialswith Genericor PluralConcreteObjectInternalArguments

# PP Adverbial # PPAdverbial

1 amongprofessionals 1 in academia2 asartists 1 like lemmings1 below decks 1 to libertarians2 for corporations 1 within institutions

Our corpusalsocontainsmany examplesof plural or genericnounswhich denoteabstractob-

jects;Table3.25providessomeexamplesalongwith their counts.

Table3.25:PPAdverbialswith Genericor PluralAO InternalArguments

# PP Adverbial # PP Adverbial

6 at times 1 in emergencies3 by law 1 in truth105 in fact 1 on occasion4 in practice 1 on reflection2 in reality 2 over time1 in theory 2 without question

Like indefinites,the semanticsof genericandplural nounsis the subjectof currentresearch

(c.f. [Roo95b, vdB96, HK98]). Generally, thesegenericnounscanbe representedeitherakin to

indefinites,e.g. asunarypredicatesdenotinganunspecifiedentity, or akin to plural nouns,which

canberepresentedasunarypredicatesdenotinganunspecifiedsetof entities,asshown in (3.92)21.

(3.92)[[professionals]]= e x.professional(x)

21Thequantifiersuggestsiteration,which is notalwaysthecase.

122

Page 140: Thesis about discourse

Whetherthey denoteconcreteobjectsor abstractobjects,thegeneric-nessandplurality of these

nounsdoesnot causethemto retrieve entitiesin theprior discourse.Again,however, certainof the

adverbialsin Table3.22 (e.g. in fact, of course) arefrequentlyviewed asdiscourseconnectives;

asstatedabove, we arguehereonly that their semanticinterpretationdoesnot requirean abstract

objectin theprior discourse,andwill returnto this issuein Chapter5.

However, aswealsofoundfor indefinitenouns,certaingenericandpluralnounsarerelational,

andappearto take a syntacticallyoptional,or hidden, or implicit, argumentthat is anaphoricto

somesaliententity in theprior discourseor spatio-temporalcontext. Therearemany examplesof

suchgenericnounsfoundastheinternalargumentsof PPadverbialsin ourcorpus.Someexamples

areshown in Table3.26.

Table3.26:PPAdverbialswith RelationalGenericAO InternalArguments

# PP Adverbial # PP Adverbial

1 asevidence 3 in effect1 by analogy 3 in essence10 by comparison 1 in exchange28 by contrast 3 in part167 for example 6 in response70 for instance 8 in return204 in addition 5 in sum1 in comparison 117 of course2 in conclusion 3 on average1 in consequence 3 on balance16 in contrast 1 on reflection

Again, it canbe difficult to distinguishwhethera particularcasetakesa hiddenargument,or

merelya potentialadjunct.It appearsthatall of the(assumed)hiddenargumentsin Table3.26can

resolve to the AO interpretationof a VP or S in the prior discourse.For example,in (3.93), the

implicit (or explicit demonstrative) argumentresolvesnot to oneof theentitiesJohn,movies,Mary,

but to theactionof Johngoingto themovieswith Mary.

(3.93)Johnwentto themovieswith Mary. In exchange(for that),shegave him abackrub.

However, not all of theseargumentsareclearlydefinite. As in Section3.4,wherewe saw that

123

Page 141: Thesis about discourse

thedefinitenessof theoptionalargumentof donatewasto someextentdependentoncontext, sodo

we seea similar dependency for on average andin essence. For while anaverage is anaverageof

(or over) something, andan essenceis an essenceof something, only in (3.94a) and(3.95a) are

these“somethings”anaphoricto anAO interpretationin theprior context. In (3.94b) and(3.95b)

these“somethings”areinterpretedindependentlyof context.

(3.94a) Wechoosetheactors,we build thesets,andwe keepthebooks.In essence, we run the

show. (shortenedWSJexample)

(3.94b) In essence,I amahappy person.

(3.95a) Mike washesthe dishes.Mary driesthemandputsthemaway. On average, they do

aboutthesameamountof work.

(3.95b) Onaverage,Johnis ahappy person.

Furthermore,aswe saw with indefinites,many of thesyntacticallyoptionalargumentsin Table

3.26 canresolve to a set, including the optionalargumentsof instance, part, sum,average, and

essence. Wewill discusssetsin greaterdetail in Section3.6andChapter5.

3.5.4 PP and ADJP Modifiers

Sofarwehaveonly seenactualcorpusexamplesof singlenouninternalarguments.Modifying these

internalargumentswith PPsandADJPscanalsocreatediscourseadverbials. Table3.27presents

approximatecountsin our corpuswherethe internalPPargumentis a singlenounmodifiedby a

singlePPor ADJ.

Table3.27:ApproximateCountsof TokensandTypesof someInternalArgumentModifiers

# Tokens # Types Modifier

1491 1400 singlenounwith PPmodifier1926 1143 singlenounwith ADJ modifier

Clearly, PPmodifiersof theinternalargumentsof PPadverbialscancreatediscourseadverbials

becausePPmodifiersthemselvescontainan internalargument,andit canbeanalyzedin thesame

way that we have alreadybeenanalyzingthe single noun internal argumentsof PP adverbials.

124

Page 142: Thesis about discourse

As a particularlyrelevant example,in Tables(3.29)-(3.28),we provide corpuscasesin which the

syntacticallyoptionalsemanticargumentsof binarydefinite,indefinite,andpluralor genericnouns,

respectively, aremadesyntacticallyovert by aPP.

Table3.28:BinaryDefiniteInternalArgumentwith Overt Argument

# PP Adverbial

1 at thebeginningof thehippodrome1 at theendof theday1 at theinsistenceof Arturo Toscanini1 at theoutsetof his career1 in thecaseof academicpersonnel1 in themidstof it all1 in thecourseof this1 on thebasisof this carefulreading1 sincethestartof thedecade1 with theexceptionof satiresof circumstance

Table3.29:Binary IndefiniteInternalArgumentwith Overt Argument

# PP Adverbial

1 asanevocationof timepast1 asanexampleof this lastfacet1 asanillustrationof theprincipleof simplicity1 asanindicatorof thetight grainsupplysituationin theU.S.1 asanintroductionto Americanpolitics8 asamatterof fact1 asapartof overall efforts to reducespending1 asa resultof thatattitude1 in aseriesof fairy talesandfantasies

The NPswithin thesesyntacticallyoptionalPPsagaindisplaya variety of novel abstractob-

jects,conveyed in a varietyof forms. For example,in theabove two tableswe find demonstrative

pronouns(e.g. this in in thecourseof this), demonstrative nouns(e.g. that attitudein asa resultof

thatattitude), genericnouns(e.g.simplicity in theprincipleof simplicity in asan illustration of the

principle of simplicity), etc. As alreadynotedabove in constructedexamples,makingthesyntacti-

cally optionalargumentof arelationalnounovertmaymovetheburden(of classifyingtheadverbial

125

Page 143: Thesis about discourse

containingit asa discourseor clausaladverbial) to theinterpretationof this overt argument.More

generally, regardlessof whetheror notaPPmodifieris asemanticargumentof thenounit modifies,

it introducesadditionalentitiesthatmayor maynot referto AOsin theprior discourse.

Table3.30:BinaryGenericor PluralInternalArgumentwith OvertArgument

# PPAdverbial

5 aspartof theagreement1 by meansof this socialcontrol1 in accordancewith legislationpassedat thelastsessionof Congress1 in additionto freemassages1 in anticipationof thatshift1 in caseof adeadlockbetweenprisonboardsandinmates1 in connectionwith thisconference1 in continuationof thesetheoreticalstudies1 in contrastto all this1 in exchangefor higherpricesupports1 in light of all this4 in point of fact1 in reactionto proposedcapital-gainslegislation1 in responseto this3 in spiteof this1 in termsof volume2 on topof all this

Adjectival modificationof aninternalargumentcanalsocauseit to beinterpretedwith respect

to discourseor spatio-temporalcontext. Thesimplestcasesin which to seethis arePPadverbials

whoseinternalargumentis modifiedby an adjective whoseinterpretationdependson the spatio-

temporalpropertiesof eitherthe discourseor the context. Someexamplesareshown Table3.31,

modifying anounthatcanbeinterpretedasanabstractobject.

Table3.31: InternalArgumentwith aSpatio-TemporalADJ

# PP Adverbial # PP Adverbial

3 undercurrentrules 1 for presentpurposes22 in recentyears 1 undermodernconditions

Therearemany otherdifferentclassesof adjectivesin our corpus;adjectivesareasvaried(and

126

Page 144: Thesis about discourse

asdifficult to imposeclassificationon) asadverbs.Onemajorfunctionalclassfoundin our corpus

includesadjectives([MW]) thatcanfunctionasdeterminers,includingdistributives(all, both,either,

each, every...), cardinalnumbers(one, two...), otherquantifiers(few, some, any...), and“dif ference

words” ([gra]) or “alternative phrases”([Bie01]) (another, other, such, same).

Somedistributivescausetheir associatednounto beinterpretedwith respectto anentity in the

prior discourse.Whenthis associatednoun is an abstractobject, it is interpretedwith respectto

an AO in the prior discourse.Someexamplesof PPadverbialscontainingsuchdistributives and

anassociatedabstractobjectareshown in Table3.32. In general,nounsmodifiedby distributives

referto groupsor individualsin groupsfrom in theprior discoursecontext ([gra]). [HK98] usesof

partialfunctionto representthesedistributivessemantically, akin to thedefinitearticle,but wherea

specifiednumberor setof entitiesis denoted.

Table3.32: InternalArgumentwith ReferentialAdjective

# PP Adverbial # PP Adverbial

1 by bothstandards 7 in eachcase5 in bothcases 1 in eachinstance1 in bothrespects 5 in eithercase1 in bothconditions 1 in eitherevent

Otherdistributives, alongwith the cardinalsandquantifiersnotedabove, do not elicit an en-

tity in theprior context for their interpretation.Someexamplesfrom our corpusof PPadverbials

containingthesedeterminersandan associatedabstractobjectareshown in Table3.33. Because

theNP they produceis non-referential,[HK98] usestotal functionsto representthesedeterminers,

suchthatoneandsomeNP denoteanunspecifiedentityakin to indefinitenouns,andtheremainder

denotesetsof unspecifiedentities,akin to plural nouns.Nevertheless,certainof theadverbialsin

thesetableshave beentreatedasdiscourseadverbials,including in anyevent. We will investigate

possiblecausesof this in Chapter5.

Many of thesedeterminerscanfunctionasnounmodifiers;thefirst columnof Table3.34shows

NPscontainingthesemodifiersandadditionaldeterminers.Of course,thesemanticsof thedeter-

miner mustnot contradictthesemanticsof themodifier (e.g. “few onehand”),but mayeffect its

127

Page 145: Thesis about discourse

referentialproperties,asin on theonehand, wherethemaycauseonehand to refer to a situation

(metaphorichand) in the prior context. Leaving the modifiedargumentimplicit asshown in the

secondandthird columnsmay alsoeffect the referentialpropertiesof theseadverbials;someand

onerequirethisargumentto beresolved.Wewill discussotheranalysesof after all in Chapter5.

Table3.33: InternalArgumentwith Non-ReferentialAdjective

# PP Adverbial # PP Adverbial # PP Adverbial

7 at onepoint 1 by all accounts 6 at any rate21 for onething 1 by all means 1 by any measure1 in onecase 4 in all cases 24 in any case1 in onesense 2 in all fairness 13 in any event2 in oneway 1 in all probability 1 in any instance1 in two cases 3 by someestimates 7 in mostcases1 on two occasions 2 for somereason 1 by mostaccounts1 for many reasons 5 for sometime 1 in severalinstances7 in many cases 4 in somecases 5 with few exceptions1 in many instances 4 to someextent 2 in certainrespects2 in many ways 1 in every period 1 undercertaincircumstances

Table3.34: InternalArgumentwith DeterminerandNon-ReferentialAdjective

# PP Adverbial # PP Adverbial # PP Adverbial

1 in a few instances 9 above all 1 to some1 in acertainsense 49 afterall 1 for thosefew7 for themostpart 3 in all 1 at themost8 on theonehand 2 in one 2 for one

Ordinalsandordinal-like adjectivesarealsofound in our corpus,with or without determiners,

asshown in Table3.35. As shown in the third column, their argumentcanbe implicit. We will

discussordinalsfurtherin Section3.6.

“Dif ferencewords”, or “alternative phrases”alsocausetheir associatednounto be interpreted

with respectto anentity in theprior discourse.Whenthis associatednounis anabstractobject,it

thuswill beinterpretedwith respectto anAO interpretationin theprior discourse.Someexamples

of PPadverbialscontainingalternative phrasesandanabstractobjectareshown in Table3.36.

Anotherandother invoke something“dif ferent,remaining,or additional”, thusthey arecalled

128

Page 146: Thesis about discourse

Table3.35: InternalArgumentwith OrdinalAdjective

# PP Adverbial # PP Adverbial # PP Adverbial

1 in afirst step 1 on secondthought 17 at last6 for thefirst time 1 at lastreport 16 at first2 in thefirst instance 1 in thesecondplace 3 in thesecond

Table3.36: InternalArgumentwith Alternative Phrase

# PPAdverbial # PP Adverbial # PPAdverbial

1 in anotherapproach 20 amongotherthings 1 in suchcases1 in anothercase 24 in otherwords 2 in suchcircumstances1 in anotherrespect 2 on othermatters 2 in suchsituations1 in anothersense 1 amongotherissues 1 onsuchoccasions

“dif ferencewords”in [gra]; otherandsuch alsocreatewhatarecalledalternativephrasesin [Bie01].

As determiners,anotheris usedwith singular, andotherwith pluralnouns,but they alsofunctionas

modifiers;thefirst columnof Table3.37showsNPscontainingthesemodifiersandadditionaldeter-

miners.Thesecondcolumnpresentsanotheranaphoric“dif ference”adjective,same, thatrequiresa

determiner. Thethird columnshows thatthemodifiednouncansometimesbeimplicit.

Table3.37: InternalArgumentwith DeterminerandAlternative Phrase

# PP Adverbial # PP Adverbial # PP Adverbial

49 on theotherhand 71 at thesametime 1 for another1 in many otherinstances 2 in thesameway 2 on theother

[WJSK03, Bie01, Mod01] discussthesemanticinterpretationor resolutionof other, such, and

otheralternative phrasesin detail.They views theform otherX asa lexical anaphorwhichrefersto

theresultof excludinganentity or setof entitiesfrom a contextually relevantset,andpresupposes

that theexcludedentity or entitiesbelongto thatset. For example,manyother instancesrefersto

thesetof instancesthatresultfrom excludingone(or more)instancein thediscoursecontext from

129

Page 147: Thesis about discourse

alargerpresupposedsetof instances.Theform anotherX (e.g.anotherX) canbetreatedsimilarly;

for example,anotherapproach refersto the approachthat result from excluding one (or more)

approachesin thediscoursecontext from a largerpresupposedsetof approaches.[Bie01] treatsthe

form such X asincluding, ratherthanexcluding, membersof a set. In his terms,for example,such

situationsrefersto thesetthat resultsfrom usingone(or more)situationsin thediscoursecontext

asan exampleof a presupposedsetof situationsthat alsoincludesthe setreferredto by the such

phrase. The form the sameX appearsto be a direct reference,akin to that X, althoughthereis

somesubtletyafoot in thata new entity canbeintroducedvia thedeterminer, albeit identicalin all

respectsto theoriginal.

[Bie01]’s analysisalso incorporatesthe non-anaphoriccounterparts(e.g. X’s other than/such

as/thesameas Y) of thesephrases.Thesecounterpartsdo not appearin our data,andsoundrel-

atively awkward as internalargumentsof PPs(e.g. in manyother instancesthan this). Bierner,

following [McC88], treatstheseoptionalPPphraseswhich instantiatetheexcludedor includedel-

ementasadjunctsratherthanarguments([Bie01], pg. 28),andusestheAI planningheuristic,“use

existingobjects”[Sac77] to resolve them.In bothBierner’sandour terms,theexcludedor included

anaphoricelementis representedasa hiddenargument. Its interpretationmay be partially deter-

minedby themodifiednoun(e.g. it’s an instancein otherinstances; however, theinterpretationof

casessuchason theother is wholly dependentoncontext, becausethemodifiednounis implicit.

Bierner’s analysisalsoextendsto otheradjectives encounteredin our corpus. Theseare the

comparative andsuperlative adjectives,which by definitionaredependentin an idiosyncraticway

onreferenceto at leastoneotherobjectthatmaybefoundin or inferablefrom in theprior discourse

or spatio-temporalcontext (c.f. [Bie01, WJSK03]) . Someexamplesof abstractobjectsmodified

by theseadjectivesareshown in Table3.38.

Table3.38: InternalArgumentwith Comparative/Superlative Adjective

# PP Adverbial # PP Adverbial # PP Adverbial

1 to betterpurpose 1 with greaterprecision 1 on furtherreflection1 in broadestterms 1 with minor exceptions 2 in theshortrun1 in earlierreshufflings 1 on adeeperplane 1 for smallernewspapers

130

Page 148: Thesis about discourse

Therearesimilarcasesfoundin ourcorpuswherethemodifiedelementis left implicit andthere

is oftenno determiner. Someexampleareshown in Table3.39; thesewill bediscussedfurther in

Section3.6.

Table3.39: InternalArgumentwith OtherSet-Evoking Adjectives

# PPAdverbial # PPAdverbial # PP Adverbial

18 in general 1 in themain 41 at least17 in particular 2 on thewhole 11 asusual3 atbest 1 atworst 13 in short

While thereis amuchmorevariedarrayof adjectivesdisplayedin ourcorpusthanthereis space

to cover here,majorclasseshave beendiscussedwhich causeabstractobjectsto refer to theprior

discourse.Many of theotheradjectivesin ourcorpus(in additionto someof thosediscussedabove),

fall to a greateror lesserextentinto theadverbmodificationtypesdiscussedin Section3.3.

3.5.5 Other Ar guments

S and PPsare also found as internal argumentsof the PP adverbials in our corpus. Moreover,

in addition to internal argumentmodifiers, thereare also adverb modifiersof the P headitself.

The remainderof our corpusconsistsof complex combinationsof the internalargumentsalready

mentioned.Corpuscountsof thesetypesareshown in Table3.40.

Table3.40:ApproximateCountsof TokensandTypesof someInternalPParguments

# Tokens # Types Inter nal Ar gument

692 679 S arguments192 180 PParguments166 166 ADVP modifiers2526 2361 complex combinations

For our purposes,the analysisof PPinternalargumentsis identical to the analysisof the PP

modifiersdescribedabove,with respectto abstractobjectsanddistinguishingdiscourseandclausal

adverbials. However, becauseannotatorsfrequentlyconfusePPadverbialscontainingPPinternal

131

Page 149: Thesis about discourse

argumentswith ADVP adverbials,we will investigatesomeparticularcasesin the next section.

AnnotatorsalsofrequentlyconfusePPadverbialscontainingadverb-modifiedP headswith ADVP

adverbials;thesetoowill bediscussedin thenext section.

Also for ourpurposes,S-modifyingphrasestakingS internalargumentssimplyprovideanother

form in whichtwo AO interpretationsarerelated.However, basedonthePennTreebankI POStag-

ging andbracketingguidelines,all PPadverbialstakingS argumentsshouldnot have beenpresent

in our data. For thoughPennTreebankPOStagsmake no distinctionbetweenprepositionsand

subordinatingconjunctions(both aretaggedasIN), the argumentto the latter is a clauseandthe

entirephraseis bracketedasa subordinatingclause[PT]. However, anexplicit list of subordinating

conjunctionsis not provided. In [Lit98], a subordinatingconjunctionis definedasa linguistic form

that makesa clausea constituentof anotherclause,and the list of subordinatingconjunctionsis

viewedasa researchquestion;[Lit98] identifies107mainentriesand374sensesfor subordinating

conjunctionsin aWebster’s dictionary, includingprepositional,adverb,andverbphrases.

Roughlyhalf of our S argumentscontaineda verb participleand lacked a subject,andwere

parsedasreducedclauses.Someshorterexamplesareshown in Table3.41.

Table3.41:PPAdverbialwith ReducedClauseInternalArgument

# PP # PP

1 afterbeatingthem 1 without sayingso1 asalreadynoted 1 for winning2 asexpected 3 in doingso1 asmight beexpected 1 in sodoing2 aspreviously reported 1 on delvingdeeper

Thesecasesareinterestingbecauseit canbe difficult to decidewhetherto treatsomephrase

asreducedclausesor sub-clausalconstituents,especiallywhenmodifiedby a prepositionthatalso

functionsasa subordinatingconjunction.This is particularlyan issuefor argumentslackingboth

a verbanda subject,suchas“while in hiding” or “as usual”. Moreover, somelexical items,such

asas, canbe both prepositionsandadverbs(e.g. twice as long). Adverbscanmodify verbs,and

thuswe could representphrasessuchasas expectedeitherassubordinatingclausesor asADVP

132

Page 150: Thesis about discourse

adverbials. [Kno96], in fact, treatsthis expressionasan adverbial discourseconnective without

consideringits composition. He doesnot, therefore,addressthe closely-relatedPP, as might be

expected, or its numerousotherpossiblevariations.

TheremainingPPadverbialsin our corpusconstitutecasesthatcombinethemodificationand

argumenttypesdiscussedabove,includingmultiplemodifications,coordinatedphrases,andrelative

clauses,etc.And again,theirsemanticfunctionasdiscourseor clausaladverbialscanbedetermined

accordingto their composition,asdiscussedabove.

3.5.6 Summary

Table3.42:PPAdverbialSummary

Inter nal Ar gumentType PP Adverbial

propernoun in New Yorkanimatepronoun to mepotentialDD afterthatdefiniteAO at thetimedemonstrative AO for thatreasonindefiniteAO withhiddenAO argument asa result(of that)indefiniteAO withhidden � AO � argument asanexample(of that)definiteAO withhidden � AO � argument at thevery least

In thissectionsaw thatthemajority of PPdiscourseadverbialsdo notoccurfrequently. Rather,

therearea few “stock” discourseconnectives,suchasas a result, anda wide variety of otherPP

adverbialswhich containa semanticargumentthatmustberesolvedto anAO interpretationin the

prior discourseor spatio-temporalcontext. We have presenteda wide variety of abstractobjects

found in our corpusof PP adverbials. We have shown that PP clausalanddiscourseadverbials

canbedistinguishedby theinterpretationthesemanticargumentsthey contain.We have presented

a variety of semanticmechanismsthat cancausetheseargumentsto be interpretedwith respect

to the discourseor spatio-temporalcontext. Ignoring the openquestionof small clauseinternal

arguments,asummaryof somebasicmechanismswhichcancausetheseargumentsto referencean

133

Page 151: Thesis about discourse

abstractobjectin thesurroundingcontext areshown in Table3.42.

3.6 S-Modifying ADVP Adverbials

As discussedin Section3.2,andrepresentedin Figure3.6,S-modifyingADVP adverbialsadjointo

anexternalS argumentandarefrequentlycomposedof a singleadverbhead,but mayadditionally

containaninternalargumentandoneor morespecifiers.

SÍ Í Í ÍÎÎÎÎADVP� � � �ÝÝ����

Spec ADV Arg�6�* S

Figure3.6: SyntacticStructureof S-Modifying ADVP Adverbials

In prior sectionswe saw [Moo93]’s useof predicatelogic to representS-modifying ADVP

thattakeonly onesemanticargument;hereweillustratetheinterpretationof suchADVP in lambda

calculus(c.f.[HK98]) asshown in (3.96),usingtheADJ (or NP)derivativeof theadverbto represent

thepropertyadjective they supplyto their (external)semanticargument� . For example,if wise is

taken to representthepropertysuppliedby theadverbwisely, thentheinterpretationof theADVP

wiselycanberepresentedasin (3.97). We resolve � to thesemanticinterpretationof theexternal

S argument. If engage(j,s) is taken to representthe semanticinterpretationof the sentenceJohn

engagedthesafety, thentheinterpretationof theS-adjoinedADVP adverbial is asin (3.98).

(3.96)[[ADVP]] = ß y.adjective(y)

(3.97)[[wisely]] = ß y.wise(y)

(3.98)[[Wisely, Johnengagedthesafety]]= wise(engage(j,s))

If we exchangewise for additional, however, we producethe interpretationof the discourse

adverbialadditionally. In theremainderof this sectionwe focuson how thesemanticmechanisms

underlyingtheinterpretationsof theadverbsin ourcorpusdistinguishADVP clausalanddiscourse

adverbials. We will show that many S-modifying ADVP adverbialstake an additionalsemantic

134

Page 152: Thesis about discourse

argumentandcanberepresentedasbinarypredicates,andwewill alsoseemorecomplex semantic

representationsfor certainadverbs. However, the ADVP adverbialsin our datasetarea motley

crew, dueto thewide variability of adverbs.For example,while many adverbsconstituteanentire

ADVP (e.g.carefully), a largenumberof ADVP adverbialsin ourcorpuscontainadverbsmodified

by anotheradverb(e.g. more carefully) thatcanalsoaffect their interpretation.Moreover, we find

frequenttaggingerrorsin ourADVP corpuswhichserveto illuminatetheinterpretationsof adverbs,

asthey displaypatternsalsocommonto correctlytaggedADVP adverbials.

3.6.1 Syntactically Optional Ar guments

Approximatecountsin ourcorpusof thetokensandtypesof theADVP adverbialsdiscussedin this

sectionareshown in Table3.43.

Table3.43:ApproximateCountsof TokensandTypesof someADVP Adverbials

# of Tokens # of Types Category

85 55 PP-like ADVP with optionalarguments78 18 relationalADVP with optionalarguments

Therearea numberof constructionsin our ADVP corpusthat shouldhave been(and in our

PPcorpushave been)taggedasPPadverbials.Someexamplesareshown in Table3.44. Thefirst

columnlists theinitial preposition-preposition construction(COMB); whenthesecondpreposition

varies,we list it asP. Thesecondandthird columnsindicatethenumberof timesthis combination

wastaggedasPPandADVP, respectively. Thelastcolumnprovidesacorpusexample.

Table3.44:Mis-TaggedPPAdverbials

COMB # PP # ADVP Example

asP 94 3 asfor themeritsbecauseof 12 54 becauseof thisexceptfor 12 1 exceptfor theembarrassmentprior to 4 6 prior to that

TheasP constructionis likely mistaggedasanADVP dueto theuseof theadverbas in other

135

Page 153: Thesis about discourse

constructionsdiscussedbelow. Thebecauseof andexceptfor constructionsmaybemistaggedas

ADVP adverbialsdueto their similarity in form to theADVP constructionshown in Table3.45,in

whichanADVP adverbialtakesanoptionalPP. TheseADVP adverbialswerealsomistaggedasPP

adverbials,asshown.

Table3.45:PP-likeADVP Adverbialswith Overt Arguments

COMB # PP # ADVP Example

insteadof 7 37 insteadof thatratherthan 2 11 ratherthantemptpeopleto buy moreregardlessof 0 5 regardlessof rightsandwrongs

However, asshown in Table3.46, andunlike the PPinternalargumentsof the PPadverbials

shown in Table 3.44, the PP of theseADVP adverbialsare syntacticallyoptional. Although as

shown regardlessdoesn’t occurin ourcorpus(sentence-initially) without its accompanying PP, this

authorrecentlyheardsuchauseonNationalPublicRadio.

Table3.46:PP-like ADVP Adverbialswith HiddenArgument

# ADVP Adverbial

18 instead14 rather0 regardless

TheseADVP adverbialscanberepresentedasbinarypredicates,akin to PPadverbials. How-

ever, recall that [Bie01, McC88] do not treattheoptionalthanPP in other thanphrasesasa com-

plement,but ratherasan adjunct,and[Bie01] thenusestheAI heuristic“useexisting objects”to

resolve thehiddenargumentto theinternalargumentof thePP. While therepresentationof this PP

mayvary dependingon thechosenformalism,we neverthelessarguethat theseADVP adverbials

take a definitehiddensemanticargument. Moreover, the interpretationand/orresolutionof this

hiddenargumentcanbeanconcreteobject,anabstractobjectNP, or anabstractobjectreifiedfrom

a VP or S in theprior discourse.(3.99)- (3.101)show thatalthoughtheargumentcanbe explicit

or implicit (a-b cases)andstill resolve to theAO interpretationof VP or S, themodifiedsentence

136

Page 154: Thesis about discourse

cannotbeinterpretedif theargumentis not resolvablein thediscourseor context ( i cases).

(3.99a) I wantedto go to themovies. Insteadof that,I wentto work.

(3.99b) I wantedto go to themovies. Instead,I wentto work.

(3.99c) *Instead,I wentto work.

(3.100a)Michaelwon’t studybiology. Ratherthanthat,hedreamsaboutit.

(3.100b) Michaelwon’t studybiology. Rather, hedreamsaboutit.

(3.100c) *Rather, Michaeldreamsaboutbiology.

(3.101a)Mary might wantto comeby today. Regardlessof that,I’m goingto themovies.

(3.101b) Mary might wantto comeby today. Regardless,I’m goingto themovies.

(3.101c) *Regardless,I’m goingto themovies.

Regardingtheprior to constructionshown in Table3.44,thereareothersimilar constructions

taggedasADVP adverbialsin our ADVP corpus,shown alongwith their countsin Table3.47.All

of theseconstructionscontainrelationaladjectives with overt PParguments. [MW] lists all five

constructionsas(complex) prepositionalphrases.Only prior to is containedin our PPadverbial

corpus,however while all fivewerecontainedin ourADVP adverbialcorpus,albeitonly once.

Table3.47:RelationalADJPwith Overt Argument

# Relational ADJP

1 contraryto theseexpectations1 dueto theearthquake in SanFrancisco1 irrespective of theoutcomein centurieselapsedsincesplitting1 shortof fleeingto Warrenton,Virginiaor Rockville,Maryland1 subjectto certainconstitutionalrestraintsin favor of fair trials

As discussedin Section3.2,many adverbsarederived from adjectivesor nouns;someADVP

discourseadverbials,not surprisingly, containadverbsderived from relationaladjectivesor nouns.

Somecorpusexamplesareshown in Table3.48.Thethird columnshowsthe(or oneof thepossible)

relationalnounor adjective from which eachadverbis derived,alongwith a demonstrative instan-

tiation of its argument. Most adverbsderived from relationalnounsor adverbssoundawkward

with thisargumentmadeexplicit, andthusdon’t appearin ourcorpus.Simultaneouslyhowever did

137

Page 155: Thesis about discourse

appearin the corpuswith an overt PPargument,asshown (partially) in the third column. Many

of theserelationalnounsandadjectivesalsohave correspondingPPadverbials(e.g. in addition to

this).

Table3.48:RelationalADVP Adverbialswith HiddenArgument

# Relational ADVP Relational ADJP/NP Derivative

12 accordingly accordingto this3 additionally additionalto this1 alternately analternateto this1 analogously analogousto this8 consequently aconsequenceof this5 conversely aconverseto this1 contrarily contraryto this3 currently currentwith this6 incidentally incidentalto this12 recently recentto this2 previously previousto this2 separately separatefrom this12 similarly similar to this8 simultaneously simultaneouslywith theanodesurfacetemperature....1 subsequently subsequentto this

We saw in theprior sectionthatmakingtheoptionalargumentof a relationalnounexplicit can

move the burden(of classifyingthe adverbial asa discourseor clausaladverbial) to this explicit

argument,becauseit cantake a variety of forms thatmay or may not dependon context for their

interpretation,andmay or may not be an abstractobject. The samevariety canbe displayedby

theexplicit optionalargumentsof theserelationaladjectives. Interestingly, simultaneouslycanfind

both argumentswithin the clauseit modifies,as in (3.102a), wherethe eventsof Mary hearing

thenoiseandMary’s husbandhearingthenoiseareinterpretedassimultaneous. Thecoordinating

conjunction,which alsoassertsthetwo eventssemanticallyandtakesbothargumentsstructurally,

mayplayarolehere;see[CFM�

02] for discussion.Thesameuseof simultaneouslycanalsoresolve

its hiddenargumentto theprior discourse,asin (3.102a), whereFredseeinga bright flashcanbe

interpretedastheevent that is simultaneouswith theevent(s)of Mary andher husbandhearinga

noise.Discussionof clause-internalresolutioncanalsobefoundin [WJSK03] andChapter4.

138

Page 156: Thesis about discourse

(3.102a)Simultaneously, Mary andherhusbandheardanoise.

(3.102a) Fredsaw abright flashof light. Simultaneously, Mary andherhusbandheardanoise.

3.6.2 Context-DependentADVP Adverbials

Approximatecountsin ourcorpusof thetokensandtypesof theADVP adverbialsdiscussedin this

sectionareshown in Table3.49.

Table3.49:ApproximateCountsof TokensandTypesof someADVP Adverbials

# Tokens # Types Category

687 134 PP-RelatedADVP919 26 DeicticADVP

Justaswe saw in our PPcorpus,in our ADVP corpustoo therearea few cases,exemplifiedin

(3.103),wherea subordinatingconjunctioncontainingan internalS argumentandmodifying anS

is misparsedasanADVP adverbial. In many othercasesfoundin ourADVP corpus,however, such

a conjunctionappearssentence-initiallyandthe modifiedS is its internalS argument. The other

conjoinedclauseis to befoundin theprior discourse.Countsof suchcasesareshown in Table3.50.

(3.103)Onceyou getthefeelof it

Table3.50:ADVP AdverbialConjunctions

# ADVP Conjunction # ADVP Conjunction

2 although 15 though212 however 1 unless88 so

Thesecasesaregenerallyviewed as adverbials(and thusmay be correctly tagged,although

PennTreebankdoesnot explicitly indicatetheir treatment),becauseoneargumentis not structural

(with respectto thesentencestructure).DLTAG currentlytreatstheadverbialhowever (andlikely

theothersaswell) asananaphoricdiscourseconnective. Theform of theseadverbialscannotbede-

composedwithout a sincerehistoricalanalysis;they neverthelessclearlyrequireinformationin the

139

Page 157: Thesis about discourse

prior discoursefor their interpretation.Only with however, asshown in (3.104),doesthemeaning

changesperceptiblybetweenthetwo forms. Its meaningasa subordinatingconjunction(3.104a)

conveys a “manner”,while its meaningasanadverbial(3.104b) conveys denialof expectation(c.f.

[Kno96])22. Theremay alsobe a differencein the temporalorderingbetweenthe two; the event

modifiedby adverbialhowever is interpretedasoccurringafter the theeventdescribedin thefirst

sentence,while the event modifiedby conjunctive however is interpretedasoccurringbefore the

eventdescribedin thefirst sentence.

(3.104a)Johncannotseemto learnto tie his shoes,however (much)hetries.

(3.104b) Johncannotseemto learnto tie his shoes.However, hetries.

Also aswe saw in our PPcorpus,many subordinatingconjunctionsfunctionasPPadverbials

whentheir internalargumentis asub-clausalconstituent.TwelvesuchcasesaremisparsedasADVP

adverbialsin ourADVP corpus;in fivecasesthereis aprecedingadverbmodifying thepreposition

head.An exampleis shown in (3.105)

(3.105)ShortlyaftertheValeincident

However, threelexical itemsthat function assubordinatingconjunctionsandprepositionsare

alsofound in our corpusfunctioningasadverbs,in that they appearwithout an internalargument

andin [MW] aretreatedasadverbs(whenmodifying VP or S). For our purposes,theseadverbials

containdefinitesemanticargumentsthatmustbe resolved to informationin theprior discourseor

context. In our corpusthesecasesoccursix times,alwayswith a specifier. Threeexamplesare

shown in (3.106)-(3.108).Moreover, theprepositionafter alsohasansynonymousadverbial form

in which the internalargumentis hidden,shown in (3.109),which occursin our corpusfour times

andwhosehiddenargumentis, like theothers,anaphoricto informationin thediscourseor context.

(3.106)Twentyyearsbefore

(3.107)Shortlyafter

(3.108)Eversince

(3.109)Afterwards

22It would beinterestingto tracethedevelopmentof thesetwo distinctinterpretations.

140

Page 158: Thesis about discourse

Similarly, therearelexical itemsthatdon’t functionassubordinatingconjunctionsbut do func-

tion asprepositions,andthatarefoundin ourcorpusfunctioningasadverbs,in thatthey tooappear

without anovert internalargument.Again, this informationmuststill beresolved in theprior dis-

courseor context. One item, shown in (3.110),appearstwenty-four times. Therearenumerous

otherexamplesthatoccuronceor twice, includingbelow, throughout,outside23, althoughthelastis

ambiguouslya noun24.

(3.110)Besides

Althoughsyntactically, prepositionsthatmodify aVP or Smaybetreatedasadverbswhentheir

syntacticargumentis missing25, at thesemanticlevel thereis no reasonto changetheir representa-

tion; in thelattercasetheir hiddenargumentsimply doesn’t comestructurally.

[Bie01] provides a semanticanalysisof “Xs besidesY” (e.g. other search enginesbesides

BidFind) asan alternative phrasethat is similar to his treatmentof the phrase“other Xs thanY”;

otherandbesidesin theseconstructionsexcludeX (theground) from Y (thefigure) andpresuppose

that both X andY belongto the samesetof alternatives. However, while Biernerconsidersthe

anaphoricform, “other Xs”, whereY is resolved anaphorically, he only considersthe structural

form of besides, not caseswhereY is implicit, becausehe only studiedsub-clausalmodification,

andthis doesnot occurwhenbesidesmodifiessub-clausalconstituents.It doesoccur(twenty-four

timesin ourcorpus),however, whenbesidesmodifiesS.An examplefrom WSJis shown in (3.111).

(3.111)Mr. Langsayshe isn’t scoutingnew acquisitions,at leastfor now. “We would have to

go outsideto banksto get the money andI am not readyto do that,” he said. “Besides, we have

enoughon ourplate.”

Y, theinternalargument,couldhave beenexplicit, e.g.besidesthat. Bierner’s semanticsmayapply

to S-modifyingbesides; in this case,however, the figurewould be an AO from thediscoursethat

is excludedfrom a largerpresupposedsetof AOs,ratherthananconcreteobject. For example,in

(3.111), the useof besidesappearsto invoke a setof reasonswhy Mr. Lang isn’t scoutingnew

23see[Par84]for a discussionof such“context-bound”modifiers.24A few (frequentlymistaggedasADVP) nounscanmodify S,asdiscussedin Section3.2,(e.g. thisyear).25andit is not clearfrom PennTreebankor [MW] thatthis is alwaysthecase

141

Page 159: Thesis about discourse

acquisitions,including Y from the context (having to go outsideto banks),and the modified X

(having enoughon our plate). ExtendingBierner’s semanticsto S-modifyingbesidesrequiresan

annotationstudy, however, suchasthatdescribedin Chapter4.

Frequently, misparsedPPadverbialsfoundin our ADVP corpusarethosein which a particular

prepositionheadis modifiedby a particularadverb26. Table3.51exemplifiessuchconstructions.

Thefirst columnlists thephrase-initialadverb-prepositioncombination(COMB); whenthesecond

prepositionin the combinationvaries,we list it asP. The secondandthird columnsindicatethe

numberof timesthis combinationwastaggedasa PPandanADVP, respectively. Thelastcolumn

providesanexampleof theconstructionasfoundin thecorpus.

Table3.51:Mis-TaggedPPAdverbialConstructions

COMB PP ADVP Example

away from 0 2 away from thegeneralobligationsectoralongwith 9 3 alongwith thenoteapartfrom 3 8 apartfrom racialproblemsasidefrom 10 8 asidefrom thisbackP 4 9 backin theU.S.A.earlyin 7 7 earlyin herlifeelsewherein 1 2 elsewherein theoil sectorfar P 1 6 far from beingminimalist

In [MW], someof theseconstructionsarelistedas(complex) prepositionalphrases.However

mostPP-modifyingadverbsarealsofoundmodifyingadverbsandnounsin ourcorpus,asshown in

(3.112)-(3.114);far is specifiedby theanaphoricso, creatinganoft-usedphrase.

(3.112)Backthen

(3.113)Immediatelythereafter

(3.114)sofar thismonth

More interestingfor our purposes,however, is the fact that many of thesemodifying adverbs

canalsooccuraloneor with only a specifier, in which casethey areinterpretedasspatio-temporal

26alongalsofunctionsasa preposition,sothis couldbea PPadverbialcontainingan internalPPargument;however,it is similar in meaningto theothersincludedin this table.

142

Page 160: Thesis about discourse

anaphorsor deixis,or equivalentlyashaving a hiddenargumentthat is interpreteddespitethefact

thatit is notovert,asexemplifiedin Table3.52.

Onesimilar adverb,ago, can’t modify a preposition,only anNP; moreover, it requiresa spec-

ifier. It appeared132 timesin our corpuswith 71 differentspecifiers.NP modificationandsome

morecommonspecifiersareshown in Table3.53.

Table3.52:Spatio-TemporalADVP Adverbials

# ADVP Adverbial # ADVP Adverbial

1 ascanthalf mile away 3 almostimmediately1 acoupleof yearsback 2 immediately2 elsewhere 2 prettysoon2 shortly 6 thusfar15 soon 32 sofar

Table3.53:AnotherSpatio-TemporalADVP Adverbial

# ADVP Adverbial # ADVP Adverbial

1 ayearagothis fall 12 two yearsago1 seventeenyearsagotoday 8 two weeksago13 ayearago 7 a few yearsago

TheseADVP, alongwith theothertemporalS-modifiersin Table3.52andthosein Table3.54

whicharenotalsousedasmodifiers,supplytemporalinformationto themodifiedclausethey mod-

ify in relationto someothertemporalinformation.

Table3.54:OtherSpatio-TemporalADVP Adverbials

# ADVP Adverbial # ADVP Adverbial

14 already 10 presently24 meanwhile 7 eventually

Thisothertemporalinformationcanberesolvedanaphorically, asin (3.115),whereimmediately

refersto the time after the storefiled for bankruptcy, or resolved deictically, as in (3.116)(from

WSJ),whereimmediatelyrefersto the time right after the time the text wasread.For example,if

143

Page 161: Thesis about discourse

thisdiscoursewerereadin today’s paper, thenimmediatelywould referto just aboutnow.

(3.115)Thestorefiled for bankruptcy. Immediatelycustomersflockedto its closingsales.

(3.116)This is but oneknot in astringof troubles.Lastyearthestorefiled twicefor bankruptcy.

Immediatelyit will besubjectto foreclosure.

Similar resolutionpossibilitiesarefoundfor modifiersin ourcorpus,exemplifiedin Table3.55,

thatareusuallyviewedas“manner”(VP) modifiers,thoughthey toosupplytemporalfeatures.

Table3.55:Spatio-TemporalMannerADVP Adverbials

# ADVP Adverbial # ADVP Adverbial

1 abruptly 1 instantly2 briefly 1 quickly7 gradually 6 slowly1 hastily 23 suddenly

Adopting Ernst’s distinctionbetweensituation (S) modificationandevent (VP) modification

discussedin Section3.4,we find thatthesemodifierscaneitherrelatethetwo temporalboundaries

of theevent,or canrelatetheinitial temporalboundaryof thesituationanda time in thediscourse

or context For example,in (3.117),quickly caneithermeasuresthe time betweenraisingthe eye

dropperandblinking, or the time betweenthestartandfinish of blinking. Briefly, moreover, can

relatethetwo temporalboundariesof aspeechact,asshown in (3.36),repeatedfrom Section3.3.

(3.117)Thedoctorraisedtheeyedropper. Quickly Johnblinked.

(3.36b) (YouaskedmewhatJohnsaid.)Briefly, Johnsaidhewouldstopby.

We have beenaddressingadverbsthathave ananaphoricor deicticquality with respectto the

fact that they mustbe interpretedwith respectto thediscourse27, equivalently viewed asa hidden

semanticargument. True deixis, however, is also frequently found in our corpus,as shown in

Table3.56alongwith their counts28. Therearealsocasesin our corpusin which thesedeicticsare

mistaggedPPmodifiers(e.g.here in Morgenzon...), or arethemselvesmodified(e.g.right now).

27see[Par84]for furtherdiscussion28PennTreebanktagstoday, tomorrow etc.asNPs,although[MW] views themasadverbs.

144

Page 162: Thesis about discourse

Moreover, from many of thesedeictics(e.g. hence, then, still, yet, thus) have beenderived

homonymousrelationaldeixis(discourseadverbials);mostof theoccurrencesin ourcorpusof these

homonyms likely correspondto their discourseadverbial use. The differencebetweentheseuses

is illustratedusing then in (3.118)29. In (3.118b), thenordersthe event it modifiesin a temporal

sequencerelationwith theeventdescribedin (3.118a). In (3.118c), thenmakesdiscoursedeictic

referenceto theeventdescribedin (3.118a) andtheit-cleft assertsthatthetemporalcoordinatesof

thiseventarethesameasthetemporalcoordinatesof theeventdescribedin (3.118c).

(3.118a)JohnandMary haddinner.

(3.118b) Thenheaskedherto marryhim.

(3.118c) It wasthenthatheaskedherto marryhim.

Table3.56:DeicticADVP Adverbials

# ADVP Adverbial # ADVP Adverbial

26 hence 41 still59 here 114 thus189 now 3 today292 then 1 tomorrow39 there 80 yet

Frommany of thesedeicticsstill otherotherdeicticsandrelationaladverbshave beenderived

aswell, asdiscussedin [Kno96]. Thosefoundin ourcorpusareshown in Table3.57.

Table3.57:Deictic-DerivedADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

1 hereby 4 thereafter 1 therefore1 heretofore 1 thereby 1 therein1 nowadays 47 therefore 1 thereupon1 thenceforth

[Kno96, WJSK03]provide detaileddiscussionsof the lexical semanticsof the discoursead-

verbialsin thesetables.That many have retainedtheir deicticquality is exemplifiedin (3.118b),

29This distinctionwasoriginally pointedout by Dr. EllenPrinceto theDLTAG group

145

Page 163: Thesis about discourse

wherethediscourseadverbialcanalsoretrieve therelatedAO from thespatio-temporalcontext. For

example,supposetwo parentsarewatchinga child playing in mudat theplayground.Oneparent

canmake thecommentin (3.119),retrieving this AO from thevisualcontext.

(3.119)“Then/Thereforehe’ll comerunningover hereandputhis handsall over me.”

3.6.3 ComparativeADVP

Approximatecountsin ourcorpusof thetokensandtypesof theADVP adverbialsdiscussedin this

sectionareshown in Table3.58.

Table3.58:ApproximateCountsof TokensandTypesof someADVP Adverbials

# of Tokens # of Types Category

384 110 Atomic Comparatives77 77 Comparative Constructions

NumerousADVP adverbialsin ourcorpusarecomparativeor superlative,andthusby definition

dependentin anidiosyncraticway on referenceto at leastoneotherobjectin theprior discourseor

spatio-temporalcontext (c.f. [Bie01, WJSK03]). In somecases,a comparisonis madebetween

abstractobjects.Most of thecomparativesadverbsin our corpusaretemporalcomparatives. Fre-

quently, they appearmodifying anS-modifier;themodifiedheaddetermineswhatpropertiesof the

objectsarebeingcompared(e.g. equallyoften), or mayclarify thesourceof thecomparison(e.g.

earlier (thanthat) this year). Someexamplesareshown in Table3.59.

Table3.59:Comparative AdverbModifiers

# ADVP Adverbial # ADVP Adverbial

15 earlierthis year 1 laterthatday1 farthersouth 10 morerecently1 equallyoften 5 mostrecently

Somecomparatives(e.g. farther) don’t appearat all asS-modifiers;othersdo, exemplifiedin

Table3.60. In a few instancestheelementwhosepropertiesarebeingcomparedwith thoseof the

146

Page 164: Thesis about discourse

modifiedS is madeovert with a PP, asshown in the secondcolumn. The corpusexamplewhere

more appearaloneis shown in (3.120);it comparestwo AOs,oneis reifiedfrom theinterpretation

of a clausein theprior context, andappearssynonymouswith moreover, discussedbelow.

(3.120)Within theOrganizationof AmericanStates,theremaybesomecriticismof thisunilat-

eralAmericaninterventionwhichwasnotwithout risk obviously. But therewasnocomplaintfrom

theDominicancrowdswhich lined CiudadTrujillo’ s waterfrontshouting,“V ive Yankees”! More,

theU.S.actionwashailedby aprincipaloppositionleader, Dr. JuanBosch,ashaving saved“many

livesandmany troublesin thenearfuture”. (Brown)

Table3.60:Comparative ADVP Adverbials

# ADVP Adverbial # ADVP Adverbial

2 earlier 1 evenearlierthanthat30 later 0 laterthanthat13 further 0 furtherthanthat1 more 2 morethanever

Morefrequentlyin ourcorpus,comparativeADVP S-modifiersoccurwith avarietyof specifiers

thatclarify theextentof thecomparison.Suchcasesoccur111times;in all but six theadverbhead

is earlier or later. Someexamplesareshown in Table3.61.

Table3.61:SpecifiedComparative ADVP Adverbials

# ADVP Adverbial # ADVP Adverbial

16 ayearearlier 1 muchbetter5 oncemore 7 ayearlater

Therearealsoadverbsin ourcorpusthatcannotmodify othermodifiers,but thatmayhavebeen

derived from comparatives,asshown in Table3.62. Whethertheir internalcomparative accounts

for their behavioral anaphoricityis anopenquestion.It is neverthelessclearthattheseS-modifiers

requireanAO from thecontext for their interpretation,andin somecaseswe canmake their hid-

denargumentovert (e.g. otherwisethan that). The lexical semanticsof someof thesediscourse

adverbialsarediscussedin [WJSK03, WJSK99, Kno96].

147

Page 165: Thesis about discourse

Table3.62:Comparative-DerivedADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

31 furthermore 53 moreover 3 nonetheless4 likewise 36 nevertheless 19 otherwise

Also in are corpusare various instantiationsof a comparative constructionof the form “asÛ adverbt as”, wherethe inner as PPtakesan S, NP, or othersub-clausalelementasargument.

50 occurrencesof this constructionarefound in our ADVP corpus;2 arefound in our PPcorpus.

Someinstantiatedforms areshown in Table3.63, alongwith their (ADVP) corpuscountsanda

corpusexample.Therearealsofive relatedconstructionsexemplifiedin thebottomsectionof the

table.“Justas” and“insofar as”alsoappearedin ourPPcorpus.

Table3.63:Comparative Constructions

# Construction Examples

1 ascheerfullyas ascheerfullyaspossible6 asearlyas asearlyas17768 asfar as asfar asi amconcerned10 aslong as aslong asthispoint of view prevails12 assoonas assoonasthetremorpassed1 assurelyas assurelyasaseesaw tilts4 aswell as aswell asin theme9 sofar as sofar asi know3 insofar as insofar assciencegeneratesany fear3 solong as solong asdeathwasnot violent12 justas justasin thecaseof every prodigychild7 nomatterWH no matterhow hot theday

In all cases,thephraseprecedingthe internalas (or WH) phrasefunctionsasanADVP; there

is someambiguity however as to whetherthe ADVP takes an internal argumentor the internal

phrasetakes an ADVP modifier; the category of theseconstructionsvariesin the literature. In

PennTreebank,the “as Û adverbt as” constructionsare taggedas ADJP when they function as

sub-clausalmodifiers,asin (3.121),andasNP whenthey functionasargumentsto transitive verbs,

148

Page 166: Thesis about discourse

asin (3.122).

(3.121)Mary is astall asJohn(is tall).

(3.122)They yieldedasmuchas20%of theexpectedamount.

However, [MW, Lit98, PT] all treataswell asasa coordinatingconjunction,and[Lit98] also

treatsthe no matterWH constructionalongwith someof the “as Û adverbt as” constructionsas

subordinatingconjunctionswhentheinnerphraseis anS.As mentionedin Section3.5,thereis also

somequestionasto whetherall argumentsof theinnerasarenotbestrepresentedassmallclauses.

For example,in thesameway thatwe add“is tall” in (3.121)to interpretthecomparisonof Mary’s

andJohn’s tallness,so we addadditionalmaterialto (3.123)to interpretthe comparisonbetween

Michael’s cheerfulnessandthepossibilityof cheerfulness.

(3.123)As cheerfullyas(it is) possible(to entertheroom),Michaelenteredtheroom.

Aswell appearsoncein ourcorpus,shown in (3.124),indicatingthatits semanticargumentcan

beimplicit. A numberof otherinstantiationsof “as Û adverbt as” can,at leastin spokenlanguage,

leave their argumentimplicit, includingasoften, asrecently. (3.125)alsoseemsfelicitous. More-

over, this authorusesthephraseno matterasanS-modifierwithout theaccompanying WH-phrase

anddependsoncontext for interpretation.

(3.124)FrankGilmartin, a traderwho follows insurancestocksfor Fox-Pitt Kelton, saidhis

strategy wasto sell early. Then, if the stocksfell sharply, he plannedto begin buying themag-

gressively, on thetheorythatthecompaniesthatinsureagainstpropertydamageandaccidentswill

have to raiserateseventuallyto compensatefor theclaimsthey will payto earthquake victims and

victims of last month’s HurricaneHugo. As well, reinsurersand insurancebrokeragecompanies

will have improvedprofits. (WSJ)

(3.125)Janesmiledwidely andburstinto laughter. As cheerfully, Michaelbeganto sing.

3.6.4 Setsand Worlds

Approximatecorpuscountsof ADVP adverbialsdiscussedin thissectionareshown in Table3.64.

149

Page 167: Thesis about discourse

Table3.64:ApproximateCountsof TokensandTypesof someADVP Adverbials

# of Tokens # of Types Category

137 16 Ordinal184 29 Frequency279 19 Epistemic56 31 Domain151 49 Non-Specific263 96 Evaluative/Agent-Oriented

Therearea numberof ordinalADVP in our data. Althoughonly first co-occurswith anovert

PP(first of all occurred6 times)30, thesecondcolumnof Table3.65shows thatotherscanaswell.

Table3.66shows synonymous-ly counterpartsof suchordinals.

Table3.65:OrdinalADVP Adverbials

# ADVP Adverbial Explicit Set

34 first first of all12 second secondof all7 third third of all3 fourth fourthof all1 last lastof all6 next next to that

Table3.66:Ordinal-ly ADVP Adverbials

# ADVP Adverbial # ADVP Adverbial

5 secondly 1 lastly1 thirdly 47 finally

Ordinalsindicatetheorderof themodifiedelementin a largerset.Semanticallythis setcanbe

representedasa hiddenargument,which may or maynot be madeexplicit with a PPandmayor

maynotberesolvedto a setof abstractobjectinterpretationsof VP or S in theprior discourse.For

30next to NPalsooccurredtwice,e.g.“next to theocean”;this interpretationyieldsa spatialordering,however.

150

Page 168: Thesis about discourse

example,in (3.126a), therelevantsetis interpretedasthereasonswhy Mary is a spacecadet,and

in (3.126b), therelevantsetis interpretedastheNP manythings.... Also requiredin thesemantics

of ordinalsis representationof placementin theset.For example,onsecondthoughtrequiresafirst

thought,third of all requirestwo prior elements,andnext requiresat leastoneprior element.

(3.126a) Mary is a spacecadet.First (of all thereasonswhy), shealwaysforgetsto buy milk

for themorning.

(3.126b) I wantto do many thingstoday. First (of all thosethings),I wantto buy milk.

OtherADVP adverbialsin ourcorpusthatinvoke setsare“frequency” ADVP adverbials.Some

examplesoccurringaloneandotherwith adverb modifiersareshown in Table3.67. At times,as

usual, from our PP corpus,are also frequency adverbials. Unlike ordinals,however, frequency

adverbialsinvoke a particulartypeof set:a setof times. But they evoke this settheway quantifiers

suchasfew, all evokesets;althoughthenumberof elementscontainedin thesesetsmaybespecified

(e.g. alwaysversusoccasionally), the setandits elementsmay but neednot resolve in the prior

discourse.Frequency adverbialshave received a variety of semantictreatments;in [Roo95b], for

example,setsof timesareequivalentto events.

Table3.67:Frequency ADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

4 frequently 19 once 14 usually2 always 2 twice 1 almostdaily8 occasionally 8 occasionally 37 again19 often 49 sometimes 2 all toooften

“Epistemic”ADVP adverbialshave alsobeenanalyzedasinvoking sets.Someexamplesfound

in our corpusareshown in Table3.68. In truth, in fact, from our PPcorpus,arealsoepistemic

adverbials.Epistemicadverbialsalsoinvoke aparticulartypeof set:asetof possibleworlds, which

are the foundationof a variety of intensionalsemantics(c.f. [HK98]). Setsof possibleworlds,

and the worlds containedthereinmay but neednot resolve in the prior discourse,andepistemic

adverbialsmayspecifythenumberof worldscontainedin theset(e.g.probablyversuspossibly).

OtherADVP adverbialsin ourcorpusthatcanbeanalyzedasevokingsetsare“domain” ADVP

151

Page 169: Thesis about discourse

Table3.68:EpistemicADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

83 perhaps 2 possibly 3 unquestionably48 maybe 13 probably 5 really31 actually 22 certainly 5 undoubtedly11 surely 44 indeed 1 truly

adverbials.Someexamplesareshown in Table3.69. In theory, in psychology, from ourPPcorpus,

arealsodomainPPadverbials.Thedomainsinvoked by theseadverbialsareagainnot anaphoric,

but they canbe very idiosyncratic;for example,the logical domain(e.g. wherelogic holds),the

psychological domain,etc. Becausedomainadverbsspecifya particulardomain;alternative do-

mains(where,for example,theassertiondoesnothold)maybeevoked,makingthemfeelmorelike

epistemics;we will explorein Chapter5 how suchalternativescanbeimplicated.

Table3.69:DomainADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

1 logically 1 literally 1 mathematically6 microscopically 1 physiologically 1 technically1 theoretically 1 visually 2 statistically1 psychologically 1 publicly 1 functionally

On theotherhand,therearea wide variety of ADVP andPPadverbials,exemplifiedin Table

3.70,which saynext to nothingaboutthepropertiesof thesetthey invoke or its elements.Rather,

they specifythecomparativepositionof theelementthey modify to thatset.Someareanalyzedas

focusparticles, particularlythosein thelastcolumn;wewill discussthemin Chapter5.

As discussedin Section3.3, however, many adverbsarenot so easilyclassifiedinto a single

modificationtype. Rather, they seemto have propertiesof numerousmodificationtypes. For ex-

ample,theADVP in thefirst columnof Table3.71areat oncedomainandfrequency adverbs;in

our terms,thesetsthey describehave both temporalandidiosyncraticfeatures.TheADVP in the

secondcolumnsupplylessspecifictemporalinformation,they feelmorelikeepistemics,asdoesthe

152

Page 170: Thesis about discourse

Table3.70:Non-SpecificSet-Evoking ADVP Adverbials

# ADVP Adverbial # PPAdverbial # ADVP Adverbial

2 basically 2 on thewhole 48 also3 essentially 3 in essence 2 even1 fundamentally 17 in particular 6 just3 specifically 41 at least 6 only3 significantly 1 atworst 11 too1 primarily 3 atbest2 mainly 1 in themain2 mostly 13 in short2 partly 3 in part

PPadverbialcounterpartin general. TheADVP in thethird columnof Table3.71arebothordinals

andtemporal;they describeorderedsetsof timesor events.

Table3.71:Multiply-FeaturedADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

2 historically 2 ordinarily 5 initially1 traditionally 2 typically 3 originally

11 generally 4 ultimately

Many ADVP adverbialsin our corpuswhich have beenclassifiedas“evaluative” alsoconvey

“domain”, and/or“epistemic”,and/or“spatial” properties.Someexamplesareshown in Table3.72.

Again,we canusesetswith multiply featuredidiosyncraticpropertiesto representthis.

Table3.72:More Multiply-FeaturedADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

10 clearly 4 ideally 23 obviously13 naturally 1 reputedly 2 hopefully2 evidently 1 conceivably 1 seemingly19 apparently 5 presumably 4 inevitably

“Evaluative” adverbialsarealsoeasilyconfusedwith “Agent-Oriented”adverbials. Someex-

153

Page 171: Thesis about discourse

amplesareshown in Table3.73alongwith their counts.

Table3.73:Evaluative or Agent-OrientedADVP Adverbials

# ADVP Adverbial # ADVP Adverbial # ADVP Adverbial

2 carefully 2 cautiously 1 convulsively1 deliberately 1 desperately 2 emotionally1 enthusiastically 1 stealthily 1 tardily1 relentlessly 1 gently 1 bluntly

While, asdiscussedin Section3.4,[Ern84] clarifiesthedistinctionbetweeneventandsituation

modification(e.g. Carefully JohnapproachedtheDuchesscarelessly), thereis anotherdistinction

which he overlooksthat canhelp clarify the distinctionbetweenevaluateandagent-orientedad-

verbs.This distinctioncanbeviewedasanadditionalhiddenargument,paraphrasedasto X or, in

X’s opinion. For example,Carefully (in John’s opinion),JohnapproachedtheDuchesscarelessly

describesJohn’s opinionabouthisaction,andthuscarefully is agent-oriented.But Carefully (in the

opinionof all outsideobservers), JohnapproachedtheDuchesscarelesslydescribesanomniscient

opinionaboutJohn’s action(which may or maynot includeJohn’s opinion),andthuscarefully is

evaluative. How the readerresolvesX determinesthe function of theseadverbs. [Swa88]’s view

of not only evaluative andagent-oriented,but alsoepistemicadverbsas Speaker saysADJ(S) is

similar; only thehiddenargumentapproach,however, acknowledgescaseswheretheimplicit X is

resolvedto someoneotherthanthespeaker.

Suchananalysisappliestootherevaluatives,including(not)surprisingly, luckily, (un)fortunately,

miraculously, all of which arefound in our corpus,andsomeof which, alongwith someof those

in Table3.72, are intuitively treatedby [Kno96] asdiscourseconnectives. In otherwords, their

treatmentappearsto bedueto theeffect of this hiddenopinion,namely, thatopinionsareasserted

for a reason. If somethingis assertedto be obvious, or (not) surprising, then thereexists some

reasonthat assertionis made,be it visual or otherwiseapparent,asobviousimplies, or a logical

cause,asinevitably implies. More generally, suchananalysisshouldapplyto all adverbuses.For

example,if somethingis assertedto beprobable, deliberate, obvious, whotheassertionis attributed

to is alsorelevant; it may attributed to a particularsomeone,or to everyone,as is likely the case

154

Page 172: Thesis about discourse

whensomethingis assertedto be first. Clearly, additionalsemantic(andpragmatic)complexity

is requiredto representthis effect, suchasa semanticswhich incorporatesmutualknowledgeand

speaker’s beliefs(c.f. [LA99, HK98, Hir91]).

3.6.5 Summary

In this section,we’ve presenteda wide variety of ADVP from our corpusof ADVP adverbials.

Using the samesemanticmechanismsasin Section3.5, we’ve arguedthat clausalanddiscourse

ADVP adverbialscanbedistinguishedby the interpretationof theadverbandits hiddensemantic

arguments.We’veseenthatalthoughsuchcasesarenotthefocusof theclause-level ADVP literature

discussedin Sections3.3-3.4,many ADVP adverbialsin our corpuscantake an optionalPPthat

instantiatesthishiddenargument.A summaryof someof themechanismsthatcancauseanADVP

adverbialto requireanentity from thediscourseor spatio-temporalcontext for its interpretationare

shown in Table3.74.

Table3.74:ADVP AdverbialSummary

Type ADVP Adverbial

potentialDD here/nowhiddenNP argument unfortunately(for me)hiddenAO argument consequently(on that)hidden � AO � argument first (aboutthat)

3.7 Conclusion

In thischapterwehaveshown thatin many caseswhathavebeencalled“cuephrases”or “discourse

connectives” arenot an accidentalgroupingof ADVP andPPadverbials; rather, their discourse

propertiesarisenaturally from their semantics.We have shown that whetheror not ADVP and

PPadverbialsfound in a corpusfunction asdiscourseconnectivesandareclassifiedasdiscourse

adverbialsdependsontheinterpretationof theirsemanticarguments.Wehaveshown thatdiscourse

adverbialsarevery similar to discoursedeixis, in that both requirefor their interpretationan AO

155

Page 173: Thesis about discourse

madeavailablefrom a VP or S in theprior context. In Chapter4 we will discusswaysin which the

semanticframework outlinedherecanbe formalizedandincorporatedinto theDLTAG model. In

Chapter5 we will discussotherwaysapartfor the interpretationof their semanticargumentsthat

adverbialscontribute to discoursecoherence.

156

Page 174: Thesis about discourse

Chapter 4

Incorporating Adverbial Semanticsinto

DLTAG

4.1 Intr oduction

In Chapter2, we describedsimilaritiesanddifferencesbetweendiscoursetheoriesin termsof nec-

essarymodulesin acompletediscoursemodel.We introducedDLTAG asa theoryof intermediate-

level discoursestructurethat bridgesthe gap betweenclauseand discoursetheoriesby treating

discourse connectivesas predicatesand using the samesyntacticandsemanticmechanismsthat

build theclauseto build discourse.In Chapter3, wedistinguisheddiscourseandclausaladverbials

andshowedhow thepredicateargumentstructureandinterpretationof discourseadverbialscauses

themto functionasdiscourseconnectives.

In this chapter, we investigatehow the predicateargumentstructureandinterpretationof ad-

verbialscanbe incorporatedinto a syntax-semanticinterfacefor theDLTAG model. This chapter

will be exploratoryratherthanconclusive, asa completesyntax-semanticinterfaceincorporating

all aspectsof DLTAG andall discourseconnectives requiresa thesisin its own right. In Section

4.2,wediscusstheroleof thesyntax-semanticinterface,review LTAG, theclause-level modelupon

whichDLTAG is built, andcomparesyntax-semanticinterfacesthathave beenproposedfor LTAG.

In Section4.3,we review DLTAG, discussa syntax-semanticinterfacethathasbeenproposedfor

157

Page 175: Thesis about discourse

a similar tree-baseddiscoursemodel,andexploreDLTAG extensionsof LTAG syntax-semanticin-

terfaces. In Section4.4, we discusshow the DLTAG annotationproject canbe usedto develop

anaphoraresolutionalgorithmsfor theanaphoricargumentsof discourseadverbialsandto train a

statisticalversionof theDLTAG parserto resolve ambiguousstructuralconnections.

4.2 Syntax-SemanticInterfacesat the SentenceLevel

4.2.1 The Roleof the Syntax-SemanticInterface

Naturallanguagecanbedefinedasasetof objects1. At thesentencelevel, for example,theseobjects

correspondto grammaticalandinterpretablesentences;at the discourselevel, they correspondto

grammaticalandinterpretablediscourseunits.

The goal of a languagegrammar(syntax)is to reproducethe structuresof all andonly these

objects. The grammarmust thereforecharacterizethe propertiesof this set (e.g. if it is finite or

infinite), definetheminimal units thatcomposeits members,anddefinerulesfor combiningthese

minimalunitsthatproduceall andonly thesemembers.In thesameway, thegoalof aformalmean-

ing representation(semantics)is to reproducetheinterpretationof all andonly theseobjects.Again

the propertiesof this setmustbe characterized,andthe minimal units that composeits elements

defined,alongwith rulesfor combiningthemthatproduceall andonly theseinterpretations.

In linguistics,theprincipleof compositionalityassertsthatthemeaningof awholeis a function

of the meaningof its parts. At the sentencelevel, for example,meaningfulsentencecomponents

correspondto syntacticsentenceconstituents.Thus,if in a grammarfor Englishwe have the rule

thata sentenceis minimally composedof a nounphrasefollowed by a verbphrase(e.g. S K NP

VP), then we might supposea one-to-onecorrespondencebetweensyntacticrules and semantic

rules,suchthat the interpretationof a sentenceis minimally composedof the interpretationof a

nounphrasecombinedwith theinterpretationof a verbphrase(e.g.S’ = f (NP’,VP’)). Thegoalof

thesyntax-semanticinterfaceis thento define,with respectto thesyntacticstructure,theextraction

from andassemblyof thecomponentsinvolvedin its interpretation.

1This discussionis derivedfrom [Gaz99].

158

Page 176: Thesis about discourse

4.2.2 LTAG: LexicalizedTreeAdjoining Grammar

An LTAG(see[JVS99])is alexicalizedtree-adjoininggrammarwhichitself is anextensionof atree

adjoininggrammar(TAG) (see[Jos87]). Theobjectlanguageof anLTAG (or TAG) is asetof trees,

ratherthanstrings. Treesallow the underlyingstructureof a surfacestring to be represented,as

well asthestringitself. Thelanguagesgeneratedby (L)TAGsarewell known to be“mildly-context

sensitive”, properlycontainingthe context-free languagesandproperlycontainedby the indexed

languages.

An LTAG consistsof a finite setof elementarytreesandoperationsfor combiningthem. El-

ementarytreesareassociatedwith at leastonelexical item, calledthe anchor. Elementarytrees

representextendedprojectionsof the anchorandencodethe syntactic/semanticargumentsof the

anchor. An anchormaybeassociatedwith morethanonetree,calledatreefamily, eachtreereflects

thedifferentsyntacticconstructionsin whichthatanchorcanappear. For example,theverbeatmay

beeithertransitive or intransitive; eachof theseformsis givenacorrespondingtree.

Therearetwo typesof elementarytreesin anLTAG: initial trees,whichencodebasicpredicate-

argumentrelations,and auxiliary trees,which encodeoptional modificationand must containa

non-terminalnode(calledthe foot node)whoselabel matchesthe label of the root. Examplesof

elementaryLTAG treesareshown in Figure4.1. Thefinal treein this figureis anauxiliary tree,all

theothersareinitial trees.

SÍ Í ÍÎÎÎNP! Ë VPÏ ÏÐÐ

V

walks

NP� ËNP

N

John

NPÙ ÙÚÚD Ë N

dog

D

the

VPÖ Ö××ADV

often

VPÌFigure4.1: ElementaryLTAG Trees

Therearetwo structure-building operationsin anLTAG for creatingcomplex trees,calledde-

rived trees:substitution(indicatedby Ë ) andadjunction(indicatedby Ì ).Thesubstitutionoperationis restrictedto non-terminalnodesmarkedby Ë on thetreefrontier.

Substitutionconsistsof replacingthis nodewith the treebeingsubstituted.Only initial treesor

159

Page 177: Thesis about discourse

treesderived from initial treescanbe substituted,andthe root nodeof the treebeingsubstituted

mustmatchthe label of the nodebeingreplaced. For example,the treeanchoredby the canbe

substitutedfor thenodelabeledáâË in thetreeanchoredby dog. This treecanthenbesubstituted

for theinternalargument(NP� ) in thetreeanchoredby walks, andthetreeanchoredby Johncanbe

substitutedfor theexternalargument(NP! ) in thetreeanchoredby walks. Theresultof thesetwo

substitionoperationsis thederivedtreein Figure4.2.

SÍ Í ÍÎÎÎNP

N

John

VP� � ����V

walks

NPÙ ÙÚÚD

the

N

dog

Figure4.2: LTAG DerivedTreeafterSubstitions

Theadjunctionoperationis restrictedto non-terminalnodesnotalreadymarkedfor substitution.

Adjunction consistsof building a new treefrom an auxiliary tree�

andany other tree � (initial,

auxiliary, or derived). In orderto combine�

and � by adjunction,the root nodeof�

mustmatch

the labelof thenode W in � to which it is to beadjoined. If this is thecase,the root nodeof�

is

identifiedwith W ; the subtreethat wasdominatedby W is attachedto the foot nodeof�

, andthe

restof thetreethatdominatedW now dominatestherootnodeof�

. For example,thetreeanchored

by oftencanadjoin to the VP nodeof the derived tree in Figure4.2, producingthe derived tree

correspondingto thesentenceJohnoftenwalksthedog, asshown in Figure4.3.

The treesin Figures4.2 and4.3 do not recordthe informationaboutwhich elementarytrees

werecombinedandwhich operationswereusedto combinethem.Thus,in additionto thederived

treethatrepresentstheresultsof combiningtreesto form complex trees,a -q.�o7+4ãqj`b�+'T�W treein LTAG

specifiesuniquelyhow a derivedtreewasconstructed.Nodesarelabeledaccordingto theelemen-

tary treeinvolved at that point in the derivation; the root of thederivation treecorrespondsto the

labelof theinitial treewhoseroot is S. In general,�B� labelsinitial treeswhoseanchoris ) (by defi-

160

Page 178: Thesis about discourse

nition aninitial treeis substituted),and� � labelsauxiliary treeswhoseanchoris ) (by definitionan

auxiliary treeis adjoined). A treeaddressis alsoassociatedwith eachnodein thederivation tree

excepttheroot. This addressis theaddressof thenodein the PAj/op.�Wcb treeat which thesubstitution

or adjunctionoperationhasoccurred.Theaddressof therootnodein theparenttreeis 0, theaddress

of the ä "� child of thatroot is (k), and(p.q) is theaddressof the Q "� child of anodeP .

SÍ Í Í ÍÎÎÎÎNP

N

John

VPÍ Í ÍÎÎÎADV

often

VP� � ����V

walks

NPÙ ÙÚÚD

the

N

dog

Figure4.3: LTAG DerivedTreeAfter Adjunction

Thederivation treecorrespondingto thederived treein Figure4.3 is shown in Figure4.4. As

shown, the initial tree rooting the derivation is ��������� . ��������� substitutesinto ���$�%�&�� at address

(1) (NP! ), while� ���� "!#� adjoinsinto ��������� at address(2) (VP), and �B����F substitutesinto ���$�%�&�� at

address(2.2) (NP� ). �� "��! substitutesinto ������F at address(1) (D).���������� � � � � ���������������� (1)� ���� "!'� (2) ������F (2.2)�� "��! (1)

Figure4.4: LTAG DerivationTree

4.2.3 A Syntax-SemanticInterface for LTAG Derivation Trees

[JVS99] arguethatif theLTAG operationsof substitutionandadjunctionareviewedasattachments

of one tree to anothertree, then a syntacticsentencederivation consistsof an unorderedset of

attachments,andthecorrespondingsemanticscanbuilt monotonicallyasasemanticsof attachments

161

Page 179: Thesis about discourse

anddescribedby a flat semanticrepresentation.For example,the semanticrepresentationof the

sentencecorrespondingto the structurein Figure 4.3 might be as in (4.1), where i1, i2 denote

individualsande1denotesanevent.

(4.1) john(i1) J walks(e1,i1, i2) J dog(i1) J often(e1)

[JVS99] notehowever, that the adjunctionoperationdoesnot preserve monotonicitywith re-

spectto theimmediatedominationrelation;theVP thatis immediatelydominatedby Simmediately

dominatesV in ��������� , but no longerimmediatelydominatesV afteradjunctionof �B���� "!'� .

Although under-specificationof the immediatedominancerelation would remove this non-

monotonicity, the tacit assumptionin many LTAG formalismsis ratherthat thenodeson thetrunk

of a tree(thepathfrom theroot to theanchor)arenot distinguishedsemantically; ([KJ99, Kal02,

JVS99]andreferencestherein)argue that usinga semanticsdefinedin termsof syntacticattach-

mentsremovestheneedto make useof syntacticunder-specificationif compositionalsemanticsis

definedwith respectto the derivation tree(Figure4.4), ratherthan the derived tree(Figure4.3).

Theseauthorsarguethat this is thenaturaltreeuponwhich compositionalsemanticsof sentences

shouldbebuilt, becausethepredicateargumentstructureof a lexical itemis representedonly in the

derivationtree(not in thederivedtree),andonly thederivationtreerecordsthedifferentelementary

treesinvolvedin thederivationanddistinguishesthesubstitutionof argumentsinto elementarytrees

from theoptionalmodificationof (adjunctionto) lexical itemsby otherlexical items2.

To build the compositionalsemanticsfrom the derivation tree, [JVS99] associatea tripartite

semanticrepresentationwith eachelementarytree.Thefirst partof therepresentationspecifiesthe

mainvariableof thepredication.Thesecondpart(innerbox) statesthepredication.Thethird part

(lowerboxes)associatesvariableswith argumentnodesin theelementarytrees.As anexample,the

semanticrepresentationsof (transitive) ���$�%�&�� and � ������� areshown in Figure4.5.

Thecompositionof theserepresentationsaftersubstitutionof � ������� in thesubjectpositionof���$�%�&�� is obtainedby unifying thevariable(x M ) correspondingto thesubjectnode3 in ��������� with

thevariable(x � ) which the representationcorrespondingto � ������� is j � T�k;b . After unification,the

2See[JVS99]for detailsconcerningdifferencesbetweenthedominationrelationsin derivationanddependency trees,andconstraintsonhow thederivationtreemustbetraversedwhenbuilding thecompositionalsemantics.

3This will beformalizedbelow by [Kal02] by statingpairsof variablesandcorrespondingnodeaddresses.

162

Page 180: Thesis about discourse

secondparts(innerboxes)aremerged.Theresultingcompositionis still j � T�kAb theeventof walking

representedas . M . The semanticrepresentationafter composing���$�%�&�� and ��������� is shown in

Figure4.6.���$�%�&�� � �������about:eMwalk(eM , x M , x g ) named:x �

about:x �John(x� )

x1 x2

Figure4.5: SemanticRepresentationsof �������&�� and ���������Johnwalks

about:eMnamed:x Mwalk(eM , x M , x g )John(xM )

x2

Figure4.6: SemanticRepresentationsof Johnwalks

[JKR03, Kal02, KJ99] further formalize a syntax-semanticinterfacein which compositional

semanticsdependson theLTAG derivation tree. In orderto representscopeambiguitiesin quan-

tifier adjuncts,they employ a restricteduseof multi-componentTAGs,alongwith morecomplex

flat semanticrepresentationusing ideasfrom Minimal RecursionSemantics([CFS97])andHole

Semantics([CFS97,Bos95]), which consistsof threeparts: typedlambdaexpressions,scopecon-

straints,andargumentvariables.[Kal02] furtherarguesthatthederivationtreecanbeenrichedwith

additionallinks to supporta compositionalsemanticsthatcanrepresentall thedifferentiatedscope

orderingsof quantifieradjuncts,suchasthoseproducedby quantifierandPPadjunctsof NPs4.

We will discussdetailsof this approachbelow in referenceto DLTAG; in theremainderof this

sectionwe illustratehow the compositionalsemanticsof a simpleexamplecanbe built from the

LTAG derivation tree. [JKR03, Kal02, KJ99]’s semanticrepresentationsfor ���$�%�&�� , ��������� , � ���� "!'�4[FvG01] proposean alternative solutionto someof theseproblemsthatusesinformationfrom both the derivation

andthederivedLTAG tree.

163

Page 181: Thesis about discourse

and ��������� areshown in Figure4.75. As shown, lambdaexpressions(formulas)maycontainpropo-

sitionallabels) � , holes V � (meta-variablesfor propositionlabels),aswell aspropositionalargument

variables( � (whosevaluesarepropositionallabels),andholevariables � (whosevaluesareholes).

Holesandlabelsareusedto generateunder-specifiedsemanticrepresentationsandallow for scope

ambiguities,whoseorderingis constrainedby thescopeconstraints.In �������&� , for example,thereis

a holeabove thepropositionlabel ) M which mayproducescopeambiguityconstrainedash MÆå l M .Quantifiersandadverbscanintroduceadditionalholesandlabels,asshown; scopedisambiguation

occurswhenthesesemanticrepresentationsarecombined,asdiscussedbelow. Argumentvariables

maybe of any variabletype,andmaybe linked to addressesin thesyntactictree. This linking is

explicitly representedasapair (e.g. Û x M ,(1)t , which indicatesthatx M is linkedto address(1)).���$�%�&�� � ������� �B������� ������ "!#�l M : walk(xM , x g )h Mæå l M———————————arg: Û x M ,(1)t , Û x g ,(2.2)t John(x� )

———–arg: –

Fido(xç )———–arg: –

l g : often(hg )g Mæå l g , hg å sM——————arg: g M , sM

Figure4.7: SemanticRepresentationsof ��������� , ��������� , ��������� , and� ���� "!#�

Whenthe derivation tree for the sentenceJohn oftenwalks Fido is built from its constituent

trees,thesemanticrepresentationof this sentencecanalsobebuilt. This representationis shown in

Figure4.8.

JohnoftenwalksFido

l M : walk(x� , x ç ), John(x� ), Fido(xç ), l g : often(hg )h Mæå l M , h Mæå l g , hg å l M——————————————————————arg: –

Figure4.8: SemanticRepresentationsof JohnoftenwalksFido

Combiningsemanticrepresentationsconsistsof building theunionof thesemanticrepresenta-

tionsof theelementarytreesinvolved in thederivationandassigningvaluesto argumentvariables.

Thederivationtreeindicateshow thesemanticrepresentationsareto becombinedsuchthattheirar-

5As [JKR03, Kal02, KJ99] do not discusstheir semanticrepresentationof definitedeterminers,we illustrate theirbasicapproachby replacingthedog from ourearlierexamplewith è�é"ê�ë .

164

Page 182: Thesis about discourse

gumentvariablesgetvalues.Whenatreeis substituted,its valueis appliedto theargumentvariable

pairedwith thepositionatwhich it attaches.For example,thederivationtreein Figure4.3indicates

that thevalueof x M comesfrom the semanticrepresentationof � ������� , andthe valueof x g comes

from thesemanticrepresentationof �B���6��� .Whenatreeis adjoined,however, its argumentvariablesareaddedto therepresentationandare

assignedthevaluesat thepositionat which it attaches.For example,thederivation treein Figure

4.3 indicatesthatthevaluesof sM andg� comefrom thesemanticrepresentationof ���$�%�&� . Because( M is a propositionalargumentvariablewhosevaluesarepropositionallabels,it is assignedto ) M ,andbecause M is aholevariablewhosevaluesareholes,it is assignedto V M . Theseassignmentsare

reflectedin thescopeconstraintsin thecombinedsemanticrepresentation.

Scopedisambiguationconsistsof finding bijectionsfrom holesto labelsthat obey the scope

constraints.Accordingto thescopeconstraintsin Figure4.8,h M å l g andhg å l M . We alsoknow

thatl gxt hg , becausehg appearsinsidetheformulalabeledl g . We thereforeknow thatl gÆt l M , and

we know thath Mìuv l M , becauseif h M = l M thenl M å l g andwe alreadyknow thatl g t l M .Therefore,theonly possibledisambiguationof our holesis: h M =l g andhg = l M . This yieldsthe

embeddedsemanticrepresentationin (4.2).

(4.2)John(x3) J Fido(xç ) J walk(x� , xç ) J often(walk(x� , xç ))Notethatwe have presentedin this examplethemorecomplex semanticsinvolved in thetreat-

mentof scopeambiguitiesfor thepurposeof showing how scopeambiguitiesarehandledseman-

tically in this approach;below whenwe considera DLTAG versionof this approachwe will not

considerscopeambiguities,andwe will thusfollow [KJ99] in usingsimplifiedsemanticrepresen-

tationssuchasshown in Figure4.9.���$�%�&�� �B���� "!'�l M : walk(xM , x g )———————————arg: Û x M ,(1)t , Û x g ,(2.2)t l g : often(sM )

—————arg: sM

Figure4.9: SimplifiedSemanticRepresentationof ���$�%�&�� and� ���� "!'�

165

Page 183: Thesis about discourse

4.2.4 A Syntax-SemanticInterface for LTAG Elementary Trees

[SBDP00]presentanextensionof theLTAG grammarin which thelexical entriesof motionverbs

arecraftedto make it aseasyaspossiblefor anaturallanguagegenerationsystem(SPUD,[SD97])

to selectthe verb that besthelpsits communicative goalsbe achieved. Briefly, SPUDgenerates

instructionsfor action in a concretedomain. SPUD’s desiredoutput is to mirror the naturally-

occurringaction instructionsin a selectedcorpus. SPUD’s input consistsof (1) a representation

of the context in which instructionis to be issued,(2) a set of communicative goalsdescribing

contentthattheinstructionshouldcontain,(3) adatabaseof factsdescribinggeneralizedindividuals

involved in the action (e.g. paths,placesand eventualities). When planninga sentence,SPUD

searchesthe derivationsof a true sentencethat areadmittedby the grammarfor onewhich best

achievesits communicative goalsin thecurrentcontext.

[SBDP00]constructa lexical entry, consistingof an elementarytree,a syntax-semanticinter-

face,anda semantics,for five motion verbs,basedon an analysisof their usein an instructional

corpus.The treesassociateeachanchor(verb)with its observed rangeandorderof complements

andmodifiersaccordingto their modificationtype6, by usingexhaustive nodehierarchiesin theel-

ementarytree,therebyallowing SPUDto generatetheseelementsin any orderwhile still producing

thecorrectsurfaceorder. As anexample,theelementarytreefor slide is shown in Figure4.10.

S� � ����NPÌ VP-PRPÌ

VP-DURÌVP-PTHÌÖ Ö××V

slide

NPËFigure4.10:TheElementaryTreefor slide

This treestructurerepresentstheobservedtype,order, andoptionalityin theinvestigatedcorpus

6see[KP02]) andChapter3 for discussionof thesemodificationtypes.

166

Page 184: Thesis about discourse

of theargumentsandmodifiersof slide. As shown, all optionalelements,whetherthey aredeter-

minedto beoptionalargumentsor adjuncts,arerepresentedusingadjunctionso thatSPUDis not

forcedto anticipatehow thesentenceis to be completedbeforeselectingamongalternative trees.

An exampledemonstratingtheuseof slideandtheseelementsis shown in (4.3).

(4.3a) Slide thecover acrossthepole quickly to achieve a tight seal.T � {`.Ci8b PAj/b�V -pkAo7j/b�+'T�W P;k;o8PAT>(�.(PTH) (DUR) (PRP)

SPUDusesanontologicallypromiscuoussemantics[Hob85] suchthateachlexical entryused

in thederivationof anutterancecontributesaconstraintto its overallsemantics;thesyntax-semantic

interfacedetermineswhichof theconstraintscontributedby anentrydescribethesamegeneralized

individuals. For example,given the phraseslide the sleeve quickly, the syntax-semanticinterface

guaranteesthattheevent . describedby slide is identifiedwith anevent .*í thatis quick. To express

thesemanticrelationshipsbetweenmultiple entriesin a derivation, [SBDP00] associateeachnode

in theelementarytreewith the individualsthat thenodedescribes.Whenonetreecombineswith

anotherby substitutionor adjunction,anodein onetreeis identifiedwith anodein anothertreeand

thecorrespondingentitiesareunified. Theindividualsassociatedwith eachnodein theelementary

treein Figure4.10areshown in Figure4.11.

The collectionof individuals associatedwith the nodesof a verb treearecalled its semantic

arguments. [SBDP00]’s notion of a semanticargumentis clearly distinguishedfrom the LTAG

notion of an syntacticargument. Eachsyntacticargumentcorrespondsto onesemanticargument

(or more),sincethe syntacticargumentpositionis a nodein the treeandis associatedwith some

semanticargument(s).However, semanticargumentsneednot be associatedwith syntacticargu-

mentpositions.For example,thereis no nodein thetreeinto which theeventuality . describedby

slidesubstitutes,but eventualitiesareneverthelesstreatedasa semanticargumentof boththeverb

andof eventmodifierssuchasquickly, asin [Dav67]. Moreover, optionalconstituentsspecifying

paths,durations,or purposesareusuallytreatedsyntacticallyasmodifiers,usingadjunction.Here,

althoughall optionality is treatedvia syntacticadjunction,theseadjunctionsitesmayeitherbeas-

sociatedwith referencesto only theoveralleventualityargumentof theverb,makingtheseadjuncts

167

Page 185: Thesis about discourse

semanticadjunctsaswell, or with referencesto additionalsemanticarguments(e.g.paths), making

themsemanticarguments.[SBDP00,4] notehowever that it is a substantive questionfor grammar

designwhich optionalconstituentsshouldbe treatedasspecifyingadditionalsemanticarguments

for a given verb entry; they make useof the testsdescribedin Chapter3 to distinguishsemantic

argumentsfrom semanticadjuncts(e.g.thedo so test,theextraction test,thepresuppositiontest).

S K î event: eïð ð ð ð ð ð ð ð ð ðññññññññññNPÌÈK î ind: agentï VP-PRPÌòK î event: eï

VP-DURÌóK î event: eïVP-PTHÌ�K }~~� event: e

path:p

changed:obj

�&���� � � � � � ��������V K }~~� event: e

path:p

changed:obj

� ���slide

NPËôK î ind: objïFigure4.11:TheSyntax-SemanticInterfacefor (*),+'-/.

[SBDP00]furtherspecifythesemanticrepresentationassociatedwith eachverbentry, in terms

of an assertionand a presuppositionaboutthe individuals referencedin the tree. The semantic

representationassociatedwith theverbentryfor slide is shown in (4.4).

(4.4a)Presupposition:located-at-start(obj, P ), along-surface(P )

(4.4b) Assertion:caused-motion( . , agent, obj, P )

In SPUD,theassertioncontributesnew relationshipsamonggeneralizedindividuals.For exam-

ple,slideassertsthatanagentcausesanobjectto move alonga path.Thepresuppositionindicates

backgroundknowledgeaboutadditionalrelationshipsbetweengeneralizedindividualsinvolved in

theassertion7. For example,aslidingeventpresupposesthatthepathalongwhichtheobjecttravels

7We discusspresuppositionin detail in Chapter5.

168

Page 186: Thesis about discourse

hasan origin, andpresupposesthat this path involves a surfacethat remainsin contactwith the

objectduringthesliding.

4.2.5 Comparison of Approaches

Althoughboth[JKR03, Kal02, KJ99, JVS99]and[SBDP00]proposeasemanticsandandasyntax-

semanticinterfacelinking it to the LTAG grammar, the two approachesdiffer in a numberof su-

perficial ways. First, [SBDP00]postulatesa differentsetof elementarytrees. Eachof their trees

fully specifies,ashierchicalnodes,all observedsyntacticargumentsandmodifiersthatappearwith

thecorrespondingverbin a corpus,regardlessof whetherthey aredeterminedto functionsemantic

argumentsor adjuncts.In [JKR03,Kal02, KJ99, JVS99]’s approach,only syntacticargumentsare

representedin their elementarytrees.

Second,[SBDP00] specifiesin theelementarytreethesemanticargumentscontributedby both

argumentandmodifiernodesto thesemanticrepresentation,while [JKR03, Kal02, KJ99, JVS99]

specifyonly in thesemanticrepresentationassociatedwith theentireelementarytreethesemantic

contribution of argumentnodes.

Third, [JKR03, Kal02, KJ99, JVS99] argue that thereis no semanticdistinctionbetweenthe

nodeson thetrunkof thesyntactictree(e.g.S-VP-VP-V),while [SBDP00] explicitly distinguishes

thesemanticcontribution of eachnode.

Fourth,[SBDP00]don’t specifywhether, duringsubstitutionor adjunction,theidentificationof

theentitiessuppliedby acomplementor modifierwith theentitiesassociatedwith thecorresponding

nodein theverbelementarytreeis madebasedon thederived or derivation tree. [JKR03, Kal02,

KJ99, JVS99]arguethat thederivation treeis theappropriateplaceto specifythesyntax-semantic

interfacebecausethederivedtreeis syntacticallynon-monotonicandonly thederivationtreerecords

thepredicateargumentstructureandoperationinvolvedin constructingthesyntacticrepresentation.

Although [SBDP00] do not provide detailson how the interpretationof a sentenceis con-

structed,[SBDP00]and[JKR03, Kal02, KJ99, JVS99]’sapproachesarepotentiallycomplementary.

In particular, in [SBDP00]’sapproach,eachnodehasauniqueaddress.Thereforetheidentification

of thespecifiedsemanticargumentsat eachnodewith thesemanticargumentsof treesadjoinedor

169

Page 187: Thesis about discourse

substitutedat thosenodescanbebasedon thederivationtree.

Themostsignificantdifferencebetweenthetwo approachesis [SBDP00]’s pre-specificationof

all possibleargumentsandmodifiersin theelementarytree.Thepurposeof this is to provideSPUD

with all theinformationit mayneedto exploit in a computationallyefficient way, whengenerating

anutteranceaccordingto its conversationalgoals.GiventhatSPUDappliesin alimited domainand

its outputis meantto mirror afinite corpus,thisapproachmaybemostcomputationallyefficient in

thisdomain.It is notclearhowever, thatthisapproachis bestwhenconsideringgenerationor inter-

pretationin unlimiteddomains.For example,in Chapter3, a wide rangeof modifiersdrawn from

theWSJandBrown werediscussed,someof whichcouldbecategorizedinto multiplemodification

types. It remainsanopenquestionof whetherit is possibleto createa modificationclassification

which is at oncecompactenoughto be incorporatedinto theelementarytreesof all verbs,andat

thesametime accuratelyandcomprehensively coverstheinterpretationsof (andany idiosyncratic

semanticargumentsspecifiedby) all modifiers.

We endthis sectionwith thecommentthatLTAG is not theonly syntacticgrammarfor natural

language;CombinatoryCategorial Grammar(CCG)[Ste96] is onealternative which encodesbuth

thesyntacticandsemanticpropertiesof wordsin thelexicon. For example,thetransitive verb like

might have thelexical entryin (4.5).

(4.5) lik e=syn: Sõ NP/NP

sem: ß x ß y.like(y,x)

Briefly, in CCG,/ refersto a rightward-lookingcategory and õ to a leftward-lookingcategory.

Categoriescombineusingrulesof functionapplication. (4.5) statesthat the syntacticcategory of

lik e is afunctionthatrequiresits syntacticargument,anNP, onits right. A functionapplicationrule

is applied,suchthat the NP on the right is identifiedandthe correspondingsemanticargumentis

simultaneouslyboundto theoutervariable� . A new functionis producedthatrequiresits syntactic

argument,againan NP, on its left. Anotherfunctionapplicationrule is applied,suchthat the NP

on the left is identifiedandthe correspondingsemanticargumentis simultaneouslyboundto the

variable� . Theresultis anSwith semanticslike( � , � ), where� and � havebeenboundasdescribed.

Onereasonfor usingCCG is that it removesthe needfor a syntax-semanticinterfacespecifying

170

Page 188: Thesis about discourse

how a semanticsanda syntaxarerelated.See[Ste96, JVS99]for a discussionof similaritiesand

differencesbetweenthetwo grammars.

4.2.6 Summary

In this sectionwe have overviewed the LTAG grammar, andhave presentedtwo syntax-semantic

interfacesthathavebeenproposedfor LTAG. In thenext sectionwewill focusonthedetailsof DL-

TAG, whichbuildsdiscoursegrammardirectlyon topof theLTAG clausegrammar, andinvestigate

how syntax-semanticinterfacesfor suchLTAG-baseddiscoursegrammerscanbedefined.

4.3 Syntax-SemanticInterfacesat the DiscourseLevel

4.3.1 DLTAG: LexicalizedTreeAdjoining Grammar for Discourse

Like many of the other discoursemodelsdiscussedin Chapter2, DLTAG [FMP�

01, CFM�

02,

WJSK03,WKJ99, WJSK99, WJ98] arguesthat discoursecanbe modeledin termsof syntaxand

semantics,but unlikeothermodels,DLTAG definesdiscoursestructureandmeaningin termsof the

samemechanismsthatarealreadyusedat thesentencelevel, andbuilds them,furthermore,directly

ontopof theclause.As [WJSK03, 23]note,thenotionthatdiscourserelatesto syntaxandsemantics

in a completelydifferentway thanthesentenceseemsstrangewhenwe considerexamplessuchas

(4.6),whereanentirediscourseis containedwithin a relative clause,exhibiting thesamecohesive

andargumentative connectionswhicharecharacteristicof otherdiscourses.

(4.6) Any farmerwho hasbeatena donkey andgonehomeregrettingit andhasthenreturned

andapologisedto thebeast,deservesforgiveness.

In DLTAG, discourseconnectivesaretreatedakin to verbsat theclauselevel. As discussedin

Section4.2,verbsat theclauselevel aregenerallyviewedaspredicatesthat take entity interpreta-

tions andsupplyrelationsbetweenthemto form a sentenceinterpretation.In DLTAG, discourse

connectivesarealsopredicatesthat take entity interpretationsandsupplya relationbetweenthem

to form interpretationsof larger discourseunits. Thesetof discourseconnectives in DLTAG cur-

rently includesthesubordinatingandcoordinatingconjunctions,punctuationat clauseboundaries,

171

Page 189: Thesis about discourse

anddiscourseadverbials,As we saw in Chapter3, theentity interpretationsthey relateareabstract

objects(AOs),which may comefrom non-NPconstituentswithin clauses,clausesthemselves,or

discourseunits that arecomposedfrom both clausesanddiscourseconnectives, within or across

sentenceboundaries.Wewill thususeD 5 to representthesyntacticargumentsof all discoursecon-

nectivesandtheunit thatresults(ratherthanS),bothto indicatethattheseunitsarebeinganalysed

from thepointof view of discourse,andbecausetheclause-level syntacticconsituentcorresponding

to theseunitscanvary.

DLTAG currently builds discourseusing the structuresand structure-building operationsof

(LTAG)[JVS99], which itself is widely usedto model the syntaxof sentences.As in LTAG, in

DLTAG eachelementarytree is anchoredby at leastone lexical item or correspondingfeature

structure8. Also as in LTAG, thereare two kinds of elementarytrees: initial treesthat represent

atomicunits andlocalizepredicateargumentdependencies,andauxiliary treesthat representop-

tional modification.

In DLTAG, two initial treesareproposedin the treefamily that representssubordinatingcon-

junctions,exemplified in Figure 4.12. Two treesare proposedbecauseof the syntacticalterna-

tion that subordinatingconjunctionsallow with repectto their positionrelative to their arguments

[QGLS85],asshown in (4.7).

(4.7a)Johnis hardto find, althoughheis generous.

(4.7b) Althoughheis generous,Johnis hardto find.

D 5� � � � ������D 5 Ë although D 5 Ë D 5� � � �ö ö����

Although D 5 Ë D 5 ËFigure4.12:DLTAG Initial Treesfor SubordinatingConjunctions

In LTAG, subordinatedclausesaretreatedasadjunctsbecausethey arenotpartof theextended

projections(e.g. argumentstructure)of theverbof themainclause.In DLTAG, however, it is the

extendedprojectionsof thediscourseconnective, not theverb,thatarebeingmodeled;subordinat-

8See[WJ98] for a discussionof reasonsfor treatinglexical anchorsas featurestructuresthat may or may not belexicalized.

172

Page 190: Thesis about discourse

ing andcoordinatingconjunctionsrelatetwo clausalinterpretationto form a larger discourseunit

andthusarerepresentedastaking two substitutedarguments.Subordinatingconjuctionsarethus

modeledwith initial trees;asshown in (4.8), the local dependenciesbetweentheir arguments(4.8

a) canbe stretchedlong-distancevia adjunctionof an additionalclause(4.8 b), asis alsotrue of

localdependenciesat theclauselevel (e.g.Apples,Bill saysJohnmaylike.).

(4.8a)AlthoughJohnis generous,he’s hardto find.

(4.8 b) Although Johnis generous–forexample,he givesmoney to anyonewho askshim for

it-he’s hardto find.

In DLTAG,two differenttypesof auxiliarytreesareproposed.Thefirst typeis usedto represent

simplecoordinationandanemptyconnective 0 , asexemplifiedin Figure4.13.

D 5Í Í ÍÎÎÎD 5 Ì and D 5 Ë D 5� � ����

D 5 Ì 0 D 5 ËFigure4.13:DLTAG Auxiliary Treefor andand 0

While both argumentsin thesetreescomestructurally, adjunction to a discourseunit in the

prior discourse(ratherthansubstitution)representsthe fact that and, or, 0 convey a continuation

(or optionalmodification)of somethingin theprior discourse.In otherwords,simplecoordination

provides further descriptionof a situationor of one or more entities(objects,events,situations,

states,etc.)within thesituation[WJSK03]. This is akin to thenotionof elaboration, asexemplified

in (4.9a); it alsoaccountsfor caseswhereacoordinatingconjunctionis usedto connecttwo clauses

thatsupplythesamerelationvia astructuralconnective to theprior discourse,asexemplifiedin (4.9

b), wherebothdisjunctsconvey analternative pointat whichJohnwill quit his job.

(4.9a)Johnwentto thezoo[and/.] H/hetookhis cell phonewith him.

(4.9b) Johnwill quit his job whenhewins thelottery [and/or] hemarriesa rich woman.

Thecoordinatingconjunctionsso,but convey morethansimplecontinuation,however; so, for

example,conveys a result relation in (4.10). Theseconnectives are thus representedwith initial

173

Page 191: Thesis about discourse

trees9.

(4.10)You didn’t eatyourspinachsoyouwon’t getdessert.

Thesecondtypeof auxiliary treeproposedin DLTAG representsdiscourseadverbials;asshown

in Figure4.14,therearetwo treesin thetreefamily representingdiscourseadverbialsdueto their

ability to appearS-initially andS-finally. Notehowever that,asdiscussedin Chapter3, discourse

adverbials, like all adverbials,can appearin a variety of other positions. Currently in DLTAG,

discourseadverbialsareextractedfrom S-internalpositionsandthemodeledusingtheS-initial tree

with a traceleft in their originalposition(see[FMP�

01] for details).

D 5÷ ÷øøD 5 Ì then

D 5÷ ÷øøthen D 5 Ì

Figure4.14:DLTAG Auxiliary Treesfor DiscourseAdverbials

TheseDLTAG treesarestructurallyidenticalto LTAG treesfor S-modifyingadverbials(except

for thelabelof their root node);thedifferencelies in thefact that in theDLTAG (discourse)gram-

marall discourseunitsarestructurallyrelatedto theprecedingdiscourse.A discourseadverbial is

viewedasanoptionalmodificationof theincomingunit, which suppliesanadditionalsemanticre-

lationoverandabove thesemanticrelationsuppliedby thestructuralconnection;thustheargument

a discourseadverbial modifiesis representedasadjunction. Only the modifiedargumentcomes

structurally, however. Theotherargumentinvolved in thesemanticrelationsuppliedby theadver-

bial mustbe resolved anaphorically. In orderto connecta discourseunit modifiedby a discourse

adverbial to theprior discourse,therefore,a structuralconnective mustbe employed (e.g. 0 , and,

etc.),whichsuppliesits own semanticrelation.Moreover, theremaybeadditionalinferredrelations

betweentheincomingunit andtheprior discourseover andabove boththerelationsuppliedby the

structuralconnectionandthe relationsuppliedby any discourseadverbials. For example,in DL-

TAG’sview, 0 andasa resultin (4.11)eachsupplyasemanticrelationbetweentheinterpretationsof

Johncamehomelate andMary left him; in addition,a temporalrelationbetweenthesetwo clauses

9This is notexplicitly statedfor but; however, by definitionit mustbethecase.Thesameargumentmayalsohold forcertainusesof and,or; see[WJSK03],footnote17, for discussion.

174

Page 192: Thesis about discourse

is inferred10.

(4.11)Johncamehomelate.As a result,Mary left him.

As discussedin Chapters2 and3, behavioral evidenceis presentedin [WJSK03]to supportthe

theoreticalview that discourseadverbialstake their prior argumentanaphorically. This evidence

includestherangeof waysthisargumentcancomefrom theprior discourse.For example,although

discourseconnectives aregenerallytaken assignallingdiscourserelationsbetweenadjacentdis-

courseunits,just ascanNP anaphora,discourseadverbialscanalsotake their prior argumentfrom

intra-sententialand implicit material. In (4.12) [WJSK03, 7], embeddedneverthelessrelatesthe

interpretationof the matrix clauseto the interpretationof the relative clause. This option is not

availableto subordinatingandcoordinatingconjunctionsbecausetheir argumentsareconstrained

to beadjacentdiscourseunitsof like syntactictype.

(4.12)Many peoplewho have developednetwork softwarehave neverthelessnever gottenvery

rich. (i.e. despitehavingdevelopednetworksoftware)

In (4.13)[WJSK03, 7], otherwisecanaccesstheinferredconditionof if thelight is notred. This

materialis not availableto structuralconnectives;or canonly accesstheconsequentclause(stop)

or thesentenceasa whole.

(4.13a) If thelight is red,stop.Otherwisegostraighton.

(4.13b) If thelight is red,stop,or gostraighton.

Wesaw furtherevidencein Chapter3, wherewe consideredtherangeof semanticmechanisms

underlyingthepredicateargumentstructureandinterpretationof S-modifyingadverbials.We saw

that we could distinguishdiscourseadverbialsandclausaladverbialsaccordingto whetheror not

their interpretationdependedon an abstractobject in the prior discourse.We saw that discourse

adverbialscontainsemanticargumentsinstantiatedasexplicit discoursedeicticreferenceto abstract

objects,demonstrative NP referenceto abstractobjects,comparative abstractobjects,relational

abstractobjects,etc. Becauseall of theseargumentsareanaphoricto abstractobjectsin theprior

discourse,thediscourseadverbialscontainingthemfunctionsemanticallyasdiscourseconnectives.

10SeeChapter2 for anexampleof how this inferenceis modeled(in [LA93]’ s discoursemodel).

175

Page 193: Thesis about discourse

Moreover, theanaphoricargumentinvolvedin therelationsuppliedby anadverbialmayor may

not resolve to theprior argumentinvolved in thestructuralrelation.In fact,a propertyof structural

connectives is that they do not allow crossingof predicate-argumentdependencies.For example,

while (4.14)[WJSK03, 5] is interpretableasanembeddedif S1,S2constructionwithin analthough

S1’,S2’construction,crossingthesedependenciesasin (4.15)[WJSK03, 5] is eitheruninterpretable

or at leastyieldsadifferentinterpretation;thedependenciesin theoriginal constructionsarelost.

(4.14)

a. AlthoughJohnis very generous-

b. if youneedsomemoney,

c. youonly have to askhim for it -

d. he’s very hardto find.

(4.15)

a. AlthoughJohnis very generous-

b. if youneedsomemoney -

d. he’s very hardto find-

c. youonly have to askhim for it.

It appearshowever thatdiscourseadverbialsdoallow crossingdependencies,asshown in (4.16)

[WJSK03, 6]. For thenin (d) to getits first argumentfrom (b), it mustcrossthestructuralconnection

betweenthe clausesin (c) and(d) that arerelatedby � .�i�j/k$(�. . Of course,as[WJSK03, 6] note,

anaphorafrequentlyshow crossingdependencies(e.g.JohnM told Mikeg heM wouldmeethimg later.

[WJSK03] show thatmodellingthesediscourserelationsstructurallywouldcreateadirectedacyclic

graph,which goesbeyondthecomputationalpower of LTAG and,moreoever, createsa completely

unconstrainedmodelof discoursestructure.

(4.16)

a. JohnlovesBarolo.

b. Soheorderedthreecasesof the’97.

c. But hehadto canceltheorder

d. becausethenhediscoveredhewasbroke.

It remainsto beshown however whetheror not all discourseadverbialsallow crossingdepen-

dencies;the fact that the felicity of constructedexamplescan be difficult to determinehas led

[FMP�

01] to outline a moreempirically-basedapproachto modelingdiscourseconnectives. As

176

Page 194: Thesis about discourse

their goal is to build the mostcomputationallyefficient discourseparserpossible,they arguethat

predicateargumentdependenciesshouldbedefinedstructurallywheneverpossible,regardlessof the

compositionalsemanticsof thepredicate.In otherwords,thedecisionto treata discourseconnec-

tiveanaphoricallywouldbebasedentirelyonwhethercorpusannotationindicatesit to benecessary

to avoid modelingcrossingdependenciesstructurally, ratherthanon their compositionalsemantics

or on constructedexamples. This approachmay help distinguishwhencompositionalsemantics

determineshow anadverbialretrievesits prior argumentandwhenit doesnot.

Certaindiscourseadverbials,however, arealreadyrepresentedin DLTAG astakingbothargu-

mentsstructurally. In particular, DLTAGdistinguishesparallel constructions,conveying disjunction

(“either...or”), contrast (“on theonehand...ontheotherhand”),addition“not only...but also”), and

concession(“admittedly...but”). Becauseof the interpretedinter-dependency of thediscoursecon-

nectivesin theseconstructions,they aremodeledwith initial trees,asin Figure4.15.

D 5� � � � � � ��������D 5Í Í ÍÎÎÎ

Ontheonehand D 5 Ë D 5Í Í Í ÍÎÎÎÎOn theotherhand D 5 Ë

Figure4.15:DLTAG Initial Treefor AdverbialConstructions

However, thesediscourseadverbialsdo not alwaysappearin theseparallelconstructions.The

majority of them,morefrequentlythesecondin thepair ([WJSK03]), canalsoappearalone,asin

(4.17),in whichcasethediscourseadverbialtakesthefirst auxiliary treein Figure4.14.

(4.17)Mike likesicecream.On theotherhand,hehatesmilk.

We have alreadydiscussedhow the two structurebuilding operationsin (D)LTAG work, we

now illustratewith theexamplein (4.18)theDLTAG treesthey produce.Wewill addressadditional

examplesin subsequentsections.

(4.18)On theonehand, I noticedthesolitudein New Hampshire,and thenI hopedwe could

stay. On theotherhand,whenI noticedthelackof multi-culturalism,I hopedwe would leave.

Werepresentclausesusingthesymbolfor initial treessubscriptedby thenameof themainverb

177

Page 195: Thesis about discourse

(e.g. �B�@����! ). In fact,oneof thebenefitsof DLTAG is that it parsesdiscourseon top of theclause

level parse;thuseachatomicclauseis itself a complex tree11. Thederivedtreefor (4.18)is shown

in Figure4.16.

D 5ð ð ð ð ð ð ð ð ð ð ð ðññññññññññññD 5� � � � � �������

On theonehand D 5� � � � �ùù������ ���� ���34! and D 5Ö Ö××then � ' "��9

D 5� � � � � �������On theotherhand D 5� � � �����

when � �7�� ���34! �B�&!���=�!Figure4.16:DLTAG DerivedTreefor Example(4.18)

The derivation tree for our exampleis shown in Figure4.17. We representinitial treeswith

discourseconnective anchorsase.g. ������� (= on theone/otherhand), auxiliary treeswith discourse

connectiveanchorsase.g.� "��!#� , andsubstitutionandadjunctionaddressesarerepresentedasabove

(i.e. with respectto theparenttree(e.g.(2.2)). Thereadershouldreferto theabove figuresto view

theelementarytreesthatcorrespondto eachdiscourseconnective.�B������ � � � ������� �7�� ��63'! (1.2)� ���7� (0)� 4 "��9 (3)� "��!#� (0)

���$�C!'� (2.2)Í Í ÍÎÎÎ� ���� ���34! (2) �B�&!���=%! (3)

Figure4.17:DLTAG DerivationTreefor Example(4.18)

11See[FMPÑ 01] for detailsof theDLTAG parser.

178

Page 196: Thesis about discourse

4.3.2 Syntax-SemanticInterfaces for DerivedTrees

In Chapter2 webriefly discussedtwo similarTAG-basedapproachesto incrementaldiscoursestruc-

ture, one([Web91]) which outlines“right frontier” constraintson discoursedeixis reference,and

another((LDM - see[Pol96, vdB96]) which outlineshow the“right frontier” andthediscoursere-

lation betweendiscourseunitsaccountfor constraintson theantecedentsof NP anaphorain local

discoursespans.Bothof theseapproachesemploy thesametwo operations([Gar97b]) for combin-

ing incomingelementarydiscourseunits(clauses),althoughtheir terminologyvaries,andbothusea

similarsyntaxsemanticinterfacewith respectto thederivedtree.Theoperations,calledattachment

andadjunctionin [Web91], andtheireffecton thesemanticinformationateachnode,areillustrated

againrespectively in Figure4.18.

Figure4.18: Illustrationof [Web91]’sAttachmentandAdjunctionOperations

In thefirst treefor theattachmentoperation,two nodeshave alreadycombined(by rootadjunc-

tion), andtheir semanticinformationhasbeencombinedin theparentnode,e.g. “(1,2)”. Attach-

mentof a third nodeto thisparentnodecreatesthesecondtreeandcausesthesemanticinformation

“(3)” associatedwith this incomingnodeto be incorporatedinto the semanticinformationof the

parentnodeto which it attachesto yield “(1,2,3)”. Roughly, attachment,e.g. inclusionin an ex-

isting discoursesegment,correspondsin bothapproachesto thesemanticlist relation.A discourse

correspondingto theresultof attachmentis shown in (4.19).

(4.19)I like summerandI like winterandI like autumn.

In thefirst treefor theadjunctionoperation,againtwo nodeshave alreadycombined(by root

adjunction),and their semanticinformationhasbeencombinedin the parentnode,e.g. “(1,2)”.

Adjunctionof a third nodeto this parentnode(root) createsa new discoursesegmentnode(whose

179

Page 197: Thesis about discourse

daughtersaretheoriginal parentandthe incomingnode)asshown in thesecondtree,andcauses

thesemanticinformationof bothchildrento beassociatedwith this new nodeto yield “((1,2),3)”.

Roughly, adjunction,e.g. creationof an embeddeddiscoursesegment,correspondsin both ap-

proachesto beginning a list relation,begining a temporalprogression,a causalrelation, etc. A

discoursecorrespondingtheresultof adjunctionis shown in (4.20),wherethefirst two clauseswill

beembeddedunderacausenode.

(4.20)Johnjoined thesoccerteamandMike joined the football teambecausethey wantedto

impresstheir fathers.

In bothapproaches,theseoperationsareconstrainedto applyto nodeson theright frontier (the

smallestsetof nodescontainingthe root suchthat whenever a nodeis on the right frontier, so is

its rightmostchild[Web91]). Wheretheapproachesdiffer is with respectto upwardpercolationof

semanticinformationasaresultof adjunctionatnodesotherthantheroot. [Web91] doesnotprovide

formal detailsconcerningthe syntaxsemanticinterfaceor explicitly addressthe issueof upward

percolation,however her examplesof adjunctionat a leaf show upward percolationof semantic

informationafteradjunctionat a leaf,asshown in figure. In this tree,afteradjunctionof i to � , not

only doestheir new parentnodecontainthe information(b,c), but this informationhaspercolated

up to therootnode,replacing(a,b)in thefirst treewith (a,(b,c))in thesecondtree.

Figure4.19:Webber’s AdjunctionataLeaf

In contrast,althoughLDM explictly defineshow theadjunctionoperationcombinestheseman-

ticsateachchild nodein thenew parentnode,they donotaddresstheneedfor upwardpercolation.

[Gar97b] arguesthat this lack of upward percolationis a problemfor LDM, becausethereis no

obviousway to readsemanticsoff treesandthereis no way to retrieve theavailableantecedentsof

180

Page 198: Thesis about discourse

discoursedeixisoff theright frontier of thetree,becausethenecessarysemanticinformationis not

madeavailablethere.Gardentillustratesthisusing[Web91]’sexample,first presentedin Chapter2

andrepeatedin (2.41)-(2.42).As Webbernotes,thediscoursedeicticreferencein (2.42)is ambigu-

ous;it canreferto any of thenodeson theright frontier of the(derived)tree:(thenodesassociated

with) clause(2.41e),clauses(2.41d)-(2.41e),clauses(2.41c)-(2.41e),clauses(2.41a)-(2.41e).

(2.41a)It’s alwaysbeenpresumedthat

(2.41b)whentheglaciersreceded

(2.41c)theareagot very hot.

(2.41d)TheFolsummencouldn’t adapt,and

(2.41e)they diedout.

(2.42)That’s what’s supposedto have happened.It’s thetextbookdogma.But it’s wrong.

UnderLDM’ s approach,the treefor (2.41) is shown in Figure4.20. Neitherthe root nor the

right frontier describesthesemanticsof this discoursecorrectly. For example,whatwe interpretas

presumedis not only � , but all of � - . , andwhatwe interpretwhenrelatingis not � and i , but � andi - . . Moreover, resolutionof that in (2.42)to � - . will incorrectlyyield asa semanticsfor b�VXj/b the

propositiondenotedby when(b,c)12.

presume(a,b)ú úûûa when(b,c)ú úûû

b cause(c,d)ú úûûc cause(d,e)� ���

d e

Figure4.20:DerivedTreefor Example(2.41)

[Gar97b] proposesDTAG,agrammarthatusesmodifiedversionsof thesyntactictree-construction

operationsand featurestructuresin feature-basedTAG (FTAG) (see[Gro99, Shi86]) alongwith

LDM’ ssemanticsandlexicon to solve LDM’ sproblemsarisingfrom lack of upwardpercolation.

12Lack of percolationis nota problemfor rootadjunctionbecauseit retainsthesemanticinformationof its children.

181

Page 199: Thesis about discourse

FTAG is identical to LTAG with respectto its structurebuilding operations.However FTAG

associateseachtreenode W , exceptfor substitutionnodes,with a top ( b � ) andbottom( � � ) feature

structure.Top featurescapturethe relationof W to its super-treeandbottomfeaturescapturethe

relationof W to its sub-tree.Substitutionnodeshave only top featuressincethetreesubstitutingin

logically carriesthebottomfeatures[Gro99, 8]. As shown in Figure4.21,aftersubstitution,thetop

featuresof thesubstitionnodeunify with the top featuresof thesubstitutednode,andthebottom

featuresareprovidedby thesubstitutednode.As definedin [JVS99], unificationcreatedtheunion

(U) of thespecifiedfeaturesandthereplacementof any featurevariableswith featuresthatcontain

valuesfor thesevariables.Onceprocessingatanodeis complete,its topandbottomfeaturesunify.

Y� b 9 g� 9 g �ü üýý + Xú úûû

Z YË [ b 9 M ] K XÍ Í Í ÍÎÎÎÎZ Y

� b 9 M U b 9 g� 9 g �ü üýýFigure4.21:Substitutionin FTAG

As exemplifiedin Figure4.22,after adjunction,the nodein X beingadjoinedinto splits, and

its top featuresunify with the top featuresof the root adjoiningnode,while its bottomfeatures

unify with thebottomfeaturesof thefoot adjoiningnode.Two setsof featuresallow thesemantic

relationshipsbetweenX andits sub-treeandsuper-treeto be maintainedafter adjunction.Again,

onceprocessingatanodeis complete,its topandbottomfeaturesareunified.

Xú ú úûûûA Y

� b 9 M� 9 M �þ þÿÿ+ Y

� b 9 g� 9 g �� � ����B YÌ � b 9 �� 9 � �

K XÍ Í ÍÎÎÎA Y

� b 9 M U b 9 g� 9 g �Í Í Í ÍÎÎÎÎB Y

� b 9 �� 9 M U � 9 ���þ þÿÿFigure4.22:Adjunctionin FTAG

182

Page 200: Thesis about discourse

In [Gro99]’sdescription,featurescapturehow predicatesconstrainor assignsemanticattributes

suchascase,agreement,numberetc, of the lexical items they take asarguments. For example,

prepositionsassignaccusative caseto their substitutedinternalNP arguments.In a PPelementary

tree,this assignmentis explicitly statedby an“assigncasefeature”[Gro99, 24], becauseNPsare

givena default casefeaturevalue.Without this feature,feature-unificationaftersubstitutionwould

notalterthedefault valueof thecasefeaturesof asubstitutedNP.

DTAG usesfeaturestructuresandoperationsto allow upwardpercolationof semanticinforma-

tion. DTAG contains1) asetof “discoursebasic”(B) trees,whicharesingle-nodetreesimplement-

ing the elementarydiscourseunits (DCUs) andtheir typedfeaturestructuresin LDM, 2) a setof

“discourserule” (R) trees,eachof which implementsa discoursegrammarrule of LDM, and3)

operationsfor combiningthem. Eachnodeof a DTAG treeis associatedwith two setsof feature

structuresmodeledafterthosein FTAG.

RecalltheLDM typedfeaturestructureof theDCU Johnsmiledrepeatedin Figure4.23,where

basicrepresentsDCU typeand (/z�{q| representsthesemanticsof Johnsmiled13.}~~��� j�(�+'iSEM (qz�{/|SCHEMA (qz�{/|

�&���Figure4.23:LDM ElementaryDCU

In DTAG, this featurestructureis identifiedwith boththetopandbottomfeaturesof anelemen-

taryDCU, asshown in Figure4.24.}~~~~~~~�TOP

}� � jq(C+'iSEM (qz�{/| ��BOTTOM

}� � jq(C+'iSEM (qz�{/| ���&��������

Figure4.24:DTAG ElementaryDCU

13Thevalueof theSCHEMAfeatureis identicalto theSEM featurein LDM elementaryDCUs.

183

Page 201: Thesis about discourse

RecallfurthertheLDM list grammarrule from Chapter2 shown againin Figure4.2514, which

statesthat any two discoursetreescancombineto form a new treeof type list that representsa

list relation(indicatedby SEM). The valueof an additionalSCHEMA featureis requiredto be a

non-trivial generalizationover themeaningof thetwo trees.}~~� ),+#(CbSEM ),+#(Cb�z4��.�_ M �%� .�_�gC|SCHEMA ^q.�W�z4� .�_�M*�%� .�_ g |

�&��� K}~~� -pb�op.C.SEM � .�_ MSCHEMA ��i�V;.�_�jXM

�&��� ,

}~~� -pb�op.C.SEM � .�_�gSCHEMA ��i�V;.�_�j g

�&���Figure4.25:LDM List Rule

Again, in DTAG eachnodeis expandedto containboth top andbottomfeatures,therebypro-

ducingthegeneralizedDTAG R treeversionshown in Figure4.2615. In this tree,R is instantiated

by thediscourserelationholdingbetweendaughters,(e.g. list, cause, etc.). As shown, thebottom

featuresof theparentnodeis theconjunctionof theapplicationof R to thebottomfeaturesof the

childrennodes(A andB) with top features(TA andTB) of thechildrennodes.}~� TOP î SEM T ïBOTTOM î SEM R(A,B) J TA J TB ï �&��ð ð ð ð ð ð ðñññññññ}~� TOP î SEM TAï

BOTTOM î SEM Aï � ��}~� TOP î SEM TBïBOTTOM î SEM Bï � ��

Figure4.26:DTAG R Tree

Gardentstatesthat her versionof the FTAG substitutionoperation,called - -substitution,is

unchangedexceptthatit is restrictedto thesubstitutionof any treeinto theright leaf of any � tree.

Her illustrationof the - -substitutionoperationis shown in Figure4.27.

DTAG’s versionof the FTAG adjunctionoperation,called - -adjunction,is also unchanged,

exceptthat it is limited to the right frontier andis no longerlimited to “recursive structures”(e.g.

14Thesegrammarrulesaresimplified; see[Pol96] for additionalinformation that is containedin a completeLDMfeaturestructure.

15Gardentdoesnot includetheSCHEMAfeature.

184

Page 202: Thesis about discourse

optionalmodificationstructureswhoseroot andfoot nodesareidentical).Her illustrationof the - -adjunctionoperationis shown in Figure4.28.Althoughnotshown in thefigure,[Gar97b, 12] states

that “on adjoining,the treedominatedby the leftmostdaughterof the local treebeingadjoinedis

‘closed-off ’ in thatall its b and � categoriesareunified(which in effect preventsany adjunctionto

thissubtree)”.This meansnodesnoton theright frontier arenotavailablefor processing16.

Figure4.27: [Gar97b]’s - -Substitution

Figure4.28: [Gar97b]’s - -Adjunction

As illustration,consider[Gar97b]’sexplanationof thederivationsof thediscoursein (4.21).For

clarity, sheremovesbracketingandfeaturestructurelabeling.

(4.21a) Dick doesnot cometo work

(4.21b) becausethetrainsaren’t running

(4.21c) andbusesaren’t either.

Gardentassertsthat theDCU for (4.21b)canbe - -substitutedinto thecausalR tree. This tree

is then - -adjoinedto the DCU for (4.21 a), making it unavailable for further processing,and its

16Gardentalsoarguesthattheattachmentoperationcanbemodeledusing ê -adjunction.

185

Page 203: Thesis about discourse

featuresunify. The result is shown in the first tree in Figure 4.29, where j and � representthe

eventualitiesdescribedby (4.21a)and(4.21b)respectively. TheelementaryDCU for (4.21c)is then- -substitutedinto thelist R tree.This treeis then - -adjoinedto the(b) nodeof thefirst tree,making

it unavailablefor furtherprocessingandits featuresunify, yielding thesecondtreein Figure4.29.

Whenprocessingis complete,all remainingtop andbottomfeaturesunify. The semanticsof the

rootbecomes:cause(a,b)J a J list(b,c) J b J c.�T

cause(a,b)J a J TB�Ï ÏÐÐ�a

a� �TB

b �K �

T

cause(a,b)J a J TB�� � � ������a

a� �TB

list(b,c) J b J TC�Ï ÏÐÐ�b

b� �TC

c �Figure4.29:First DTAG Derivationof Example(4.21)

Thereis anotherinterpretationof (4.21),however, namelythat both � and i togethercausej .Gardentarguesthatthis interpretationcanbederivedif (4.21c) first - -substitutesinto thelist R tree,

which then - -adjoinsto (4.21b), yielding thefirst treein Figure4.30. This treethen - -substitutes

into causalR tree.Theresultis then - -adjoinedto (4.21a),yielding thesecondtreein Figure4.30.

Whenprocessingis complete,all remainingtop andbottomfeaturesunify. The semanticsof the

rootbecomes:cause(a,(list(b,c)J b J c)) J a J list(b,c) J b J c.�T1

list(b,c) J b J TC�Ï ÏÐÐ�b

b� �TC

c �K �

T0

cause(a,(list(b,c)J b J TC)) J a J T1�� � � ������a

a� �T1

list(b,c) J b J TC�Ï ÏÐÐ�b

b� �TC

c �Figure4.30:SecondDTAG Derivationof Example(4.21)

186

Page 204: Thesis about discourse

In its essence,DTAG presentsa viable approachto obtainingthe compositionalsemanticsof

a discoursefrom its discoursetree. However, it is not clear that the featurestructuresin Figures

4.29and4.30havebeenconstructedaccordingto thedefinitionsof - -substitutionand - -adjunction.

“Irrelevant information” hasbeenomittedfrom thesefigures[Gar97b, 15]; if we stepthrougheach

stage,however, it is not clearthatwe would yield thesamefeaturestructures.Consideragainthe

secondderivationof (4.21).To avoid conflatingtheidentitiesof theTB variablesin thetwo R trees

involved in this derivation,we will useTX in onetree,andTZ in theothertree. In thefirst step,

theDCU for (4.21c) - -substitutesinto theright daughterof the list R tree,which by definitionof- -substitutionproducestheunionof thetopandbottomfeaturesasshown in Figure4.31.�T1

list(X,C) J TX J TC�ú úûû�TX

X � �TC

C �+�c

c� K �T1

list(X,C) J TX J TC�� � �����TX

X � �TC U c

C U c �Figure4.31:StepOnein theSecondDTAG Derivationof Example(4.21)

In the secondstep,the list R tree - -adjoinsto the DCU for (4.21 b), which by definition of- -adjunctionproducestheunionof thetop featuresof theDCU for (4.21b) with thetop featuresof

theparentnodeof the list R tree,andtheunionof thebottomfeaturesof theDCU for (4.21b) with

thebottomfeaturesof theleft daughternodeof the )r+�(�b R tree,asshown in Figure4.32.�T1

list(X,C) J TX J TC�� � �����TX

X � �TC U c

C U c �+�b

b� K �b U T1

list(X,C) J TX J TC�Í Í ÍÎÎÎ�TX

b U X� �TC U c

C U c �Figure4.32:StepTwo in theSecondDTAG Derivationof Example(4.21)

The result is not identicalto thefirst treein Figure4.30,which is Gardent’s representationof

this stage.Part of thedifferenceis dueto thefact that the leaf correspondingto (4.21b) hasbeen

pushedoff the right frontier, so accordingto the definition for - -adjunction,its top and bottom

187

Page 205: Thesis about discourse

featuresunify. Theresultis thatTB andB areinstantiatedasthesingleavailablevalue, � . We can

thenreplaceall otherinstancesof thesevariableswith � , producingthetreein Figure4.33.�b U T1

list(b,C) J b J TC�� � �����b

b� �TC U c

C U c �Figure4.33:StepThreein theSecondDTAG Derivationof Example(4.21)

This treeis still not identicalto Gardent’s version.Onereasonis thatwe have not fully unified

the rightmostleaf, becauseit is not clearwhat Gardenthasdone. ShehasreplacedC with i , but

shehasnot replacedTC with i . Both TC andC arevariablesand i is a value,andall areof like

type (e.g. BOTTOM-SEM and TOP-SEM,respectively). We will thus completethe derivation

without resolvingvariablesafter theunionof featuresat right frontier nodes,andseewhatresults.

Anotherdifferenceis thatGardenthasnot shown theunion“b U T1” in theparentnode,aswould

beexpectedby - adjunction.Gardentdoessaythat - -adjunctionto a leaf is a specialcase,but this

unionwouldalsobeproducedafteradjunctionat othernodes.Wewill thusincludethisunion.

In stepfour the list tree - -substitutesinto the right-mostleaf of a causeR tree,producingthe

unionof thetopandbottomfeatures,asshown in thethird treein Figure4.34.Notethatin Gardent’s

version(thesecondtreein Figure4.30),GardenthasidentifiedZ with list(b,c)J bJ TC, andshehas

also identifiedTZ with T1, in contrastto above,whereshedid not identify TC with i .�T0

cause(A,Z)J TA J TZ�Ï ÏÐÐ�TA

A � �TZ

Z �+

�b U T1

list(b,C) J b J TC�� � �����b

b� �TC U c

C U c �K �

T0

cause(A,Z)J TA J TZ�� � � ������TA

A � �TZ U b U T1

Z U list(b,C)J bJ TC�� � �����b

b� �TCU c

C U c �Figure4.34:StepFour in theSecondDTAG Derivationof Example(4.21)

In stepfive, theresultingcausetree - -adjoinsto theDCU for (4.21a),yielding thethird treein

188

Page 206: Thesis about discourse

Figure4.35.�T0

cause(A,Z)J TA J TZ�� � � ������TA

A � �TZ U b U T1

Z U list(b,C)J bJ TC�� � �����b

b� �TC U c

C U c �

+�a

a� K �aU T0

cause(A,Z)J TA J TZ�� � � � �������TA

a U A� �TZ U b U T1

Z U list(b,C)J bJ TC�� � �����b

b� �TC U c

C U c �Figure4.35:StepFive in theSecondDTAG Derivationof Example(4.21)

Now theleafcorrespondingto (4.21a)hasbeenpushedoff theright frontier, soby - -adjunction,

its variablesareinstantiatedas j , asareall instancesof TA andA, producingthetreein Figure4.36.�a U T0

cause(a,Z)J a J TZ�� � � ������a

a� �TZ U b U T1

Z U list(b,C)J bJ TC�� � �����b

b� �TC U c

C U c �Figure4.36:StepSix in theSecondDTAG Derivationof Example(4.21)

At this point thereis no moreinput andso all top andbottomfeaturesunify. The problemis

that in mostnodescontainingvariableswe no longerhave a singlevaluewith which they canbe

identified.We do have a singlevaluein our rightmostleaf, sowe unify TC andC with i , andthen

identify all otherinstancesof TC andC with i , producingthetreein Figure4.37.

Again,however, this treedoesnot correspondto Gardent’s intendedfinal derivation for (4.21).

In orderto reproduceherintendedderivation,wehave to understandwhy topfeaturesof elementary

DCUsarenot alwaysidentifiedwith top featuresin thesubstitutednodeaftersubstitution,andwe

have to understandwhy top featuresof nodesbeingadjoinedto arenot unified with top features

in theparentnodeof theadjoiningtree. As statedabove, [Gar97b, 12] doesassertthatadjunction

to a leaf is a “specialcase”,but shedoesnot defineit, andthe sameproblemswould ariseafter

189

Page 207: Thesis about discourse

adjunctionatothernodeson theright frontier.�a U T0

cause(a,Z)J a J TZ �� � � ������a

a� �TZ U b U T1

Z U list(b,c)J bJ c�÷ ÷øø�b

b� �c

c�Figure4.37:StepSevenin theSecondDTAG Derivationof Example(4.21)

One way to reproduceGardent’s final derivation would be to explicitly statehow features

“unify”after adjunctionand substitution. We would requirethat in substitution,the bottom fea-

turesof the parentnodearereplacedby the bottomfeaturesof the substitutingnode,but the top

featuresof thesubstitutingnodearereplacedby the top featuresof the incomingnode. In Figure

4.27,thiswouldbe � � K � � and b � K b � . Wewouldrequirethatin adjunction,thebottomfeatures

of the left daughterof theadjoiningtreearereplacedby thebottomfeaturesof the treebeingad-

joinedto, but top featuresof theadjoiningtreereplacetop featuresof thetreebeingadjoinedto. In

Figure4.28,thiswouldbe � ��� K � � and b � K b � . Alternatively, wemightdefinesubstitutionnodes

ashaving only top featuresasin FTAG, andelementaryDCUsashaving only bottomfeatures;we

would thenonly have to definehow adjunctionaffectedfeatureunification.

4.3.3 A Syntax-SemanticInterface for DLTAG Derivation Trees

Becausein DLTAG, discourseconnectivesarepredicates, bothsyntacticallyandsemantically, DL-

TAG canbuild boththesyntaxand thecompositionalsemanticsof thesepredicatesusingthesame

syntacticandsemanticmechanismsthatareusedto build thesyntaxandcompositionalsemantics

of predicatesat theclauselevel.

As discussedabove, DLTAG currentlyusestheLTAG grammarto build discoursesyntax. As

such,in thissectionweexplorehow the[JKR03,Kal02, KJ99, JVS99]syntax-semanticinterfacefor

LTAG,whichemploysaflat semanticrepresentationusingideasfrom Minimal RecursionSemantics

190

Page 208: Thesis about discourse

([CFS97])andHole Semantics([CFS97,Bos95]), canalsoapply to DLTAG. This sectionwill be

exploratoryratherthanconclusive,asacompletesyntax-semanticinterfaceincorporatingall aspects

of DLTAG anddiscourseconnectivesrequiresa thesisin its own right.

As introducedin Section4.2, [JKR03, Kal02, KJ99, JVS99] argue that producinga compo-

sitional semanticsfor LTAG shouldcorrespondto establishingsemanticpredicate-argumentrela-

tionshipsusingthe LTAG derivation tree,becauseit is this tree,ratherthanthe derived tree,that

reflectstheserelationships.To illustratethe basicsof the extensionof their interfaceto DLTAG,

we begin with a few simpleexamplesof two-sentencediscoursesconnectedby structuraldiscourse

connectives.Wewill thenconsidermorecomplex discoursesanddiscusswhy [JKR03, KJ99]’suse

of multipcomponentTAGsand/or[Kal02]’s enrichedderivation treeis sometimesneededto build

thecompositionalsemanticsof discourse.

Considerfirst (4.22),wheretwo clausesarelinkedby thesubordinatingconjunctionbecause.

(4.22)JohnlikesMary becauseshewalksFido.

The elementaryDLTAG treesfor (4.22) are shown in Figure 4.38. SC representssubordi-

nating conjunction. Although the internalstructureof clausesis accessiblein the DLTAG parse

([FMP�

01]), in all of ourexamplesin thissectionwe modelatomicclauseswith elementaryinitial

trees(e.g. �B�@����! ). As discussedabove, D 5 is a genericdiscourseunit root label usedto represent

bothatomicclausesandconstituentscomposedof clausesandpredicates.

D 5� � � �����D 5 Ë SC

because

D 5 Ë D 5�B�@����! D 5��������Figure4.38:DLTAG ElementaryTreesfor Example(4.22)

Figure4.39shows thesemanticrepresentationsof theelementarytreesin Figure4.38. In these

andall of our following semanticrepresentationswewill employ thesemanticsoutlinedin [JKR03,

Kal02, KJ99]. In our discourse-level extension,however, their propositionallabels,e.g. ) � , will

now be associatedwith the semanticsof j/)E) discourseunits, both atomicandcomplex, andtheir

191

Page 209: Thesis about discourse

propositionalvariables,e.g. ( � , will now beusedasdiscourseunit variables,whosevaluesarethese

labels. Furthermore,we will employ a simplified representationof clausesemantics,akin to the

treatmentof NPsin [Kal02], in which thevaluesof theclauseargumentsarealreadyprovided.For

example,we will use ) g��p),+#äy.��2z�{p��_Å| for thesemanticvalueof theclauseJohnlikesMary.��12!�34��5*'! �B�@����!� ���$�%�&��l M : because’(sM , sg )——————————arg: Û sM ,(1)t , Û sg ,(3)t l g : like’(j, m)

—————arg: –

l � : walk’(m, f)—————arg: –

Figure4.39:SemanticRepresentationof �B12!#3'��5*#! , �B�@����!� and ���������As shown, becausetakes two substitutedargumentsthat are associatedwith positionsin its

elementarytree. Figure4.40 shows the derived andderivation treesthat result from substituting�B�@����! and ���$�%�&� into thesepositionsin �B12!#34��5*#! , alongwith thesemanticrepresentationthatresults.

D 5� � � � ������D 5�B�@����! SC

because

D 5���������B12!#3'��5*#!� � �������&���8! (1) �������&� (3)

l M : because’(lg , l � ), l g : like’(j, m), l � : walk’(m, f)———————————————————–arg: –

Figure4.40:DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.22)

As discussedin Section4.2, combiningsemanticrepresentationsin this approachconsistsof

building theunionof thesemanticrepresentationsof theelementarytreesinvolvedin thederivation

andassigningvaluesto argumentvariables,wherethederivationtreeindicateshow argumentvari-

ablesgetvalues:whena treeis substituted,its valueis appliedto theargumentvariablepairedwith

thepositionat which it attaches.For example,thederivation treein Figure4.40indicatesthat the

sM argumentvariablein thebecauseelementarytreeshouldbeassignedthe labelof �B�@����! because�B�@����! substitutesin at position(1) andposition(1) is associatedwith thesM argumentvariable. In

thesameway, thederivation treeindicatesthat sg shouldbeassignedthe label of ���$�%�&� . Theflat

192

Page 210: Thesis about discourse

semanticsshown in Figure4.40leadsto theembeddedsemanticrepresentationin (4.23).

(4.23)like’(j, m) J like’(m, f) J because’(like’(j, m), walk’(m, f))

Theanalysisof a discoursecontainingthecoordinatingconjunctionand andtwo clausalargu-

mentsproceedssimilarly17. Considerfor examplethediscoursein (4.24).

(4.24)Johnsaw Mary andhekissedher.

The elementaryDLTAG treesfor (4.24)areshown in Figure4.41. CC indicatescoordinating

conjunction. This analysisalsoappliesto theemptyconnective, 0 , asdiscussedabove.

D 5Í Í ÍÎÎÎD 5 Ì CC

and

D 5 Ë D 5� #!#! D 5�B�8���Figure4.41:DLTAG ElementaryTreesfor Example(4.24)

Figure4.42shows thesemanticrepresentationsof theelementarytreesin Figure4.41.Only the

substitutedargumentof and(sg ) is linkedto anaddress;asdiscussedin Section4.2,in asubstitution

stepat position P , the argumentvariablelinked to P will get the valueof the substitutedelement,

andin anadjunctionstep,theargumentvariableof theincomingtreeis addedto therepresentation

andassignedthevalueat thepositionwhereit attaches.� ����� � #!#! ���8�"�l M : and’(sM , sg )————————arg: sM , Û sg ,(3)t l g : see’(j,m)

—————arg: –

l � : kiss’(j, m)—————arg: –

Figure4.42:SemanticRepresentationof� ���7� , � '!�! and �B�8�"#

Figure4.43shows thederivedandderivation treesthatresultfrom substituting�B�8�"# into� ���7�

andadjoining� ���7� to � #!#! , alongwith thesemanticrepresentationthatresults.Again,thederivation

treein Figure4.43 indicatesthat thevalueof thesg argumentvariablein theand elementarytree

is assignedthe semanticlabel of ���8�"# , andfurther indicatesthat the sM argumentvariablein the

17As discussedabove,so,but take aninitial tree,andareanalysedlike because.

193

Page 211: Thesis about discourse

andelementarytreeis addedto therepresentationandassignedthesemanticlabelof � #!�! . Thisflat

semanticrepresentationleadsto theembeddedsemanticrepresentationshown in (4.25).

(4.25)see’(j,m) J kiss’(j, m) J and’(like’(j, m), kiss’(j, m))

D 5Í Í Í ÍÎÎÎÎD 5� '!�! SC

and

D 5�B�8�"#� '!�!� ���7� (0)�B�%�"� (3)

l M : and’(lg , l � ), l g : see’(j,m), l � : kiss’(j, m)——————————————————arg: –

Figure4.43:DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.24)

Considernow themorecomplex discoursein (4.26). Theargumentsof becauseareitalicized,

andthe argumentsof and arebracketed. To make it clear that thesearethe interpretedsemantic

argumentsof eachconnective,we includeacontextual question.Figure4.44shows thederivedand

derivation treesandthe semanticrepresentationfor (4.26) after substituting�����D;��� ¿�¿ 9 and �B���6�7�into ��12!�34��5*'! , substituting���"�DA#��� into

� ���7� , andadjoining� ����� to the o7T7T�b of the � .Ci�j`k�(*. tree.

(4.26)(Who is happy andwho is sad?)

[Mary is happybecauseherhusbandfounda job] and[Johnis sad].

D 5� � � � � �ú ú ú������D 5� � � � �ö ö�����

D 5���"�D;��� ¿�¿ 9 SC

because

D 5�B���6�7�CC

and

D 5�����DA#����B12!#34��5*#!� � � � � �� ����������"�D;��� ¿�¿ 9 (1) �B���6�7� (3)

� ����� (0)���"�DA'�%� (3)

l � : is-happy’(m), l ç : find’(h, j), l : is-sad’(j),l M : because’(l� , l ç ), l g : and’(lM , l )————————————————————————————————-arg: –

Figure4.44:DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.26)

Thisflat semanticrepresentationleadsto theembeddedsemanticrepresentationshown in (4.27).

194

Page 212: Thesis about discourse

This is theexpectedinterpretation;by adjoiningto therootof thebecausetree,thevalueof theand

adjunctionargumentvariableis thelabelcorrespondingto theentirebecausetree.

(4.27)is-happy’(m) J find’(h, j) J is-sad’(j)J because’(is-happy’(m), find’(h, j)) J and’(because(is-happy’(m), find’(h, j)), is-sad’(j))

But now considerthesimilardiscoursein (4.28).Again,theargumentsof becauseareitalicized,

the argumentsof and arebracketed,and to make it clear that theseare the interpretedsemantic

argumentsof eachconnective,we includeacontextual question.Figure4.45shows thederivedand

derivation treesandthe semanticrepresentationfor (4.28) after substituting�����D;��� ¿�¿ 9 and �B���6�7�into ��12!�34��5*'! , adjoining

� ����� to theleaf ( �����6��� ) of theresult,andsubstituting�B�@����! into� ���7� .

(4.28)(Why is Mary happy?)

Mary is happybecause[herhusbandfounda job] and [helikesit].

D 5� � � � � � � �������D 5���"�D;��� ¿�¿ 9 SC

because

D 5Í Í Í ÍÎÎÎÎD 5�����6��� CC

and

D 5���&���8!��12!�34��5�'!Í Í Í ÍÎÎÎÎ���"�D;�C� ¿�¿ 9 (1) �����6��� (3)� ���7� (0)�B�@����! (3)

l � : is-happy’(m), l ç : find’(h, j), l : like’(h, j), l M : because’(l� , l ç ), l g : and’(lç , l )————————————————————————————————-arg: –

Figure4.45:DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.28)

Thisflat semanticrepresentationleadsto theembeddedsemanticrepresentationshown in (4.29).

(4.29)is-happy’(m) J find’(h, j) J like’(h, j)J because’(is-happy’(m), find’(h, j)) J and’((find(h,j)), like’(h, j))

Assumingthat Mary beinghappy is causedby both her husbandfinding a job and liking it,

this is not the interpretationproducedby this approach,wholly becausetheand adjunctionis not

retrievedfrom theargumentposition(3) of becausein thederivationtree18.

18Note that theproblembecomesenormouslymorecomplex whenwe considerthe currentDLTAG parser’s “lowestadjunction”default for theemptyconnective. See[FMP Ñ 01] for detailsaboutthisdefault.

195

Page 213: Thesis about discourse

In fact,thesameproblemarisesin clause-level conjunction.Considerfor example(4.30).

(4.30)JohnlikesMary andSue.

In (4.30),theunit Mary andSueis built by adjunctionof an j`WU- treeto � j/o�� , andthesubsti-

tutionof Sueinto theandtree.Theand treehereis identicalto auxiliaryandtreein DLTAG,except

that its root, substitutionandadjunctionnodesareall labeledNP. This tree,alongwith theLTAG

derived andderivation treesfor (4.30), are shown in Figure4.46. As shown, the compositional

semanticsof theverbis likes(j,m), because� � ��<�9 is whatsubstitutesinto address(2.2).

NPÍ Í ÍÎÎÎNPÌ CC

and

NPË SÍ Í Í ÍÎÎÎÎNP

John

VPÍ Í Í ÍÎÎÎÎV

likes

NPÍ Í Í Í� �ÎÎÎÎNP

Mary

CC

and

NP

Sue

���&�6��!Í Í ÍÎÎÎ�� %����� (1) � � ��<�9 (2.2)� ����� (0)� 45*! (3)

Figure4.46:LTAG DerivedandDerivationTreesfor Example(4.30)

Onesolutionat theclauselevel is to make theNP-conjoiningj/WU- treeinto an initial tree,but

althoughthis would solve theproblemin (4.30),it is not only NP conjunctionthatyields this type

of problem;any adjunctionto a substitutedelementwill not be reflectedin thesecompositional

semantics19. This solutioncouldalsobeappliedat thediscourselevel, e.g. if DLTAG modelsand

(and 0 ) with initial trees;thentheand treein example(4.28)couldsubstituteinto position(3) in the

becausetree,andproducethedesiredinterpretation.

Thereis moregeneraltechnicalsolutionproposedin [Kal02] to dealwith scopein Frenchquan-

tifiers,whichhave beenanalyzedasadjunctsthatadjointo anNP, asillustratedin Figure4.47.We

overlookdetailsof thesemanticsof Frenchquantifiershere;roughlystated,thequantifiersemantics

19This solutionfor NP coordinationwassuggestedseparatelyby LauraKallmeyer andAravind Joshi,personalcom-munications.As Aravind Joshinotes,it is naturalfor NP coordinationto be representedin LTAG with an elementarytreebecauseany NP treein LTAG hasthe (root) label NP; however, VP, whosecoordinationcanalsobe viewed asScoordinationin which two subjectnodesareidentified,doesnot exist in LTAG asanelementarytreewhoseroot label isthesameasanelementarytreethatcouldconjoin it; rather, theroot labelof theVP treeis S. He suggestsinsteadthataVP coordinationtreebeconstructeddynamicallyduringparsing.

196

Page 214: Thesis about discourse

requiresaccessto the verb semantics,andvice versa,but thereis no link (edge)in the derivation

treebetweenthem20.

NÖ Ö××DET

chaque

N Ì + N

chien

+ S÷ ÷øøNË VP

V

aboie

K SÍ Í ÍÎÎÎNú úûû

DET

chaque

N

chien

VP

V

aboie

, aboie

chien(1)

chaque(0)

Figure4.47:Quantifiersin French

[Kal02, 104]’s solutionto this problemis to enrichthederivation treewith anadditionaledge

betweenthequantifierandtheVP, asshown in Figure4.48.

aboie

chien(1)

chaque(0)

Figure4.48: [Kal02]’s . -Edgesfor Quantifiersin French

This additionallink, it is argued,makesexplicit the intuitive dependency relationshipswe can

alreadyseein thederivationtreebetweentheVP treeinto whichchiensubstitutes,andelementsthat

adjointo chien. In otherwords,[Kal02] proposesthat in thecaseof anadjunctionat theroot node

of someelementarytree � (in this casechien), the adjoinedtree(in this casechaque) is not only

linkedby anedgeto � , but is alsolinkedby anadditionaledgeto thetreeto which � wasaddedin

somepreviousderivationstep,which is in thiscasethetreefor aboie. [Kal02] notesthatin factthis

additionaledgereflectstheunificationof featuresaftersubstitutionof annodeW (whichcreatesthe

unionof top features(seeFigure4.21))followedby adjunctionat W (whichcreatestheunionof the

top featuresin theparenttreewith thetop featuresof W (seeFigure4.22)).

More generally, [Kal02] proposesthe following definition of the e-derivation graph(it is no

longera tree): all edgesin the derivation treeareprimary e-derivation edges.Furthermore,there

20Translation:chaque= each; chien=dog; aboie=barks

197

Page 215: Thesis about discourse

is a secondary. -derivation edge( . -edge)betweentwo nodesª and�

if, in the derivation tree,

therearenodesª ’,� M ,...� � suchthat ª ’ is a daughterof ª ,

� M is a daughterof ª ’ with position0

(adjunctionat therootnodeof ª ’),� � � M is adaughterof

� � with position0 (1 � i Û n) and� � =

�.

Thisdefinitionis sketchedin Figure4.49. ªª ’� M (0)

...� � (0)

Figure4.49: [Kal02]’s . -DerivationGraph

[Kal02] then redefineshow substitutionand adjunctionyield the semanticrepresentationto

incorporatethe contribution of . -edges.By her new definitions,a variablein the semanticrepre-

sentationof a nodecanobtain its value either from a tree linked to it by a primary edgein the

derivation tree,or from a treelinked to it by a secondary. -edgein the . -derivation structure.For

example,avariablein thesemanticrepresentationof aboiecanobtainits valuefrom eitherchienor

chaque. Theonly furtherrequirement,obviously, is that thevariableandthevalueit obtainsareof

like semantictype.

Returningnow to the DLTAG example(4.28),what addingthe additional . -edgeseffectively

doesis createambiguityasto the interpretedsemanticargumentof because, becauseall variables

andvalueswe have consideredso far areof like semantictype (e.g. discourseunits). Consider

the . -derivationgraphfor (4.28),shown in Figure4.50alongwith thederivedtree. If we compute

semanticsbasedonthis . -derivationstructure,thenin additionto thesemanticrepresentationof this

discoursethatwealreadydiscussed,shown in Figure4.45(wherebecausetakesl � : ���"�D;��� ¿�¿ 9 andl ç :�����6��� asits arguments),thereis anotherpossiblesemanticrepresentationfor this discourse,shown

in Figure4.51.Thisadditionalinterpretation,which leadsto theembeddedsemanticrepresentation

shown in (4.31),is theintendedinterpretation.It is producedby theadditional. -edge,whichmakes

198

Page 216: Thesis about discourse

availablethevalueof theand treeto theargumentvariablein thebecausetree.

(4.31)is-happy’(m) J find’(h, j) J like’(h, j) J and’(find’(h, j), like’(h, j))J because’(is-happy’(m), and’(find’(h, j), like’(h, j)))

D 5� � � � � � � �������D 5���"�D;��� ¿�¿ 9 SC

because

D 5Í Í Í ÍÎÎÎÎD 5�����6��� CC

and

D 5���&���8!��12!�34��5�'!Í Í Í ÍÎÎÎÎ���"�D;�C� ¿�¿ 9 (1) �����6��� (3)� ���7� (0)�B�@����! (3)

Figure4.50:DLTAG DerivedTreeand . -DerivationGraphfor Example(4.28)

l � : is-happy’(m), l ç : find’(h, j), l : like’(h, j), l M : because’(l� , l g ), l g : and’(lç , l )————————————————————————————————-arg: –

Figure4.51:Additional SemanticRepresentationfor (4.28)dueto . -DerivationGraph

The questionthenbecomesoneof whetherwe want this ambiguity. We certainlywant some

ambiguity, the problemis that we may not want the first interpretationof (4.28) to be possibleat

all. Thatis, we don’t wantto producearepresentationfor (4.28)wherebecausetakesl � : �����D;��� ¿�¿ 9andl ç : �B���6�7� asits arguments,andand takes l ç : �����6��� andl : ���&���8! , becauseit is not clear that

this is even a possibleinterpretationof of (4.28). Nor, perhaps,do we want the . -edgein Figure

4.50to allow j`WU- to resolve its adjunctionvariableto the interpretationof because. While doing

so introducesa possibleinterpretation,in fact this is the sameinterpretationalreadyachieved by

adjunctionof theand treeto therootof because, aswe saw in (4.26).

In otherwords,thereappearto beonly two possibleinterpretationsfor (4.28),oneis achieved

by adjunctionof and to the root of the becausetree, and the other is madeavailable by the . -edgethat allows becauseto take the and tree as its secondargument. The third representation,

wherebecausetakes �����6�7� asits secondargumentandand takes �����6�7� asits first argument,is not

clearly interpretable.To prevent suchunwantedinterpretationsarisingfrom DLTAG . -derivation

graphs,we could of coursedefineconstraintson the effect that . -edgescanhave on the valueof

199

Page 217: Thesis about discourse

semanticargumentvariableswhosevaluesarediscourseunit labels. As illustration, let us define

two constraintsthatrestrictthesemanticvaluesmadeavailableby DLTAG . -edges:

Constraint(1): if thereis an . -edgebetweentwo elementaryDLTAG treesthat both take two

structuralarguments(via substitutionor adjunction)whosevaluesareboth discourseunit labels,

thenthis . -edgewill determinethevalueof theargumentof thehigherelementarytree.

Constraint(2): theargumentvalueproducedby Constraint(1) is theonly additionalargument

valuemadeavailableby . -edgesbetweenDLTAG elementarytrees21.

Webriefly illustratetheeffectof theseconstraintsby consideringthediscoursein (4.32),where

the intendedargumentsof becausearebracketed. Thederived treeand . -derivation graphfor this

exampleareshown in Figure4.52.

(4.32)[Mary is happy] because[herhusbandfoundajob andhelikesit andit paysgoodmoney].

D 5ð ð ð ð ð ð ð ðûû ññññññññD 5�����D;��� ¿�¿ 9 SC

because

D 5� � � � �ÚÚ �����D 5�����6�7� CC

and

D 5Í Í Í ÍÎÎÎÎD 5�B�@����! CC

and

D 5� ¿ ��9

�B12!�34��5*#!Í Í Í ÍÎÎÎÎ���"�D;��� ¿�¿ 9 (1) �����6��� (3)� ���7� (0)�B�@����! (3)� ���7� (0)� ¿ ��9 (3)

Figure4.52:DLTAG DerivedTreeand . -DerivationGraphfor Example(4.32)

As thederived treealreadyshows, andConstraint(1) requires,we want thehigherand to take

thevalueof the lower and asthevalueof its substitutionargumentvariable,andwe wantbecause

to take the valueof the higherand asthe valueof its substitutionvariable. Constraint(2) would

prevent the . -edgesfrom allowing any othersemanticinterpretations,suchasthelower and taking

the valueof the higherand asthe valueof its adjunctionargument– becausethis canalreadybe

achievedby adjunctionto therootof thehigherand. Essentially, however, it shouldbeobviousthat

21This cannotbestipulatedat theclauselevel, for chaqueneedsaccessto thelabelof aboie.

200

Page 218: Thesis about discourse

combining[Kal02] e-derivation graphswith constraintsthat reducethe interpretationsyieldedby

thesegraphsis anattemptto avoid asituationwheretheDLTAG derivationstructure(treeor graph)

producesthewrongresultandthusis notabetterbasisfor semanticinterpretationthansimplyusing

theDLTAG derivedtreein thefirst place.

Thereis an alternative solutionproposedin [JKR03], however, which allows us to derive all

andonly the intendedsemanticinterpretation(s)from theoriginal DLTAG derivation treewithout

requiringtemperingconstraintsor additional . -edges.This solutionemploys thenotionof flexible

directionof composition22, which we illustratefirst usinga context-freegrammar(CFG)rule such

asA K BC. To produceA, wecanview B asa functionandC asits argument,or wecanview C as

thefunctionandB asits argument.BecauseCFGsprovide string rewriting rules,in which function

andargumentare‘string-adjacent’strings,thisuseof flexible compositioneffectsneithertheweak

generative capacity(setof stringsgenerated)nor the stronggenerative capacity(setof derivation

treesgenerated)of theCFG.

A TAG, however, providestreerewriting rules. Functionandargumentin a TAG arecomplex

topologicalarguments(trees)thatare‘tree-adjacent’;thus,dependingonhow it is specified,theuse

of flexible compositionin aTAG canpotentiallyeffectbothits weakandstronggenerativecapacity.

[JKR03] specifytheuseof flexible compositionin aTAG asfollows: if a tree b composesinto atreek , k mustbeanelementarytree. If both b and k areelementarytrees,thedirectionof composition

cango eitherway. If both b and k arederivedtrees,they cannotcomposewith eachother. Roughly

stated,this definitionof flexible compositionallows thederivation treeto betraversedin a flexible

mannerwhile ensuring‘tree locality’: thederivation treecanbetraversedstartingat any node,but

as traversalcontinuesthe growing complex (derived) treecanonly composeinto ’tree-adjacent’

elementarytrees.

To take a simpleexample,considerthe constructionof theclauseFunnypeoplesmile, whose

syntaxandsemanticscanbe producedthroughcompositionof the threeelementaryLTAG trees

shown in Figure4.53alongwith theresultingderivedandderivation trees.Note thatalthoughthe

derivation treerepresentshow theseelementarytreescombine,andthederived treerepresentsthe

22This discussionof flexible compositionin TAG is derivedfrom [JKR03, 8].

201

Page 219: Thesis about discourse

resultingclausestructure,neitherof thesetreesshow thetraversalorderin which theseelementary

treeswerecomposed.By theTAG definitionof flexible composition,wecanfor examplefirst com-

pose� ¿ !�� ¿ �&! and �� � ���&! andthencomposetheresultingderived treeinto� ��5�����9 23. Alternatively,

we canfirst compose� ��5����*9 and � ¿ !#� ¿ �! , andthencomposetheresultingderivedtreeinto �B � ���! .

NPÖ Ö××ADJ

funny

NPÌ + NP

people

+ SÖ Ö××NPË VP

V

smile

K SÍ Í ÍÎÎÎNPú úûû

ADJP

funny

NP

people

VP

V

smile

, �� � ���&!� ¿ !#� ¿ �! (1)� ��5��*��9 (0)

Figure4.53:Flexible Compositionin LTAG

Althoughthesetraversalordersareall possibleby thedefinitionof flexible compositionin TAG,

only thelatterbottom-uptraversalyield thedesiredsemantics.As illustration,supposetheseman-

tics correspondingto � ¿ !�� ¿ �&! is people’(x), thesemanticscorrespondingto� ��5����*9 is funny’(y), and

thesemanticscorrespondingto �B � ���! is smile’(z). If we first compose� ¿ !�� ¿ �&! and �B � ���&! , then sis identifiedwith people’(x); subsequentcompositionof this derived treewith

� ��5����*9 identifies �with people’(x)too, yielding theclausesemantics:smile’(people’(x))J funny’(people’(x)). How-

ever, if we first compose� ¿ !#� ¿ �! with� ��5�����9 , then � will be identifiedwith people’(x), yielding

funny’(people’(x)). Substitutingthecorrespondingderivedtreeinto �� � �6�! will causes to beiden-

tified with funny’(people’(x)), yielding thedesiredsemantics:smile’(funny’(people’(x))).

In this thesis,it is the impactof flexible compositionon the computationof semanticsfrom

theDLTAG derivation treethat is our mainconcern.First, considerhow flexible compositionand

an assumptionof bottom-uptraversalof the DLTAG derivation tree producesonly the intended

interpretationof (4.28),repeatedbelow, wheretheintendedargumentsof becausearebracketed.

(4.28)(Why is Mary happy?)

[Mary is happy] because[herhusbandfounda job andhelikesit].

Thederivedandderivationtreesfor this examplearerepeatedin Figure4.54.We have already

seenhow a top-down traversalof thederivation treeyieldsanunintendedsemanticrepresentation23Allowing simultaneouscompositionis alsonecessarysoasnot to excludestandardTAG derivations[JKR03].

202

Page 220: Thesis about discourse

for this clause(seeFigure4.45),Figure4.54shows the intendedsemanticrepresentationfor this

example,which results,very simply, from abottom-uptraversalof thederivationtree.

D 5� � � � � � � �������D 5���"�D;��� ¿�¿ 9 SC

because

D 5Í Í Í ÍÎÎÎÎD 5�����6��� CC

and

D 5���&���8!��12!�34��5�'!Í Í Í ÍÎÎÎÎ���"�D;�C� ¿�¿ 9 (1) �����6��� (3)� ���7� (0)�B�@����! (3)

l � : is-happy’(m), l ç : find’(h, j), l : like’(h, j), l M : and’(lç , l ) l g : because’(l� , l M )————————————————————————————————-arg: –

Figure4.54:DLTAG DerivedandDerivationTreesandSemanticRepresentationfor (4.28)

We will now considertheanalysisof two-sentencediscoursesin which thesecondsentenceis

modifiedby adiscourseadverbial.Wewill show thattheuseof flexible composition,whichallows

a bottom-uptraversalfor DLTAG derivation trees,is sometimescrucial to obtainingthe intended

semanticinterpretation.We will further illustratehow [KJ99]’s enrichedderivation treecould be

usedinstead,but would yield greatersemanticambiguity.

As discussedin Chapter3, the distinctionbetweenclausaladverbials,which do not function

semanticallyasdiscourseconnective, anddiscourseadverbials,which do functionsemanticallyas

discourseconnectives, is derived from their predicateargumentstructureand interpretation.For

example,in (4.33),theinternalargumentof thePPdiscourseadverbial is a demonstrative AO, this

way, which refersto theinterpretationof theprior sentence.

(4.33)Thecompany interviewedeveryone.In thisway, they consideredall theiroptions.

Of course,theanaphoricityof demonstrativeNPreferenceisnotmodeledstructurallyin (D)LTAG;

asdiscussedin Chapter3, theanaphoricityof pronounscanberepresentedsemanticallyusingas-

signmentfunctions[HK98]. In general,[[ � � ]] � à is readas“x denotesanentity . via anindex + that

is mappedto . relative to anassignmentfunction j , where j is determinedby a context i ”. How-

ever, definitenounshave beenrepresentedsemanticallyusingpartial functions: thedomainof the

definitearticlecontainsonly nounsthatcorrespondto one(andonly one)entity in thesetof individ-

203

Page 221: Thesis about discourse

uals([HK98]). Althougha definitenounthuspresupposesoneandonly oneentity corresponding

to its denotation,it mayor maynot beanaphoricin thesensethat it refersto a saliententity in the

context. Demonstrative NPsaremoreakin to pronounsin that they usuallyare anaphoricin the

above-mentionedsense.However, in LTAG, demonstrative determinersadjoin to an NP andtheir

associatedfeaturestructurecontainsa +definite feature([Gro99, 159]). As statedabove, definite

descriptionsarenot discussedin [JKR03, Kal02, KJ99]; aswe will seebelow, they do discussthe

semanticrepresentationof quantifiers,modelingthemwith a“scopepart” anda“predicate-argument

part”. As definitedeterminersarenotquantifiers([Gro99, 161]),we will extrapolatetheir semantic

representation.We will alsoextend[JKR03, Kal02, KJ99]’s representationof PPsasNP-adjuncts

to PPsasS-adjuncts.

Theelementarytreesfor (4.33)areshown in Figure4.55; to build thediscoursestructure,the

elementarytreerepresentingtheemptyconnective (anchoredby 0 ) is employed.Figure4.56shows

thesemanticrepresentationsof theseelementarytrees.

D 5� � ����PP� �

P

in

NPË D 5 Ì NPÖ Ö××DET

this

NPÌ NP

way

D 5� � ����D 5 Ì 0 D 5 Ë D 5� �6�* "!#<�=���!'� D 5�B34���>4����!#<

Figure4.55:DLTAG ElementaryTreesfor Example(4.33)

� � � � "���" � ����9l M : in’(x M , sM )—————————-arg: Û x M ,(1.2)t , sM this’(pM (z))

—————arg: pM q M : way’

————-arg: –�A: � � �� "!#<�=��6!#� ��3'�#�>'�6��!'<

l g : 0 ’(s g , s� )————————-arg: sg , Û s� ,(3)t l � : interview’(c, e)

———————–arg: –

l ç : consider’(c,o)———————arg: –

Figure4.56:SemanticRepresentationof� �6� , � "���" , � ����9 , �A: , � �6�* "!#<�=���!'� and �B34���7'����!'<204

Page 222: Thesis about discourse

As shown,� �6� contains� M ; asin [Kal02], � , � , s areargumentvariableswhosevaluesareNP-

denotations.� �6� alsocontains( M , which aswe have alreadystatedis anargumentvariablewhose

valuesarediscourse-unitlabels.� "���" containsp M , which as in [Kal02] is an argumentvariable

whosevaluesareunarypredicates.� �$��9 is labeledwith theunarypredicatevalueq M . Werepresent

theinterpretationof theemptyconnective, e.g. thecontinuationof thediscourse,simply as 0 ’; the

restof its representationis identicalto thatfor and.

Figure4.57 shows the derived tree, derivation tree, and . -derivation graphfor this example,

alongwith thesemanticrepresentationthatwouldresultfrom abottomuptraversalof thederivation

tree, e.g. from adjoining� "���" to � ����9 , substitutingthe result into

� �6� , adjoining the result to�B34���7'����!'< , substitutingtheresultinto� :

, andadjoiningtheresultto � �6�* "!#<�=���!'� .

D 5� � � � � �ûûû ������D 5� �6�* "!#<�=���!'� 0 D 5� � � �����

PPú ú úûûûP

in

NPÖ Ö××DET

this

NP

way

D 5�B34���>4����!#<� � �� "!#<�=��6!#��A:

(0)�B34���7'����!'< (3)� �6� (0)� ����9 (1.2)� "���� (0)

� �6�* "!#<�=���!'��;:(0)�B34���7'����!#< (3)� �6� (0)� ����9 (1.2)� "���" (0)

l � : interview’(c, e), l ç : consider’(c,o), this’(way’(z)), l M : in’(z, l ç ), l g : 0 ’(l � , l M )———————————————————————————————arg: –

Figure4.57:DLTAG DerivedandDerivationTrees,. -DerivationGraphandSemanticsfor (4.33)

This leadsto thefinal semanticrepresentationshown in (4.34).Notethattheanaphoricityof this

wayhasnotbeenrepresentedor resolvedby thecompositionalsemantics.Anaphora/Demonstrative

NP resolutionshoulddeterminethatit refersto l � : interview’(c, e).

(4.34)interview’(c, e) J consider’(c,o) J 0 ’(interview’(c, e), in’(this’(way’(z)), consider’(c,o)))

Clearly, bottom-uptraversalof thederivationtreeyieldstheintendedsemanticinterpretation.If

insteadwe usea top-down traversalof the . -derivationgraph,we couldstill achieve this interpreta-

205

Page 223: Thesis about discourse

tion. In this case,the . -edgebetween� "���" and

� �6� is crucial; asin [Kal02], the representationof

thedeterminerthiscontainsthearguments which is identifiedwith thesubstitutionargumentatpo-

sition (1.2) in therepresentationof theprepositionvia the . -edgebetweenthem.However, asnoted

above,these. -edges,if notconstrained,allow numerousotherinterpretations.Moreover, Constraint

(1) doesnot requireus to usethe . -edgebetween�A:

and� �6� becauseonly oneof

� �6� ’s arguments

is a discourseunit variable. The interpretationproducedby consideringa top-down traversalof

primaryedgeswould bethatof two separatediscourserelations:therelationconveyedby 0 would

bebetweenthefirst andsecondsentence- the fact that in this wayadjoinsto thesecondsentence

would not bereflectedin the 0 relation. Therelationconveyed by in this waybetweenthesecond

sentenceandtheprior discoursewould bereflectedin a separate formula. Consideringthe . -edge

couldproducestill additionalinterpretations;however, Constraint(2) would preventadditionalin-

terpretationsarisingfrom this . -edge. Clearly, bottom-uptraversalprovidesa betterapproachto

deriving theintendedsemanticinterpretationfrom at leasttheseDLTAG structures.

As shown in Chapter3, however, explicit AO reference,(e.g. this way), is not theonly way an

adverbialcanfunctionasa discourseconnective. Both PPandADVP adverbialsmayalsocontain

hiddendefiniteAO argumentsthatcansometimesbemadeexplicit by aPPmodifier, e.g.asa result

(of that),consequently, for anexample(of that). Considerfor examplethediscoursein (4.35).

(4.35)Mike foundno new clients.Consequently, helosthis job.

The elementaryDLTAG treesfor (4.35) areshown in Figure 4.58. As shown, consequently

takes only a single structuralargument; the discourseunit it modifies. Again, DLTAG usesthe

emptyconnective to build thestructuralconnectionbetweenthetwo sentences.

D 5� � ����ADVP

ADV

consequently

D 5 Ì D 5� � ����D 5 Ì 0 D 5 Ë D 5�����6��� D 5�B�&��'!

Figure4.58:DLTAG ElementaryTreesfor Example(4.35)

206

Page 224: Thesis about discourse

Figure 4.59 shows the semanticrepresentationsof the elementarytreesin Figure 4.58. As

shown,� 34���7#!#?�5C!#�� "�@9 contains( M , which representsthe interpretationof adjunctionargumentvari-

able,andit alsocontains[[s � ]] � à , which representsa hiddenAO argumentthat mustbe resolved

anaphorically. Herewe experimentwith theexplicit useof anassignmentfunctionto representthis

anaphoricity;becausethe first sentenceyields the intendedvalue for + , we usethe ( variableas

generalizedvariablefor AO interpretations.� 34���>'!�?�5C!#�* "�&9 �;: �����6��� �����'!l M : consequently’([[s� ]] � à , sM )——————————-arg: sM l g : 0 ’(s g , s� )

———————arg: sg , Û s� ,(3)t l � : find’(m, c)

————–arg: –

l ç : lose’(m,j)—————-arg: –

Figure4.59:SemanticRepresentationof� 34���7#!#?�5C!#�� "�@9 , �;: , �����6��� and �����'!

Figure4.60shows thederived tree,derivation tree,and . -derivation graph,alongwith these-

manticrepresentationthatwould resultfrom a bottomup traversalof thederivation tree,e.g. from

adjoining� 3'�#�>#!#?�5C!#�* "�&9 to �B�&��#! , substitutingtheresultinto

� :, andadjoiningtheresultto �����6��� .

D 5� � � � �ûûû �����D 5�����6�7� 0 D 5� � ����

ADVP

ADV

consequently

D 5�����'!�����6�7��;:

(0)�B�&��'! (3)� 3'�#�>#!#?�5C!#�* "�&9 (0)

�����6����A:(0)�B�&��#! (3)� 34���>'!�?�5C!'�� "�@9 (0)

l � : find’(m, c), l ç : lose’(m,j), l M : consequently’([[s� ]] � à , l ç ), l g : 0 ’(l � , l M )————————————————————————————————-arg: –

Figure4.60:DLTAG DerivedandDerivationTrees,. -DerivationGraph,andSemanticsfor (4.35)

This leadsto the final semanticrepresentationshown in (4.36). Again, anaphoraresolution

shoulddeterminethat theanaphorresolvesto l � : find’(m, c). And alsoasin theprior example,the. -derivationgraphwouldproducethis,alongwith numerousother, semanticinterpretations.

(4.36)find’(m, c) J lose’(m,j) JÅ0 ’(find’(m, c), consequently’([[s� ]] ��à , lose’(m,j)))

207

Page 225: Thesis about discourse

As notedabove and in Chapter3, in PPadverbialssuchasas a result (of that), the internal

argumentcontainsahiddenor overt anaphoricAO argument.Whenthisargumentis overt,we take

thesametackasin [Kal02], wheresimilar binaryNP predicates(e.g. winnerof NP) aremodeled

usingNP elementarytreeswith a substitutedPPargument.Considerfor examplethediscoursein

(4.37).TheelementaryDLTAG treesfor (4.37)areshown in Figure4.61.

(4.37)Mike foundno new clients.Asa resultof that, helosthis job.

D 5� � ����PPÙ ÙÚÚ

P

as

NPË D 5 Ì NPÖ Ö××DET

a

NPÌ NPú ú úûûûNP

result

PP� �P

of

NPËNP

that

D 5� � ����D 5 Ì 0 D 5 Ë D 5�B��� �7� D 5�����'!

Figure4.61:DLTAG ElementaryTreesfor Example(4.37)

As shown, there is no extra elementarytree for the of preposition. Rather, it is treatedas

semanticallyvoid and is part of the elementarytree for result that selectsfor the PP containing

its internalargument.Furthermore,this treecontainsa substitutionsite for this internalargument.

Figure4.62shows thesemanticrepresentationsof theelementarytreesin Figure4.61.� �8 � � ��<�!�'5C�@ ED;��� �� "���� l M : as’(xM , sM )————————–arg: Û x M ,(1.2)t , sM a’(pM (z))

————arg: p M qM : ß y[result-of’(y, x g )]

—————————–arg: Û x g ,(2.2)t that’([[x � ]] ��à )

—————-arg: –�;: �����6��� �����#!

l g : 0 ’(s g , s� )———————arg: sg , Û s� ,(3)t l � : find’(m, c)

————–arg: –

l ç : lose’(m,j)—————-arg: –

Figure4.62:SemanticRepresentationof� �8 , � � , ��<�!�'5C�@ ED;��� , �� "���� , � : , �����6�7� and �B�&��#!

As shown,� �8 contains(>M , which asusualrepresentstheinterpretationof theadjunctionargu-

208

Page 226: Thesis about discourse

mentvariable,andit alsocontainsx M , whichrepresentsthesubstitutedNP. Wealsofollow [Kal02]’s

representationof thequantifier j , exceptthatwe condensethescopeandpredicatepartsinto a sin-

gle formula; we will discussscopeissuesbelow. Note that j suppliesthe individual variablefor

thebinarypredicateresult, andtakesanargumentvariablep M to thispredicate.Following [Kal02]’s

representationof thebinarypredicatewinner-of, result-of is representedasabinarypredicatevalue;

it takesa substitutedargumentandtheentity it denotesis treatedasa boundvariablewhichwill be

instantiatedwith theindividual variablesuppliedby j (z). We treatthepronounthat asdenotingan

entity � whosevaluemustbefixedby anassignmentfunction.

Figure4.63shows thederivedandderivationtrees,. -derivationgraphandsemanticsthatresult

from adjoining� � to ��<�!�'5C�@ ED;��� , substituting�� "�C�� into theresult24, substitutingtheresultinto

� �% ,adjoiningtheresultto �B�&��#! , substitutingtheresultinto

�;:, andfinally adjoiningtheresultto �����6��� .

D 5ð ð ð ð ð ð ð���� ñññññññD 5�����6��� 0 D 5� � � � � �������

PPÍ Í ÍÎÎÎP

as

NPÍ Í ÍÎÎÎDET

a

NPú ú úûûûNP

result

PP� �P

of

NP

that

D 5�B�&��'!�����6���� :

(0)�B�&��#! (3)� �% (0)��<�!�'5C�@ ED;��� (1.2)� � ����� � (0) �� "���� (2.2)

�����6���� :(0)�����#! (3)� �% (0)��<�!�'5C�@ ED;��� (1.2)� � ����� � (0) �� "���� (2.2)

l � : find’(m, c), l ç : lose’(m,j), l M : as’(a’(ß y[result-of’(y, that’([[x � ]] � à ))](z)), l ç ), l g : 0 (l � , l M )————————————————————————————————-arg: –

Figure4.63:DLTAG DerivedandDerivationTrees,. -DerivationGraphandSemanticsfor (4.37)

This leadsto thefinal semanticsshown in (4.38).Again,anaphoraresolutionshoulddetermine

that the anaphoricreferenceof the explicit internalargumentof result resolvesto l � : find’(m, c).

24Thetraveralbeginsat frontier nodes;left or right orderingdoesnotmatter.

209

Page 227: Thesis about discourse

Notethatif wechoseinsteadto employ a top-down traversalorderusingthe . -derivationgraph,the. -edgebetween� � and

� �% would becrucial,asthequantifierintroducestheargumentneededby

thepreposition.Also asabove, . -edgescouldproduceotherinterpretations.

(4.38)find’(m, c) J lose’(m,j) JÅ0 ’(find’(m, c), as’(a’(result-of’(z, that’[[x � ]] � à ), lose’(m,j)))

As shown in Figure4.39and(4.39), if the internalargumentof result is hidden,we canrep-

resentresultsyntacticallyusingtheatomicNP tree,andrepresentits hiddenargumentonly in the

semantics.Figure4.65 shows the derived andderivation trees, . -derivation graphandsemantics

thatresultfrom adjoining� � to ��<�!�'5C�@ , substitutingtheresultinto

� �8 , adjoiningtheresultto �����'! ,substitutingtheresultinto

�;:, andadjoiningtheresultto �����6��� . Thisleadsto thesemanticsin (4.40).

(4.39)Mike foundno new clients.As a result,helosthis job.

NP

result

qM : ß y[result’(y, [[x � ]] � à )]—————————–arg: –

Figure4.64:DLTAG ElementaryTreeandSemanticRepresentationfor ��<�!�45*�@ in (4.39)

D 5� � � � � ���� ������D 5�����6�7� 0 D 5� � � �����

PP� � ����P

as

NPÏ ÏÐÐDET

a

NP

result

D 5�B�&��'!�����6�7��;:

(0)�����'! (3)� �% (0)��<�!�'5C�& (1.2)� � (0)

l � : find’(m, c), l ç : lose’(m,j), l M : as’(a’(ß y[result’(y, [[x � ]] ��à )](z)), l ç ), l g : 0 (l � , l M )————————————————————————————————-arg: –

Figure4.65:DLTAG DerivedandDerivationTrees,. -DerivationGraphandSemanticsfor (4.39)

(4.40)find’(m, c) J lose’(m,j) JÅ0 ’(find’(m, c), as’(a’(result’(z,[[x � ]] � à )), lose’(m,j)))

210

Page 228: Thesis about discourse

Now, we have shown how bottom-uptraversalof theDLTAG derivationtreeyieldsa singlese-

manticinterpretationfor theabove examples,ascomparedto the . -derivation graph,which would

yield multiple additionalinterpretationsof eachexample. The next issueto addressconcernex-

ampleswherewe may want somesemanticambiguity, to seeif it can be achieved via flexible

compositionin theDLTAG derivationtree.

Recall that discourseadverbialsdo not composewith their left argument, and, becausethe

sentence-level parseis retainedat the discourselevel, the interpretationof the right argumentof

discourseadverbialsis alwaysthe interpretationof the S they modify. Recall further that, in the

two-sentencediscoursescontaininga structuralconnective anda discourseadverbial thatwe have

sofarconsidered,bottom-uptraversalof thederivationtreealwaysyieldsthestructuralconnective’s

secondargumentastheinterpretationof thesecondsentencemodifiedby thediscourseadverbial.

Do we ever want, for example,an interpretationwherethe structuralconnective’s secondar-

gumentis insteadthe interpretationof the unmodifiedsecondsentence?[WJSK03] identify four

separatecasesconcerningthe interactionof therelationsuppliedby a discourseadverbialwith the

relationsuppliedby a structuralconnective. Case1 representsdiscoursesin which thetwo connec-

tivesareinterpretedaseachsupplyinganindependentrelationto thediscourse.Theexamplethey

usewaspresentedin Section4.3asanexampleof how discourseadverbialsallow crossingdepen-

dencies,andis repeatedbelow in (4.16). As alreadynoted,thediscourseadverbial thenrelatesthe

orderingsituationdescribedin b. to thediscoveringsituationdescribedin d., which itself precedes

(temporally)the cancelingsituationdescribedin c. The structuralconnective becauserelatesthe

cancelingsituationdescribedin c. to thediscoveringsituationdescribedin d.

(4.16)

a. JohnlovesBarolo.

b. Soheorderedthreecasesof the’97.

c. But hehadto canceltheorder

d. becausethenhediscoveredhewasbroke.

ContrastCase1 with Case2, which representsdiscourseswherethe relationsuppliedby the

discourseadverbial is interpretedastheright semanticargumentof thestructuralconnective. The

exampleof this caseusedin [WJSK03] is shown in (4.41). In this case,the interpretationcanbe

211

Page 229: Thesis about discourse

paraphrasedasIf the light is red,stop,becauseif the light is redandyoudo somethingother than

stop,you’ll geta ticket: [WJSK03]arguethattheleft argumentof otherwiseis theinferredsituation

wherethelight is redandyoudo somethingotherthanstop25.

(4.41)

a. If thelight is red,

b. stop

c. becauseotherwiseyou’ll geta ticket.

If in Case1 theright argumentof becauseis theunmodifiedclausein (4.16d), andin Case2

the right argumentof becauseis the modifiedclausein (4.41c), thenthe DLTAG derivation tree

canonly accountfor thesedifferentinterpretationsby varyingthetraversalorder, suchthatCase2

is achievedvia bottom-uptraversal,andCase1 is achievedvia top-down traversal.Allowing both

traversalsintroducessemanticambiguity into our analysesof all the examplesaddressedin this

sectionthat containa structuralconnective anda discourseadverbial, suchthat the interpretation

producedby a top-down traversalyieldstwo separaterelationsconveyed,andabottom-uptraversal

yields the discourseadverbial’s relation embeddedin the structuralconnective’s relation. Note

however thatsuchambiguityis still lessthanthatallowedby . -edges.As illustration,thederivation

treeand . -derivationgraphfor (4.41)areshown in Figure4.66.��12!�34��5�'!� � � � ����������� (1)� � ������<�!�� (1) � ' "� ¿ (3)

� F�!' (3)� �� "��!'<��$�"'! (0)

�B12!#3'��5*#!� � � � ���������"� (1)� � ������<�!#� (1) � 4 "� ¿ (3)

� F�!# (3)� �� "��!#<��A�"#! (0)

Figure4.66:DLTAG DerivationTreeand . -DerivationGraphfor Example(4.41)

Cases3 and4 correspondin somesenseto theoppositeof Case2. Case3 representsdiscourses

wherethe relation suppliedby a discourseadverbial to the secondsentenceis parasitic on the

relation that the structuralconnective suppliesbetweenthe first and secondsentence,and Case

4 concernscaseswherethe relationsuppliedby the structuralconnective is incorporatedinto the

semanticsof thediscourseadverbialasa defeasiblerule. We focusbelow on Case3; Case4 would

behandledsimilarly. An exampleof Case3 ([WJSK03])is shown in (4.42).

25See[WJSK03,KKW01b, KKW01a] for detaileddiscussionof thelexical semanticsof otherwise.

212

Page 230: Thesis about discourse

(4.42)Johnjustbroke his arm.So,for example, hecan’t cycle to work now.

Theinterpretationof thisexampleis thatJohnnotbeingableto cycle to work is oneexampleof

theresultof Johnbreakinghis arm. In otherwords,theleft argumentof for exampleis dependent

on the relation suppliedby the structuralconnective so betweenthe first and secondsentences.

[WJSK03] arguethat theinterpretationof for exampleinvolvesfirst abstractingthemeaningof the

left argumentwith respectto thethemeaningof theunit it modifies,andthenmakinganassertion

with respectto this abstraction.In their terms,if ª representstheinterpretationof hecan’t cycleto

work now, � representstheinterpretationof Johnjust broke his arm, andresult(� , ª ) representsthe

interpretationof the relationsosuppliesto thesearguments,thenthe interpretationof for example

is asin (4.43),whereexemplifyrepresentstherelationfor examplesupplies.

(4.43)exemplify ( ª , ß X. result(X, � ))

Interestingly, [WJSK03]arguethattheinterpretationof for examplethusresemblestheinterpre-

tationof aquantifier, in thatthescopeof its interpretationis wider thanis explainedby its syntactic

position. In otherwords,our semanticanalysisshouldtake into accountthe fact that in order to

interpretfor example, we appearto alwaysabstractthe interpretationof precedingpredicate.Note

thatalthugh[WJSK03] don’t give for exampleananalysisthatfollowsfrom its internalpredicatear-

gumentstructureandsemantics(i.e.,hiddenargument),but rathertreatit asanunanalysedlexeme,

theirapproachis compatiblewith ananalysisthattakesits internalpredicateargumentstructureand

semanticsinto account.

In order to understand[WJSK03]’s treatmentof for exampleas a quantifier, we must back-

track andexamine[JKR03, Kal02, KJ99]’s accountof Englishquantifiers,which we have so far

overlooked. Above we discussedhow prior analysesof Frenchquantifiersas NP-adjunctsled

[Kal02] to proposederiving semanticinterpretationsfrom the . -derivation graph,ratherthan the

derivation tree. In English,however, quantifiershave beenanalysedasboth NP adjuncts(auxil-

iary trees)andasNPsinto which genericnounssubstitute(initial trees)26. [JKR03, Kal02, KJ99]

addressthe treatmentof Englishquantifiersasinitial trees,andanalysethemashaving a “scope”

and “predicate-argument” part, which they representusing multi-componentTAGs (MC-TAGs).

26see[Kal02] for references.

213

Page 231: Thesis about discourse

Briefly, eachquantifieris associatedwith two elementarytrees:oneauxiliary treeconsistingof a

singlenoderepresentingthescopepartof thesemanticsof thequantifier, andoneinitial treerep-

resentingthepredicateargumentpart. Thesetreesareshown respectively asthefirst two treesin

Figure4.67,alongwith their semanticrepresentations.Thescopepart (shown first), introducesa

proposition(l g ) containingthequantifier, holesfor its restrictive (hg ) andnuclear(h� ) scope,andits

variable(sM ), which is assertedto bein thenuclearscopeof thequantifier(sM � h� ). Thepredicate

argumentpartintroducesaproposition(l � ) containingapredicatevariable(pM ); thispropositionis in

therestrictive scopeof thequantifier(l � � hg ). After� !#=�!'<�9 adjoinsto therootof �B12��<��� , sM obtains

thevaluel M . After �B����F substitutesinto � !#=�!'<�9 , pM obtainsthevalueq M . After � !#=�!'<�9 substitutesinto�B12��<��� , x M obtainsthevalue � . Theonly possibledisambiguationof holesis then:hM K l g , hg K l � ,h� K l M , which leadsto thesemantics:every’(x,dog’(x), bark’(x)).

SÌ NP÷ ÷øøQ

every

NË N

dog

SÖ Ö××NPË VP

V

barks� !#=�!'<�9 � !#=�!#<�9 �����#F �B12��<���l g : every’(x, hg , h� )sM�� h�——————–arg: sM

l � : p M (x)l ��� hg—————arg: Û p M , (2) t q M : dog’

———–arg: –

l M : bark’(xM )l M�� hM——————arg: Û x M , (1) t

Figure4.67:ElementaryLTAG TreesandSemanticRepresentationsof� !#=�!#<�9 , � !#=�!#<�9 , �B����F , � 12��<���

Now, aswe saw above, when[Kal02] analysesthesemanticsof Frenchquantifers,sheargues

thatMC-TAGs(e.g. associatingquantifierswith two treesto accountfor their scope)shouldn’t be

used,becauseFrenchquantifiersareusuallytreatedasNP-adjuncts,andin orderfor theirscopepart

to adjoin to S, non tree-localMC-TAGs would be required. Becausean unrestricteduseof non-

tree-localMC-TAGshasbeenshown to bemuchmorepowerful thanTAG,sheproposesinsteadthe. -derivation graph,which is closein power to TAG. On theotherhand,in [JKR03] analternative

approachis taken,usingflexible compositionin combinationwith a restricteduseof non-tree-local

214

Page 232: Thesis about discourse

MC-TAGs (e.g. only adjunctionof the scopepart of quantifiersis allowed to be non-tree-local),

whichdoesnoteffect thegenerative capacityof thegrammar.

Thesyntacticandsemanticanalysisof NP quantifersis not at issuehere;what is at issueis the

syntacticandsemanticanalysisof discourseadverbialssuchas for example. We have illustrated

theuseof MC-TAGsfor Englishquantifersbecause[WJSK03] extendthisanalysisto for example.

As notedabove, [WJSK03]arguethatfor examplebehaveslike aquantifer, in thatits interpretation

musttakescopeoveraprecedingpredicate,whichit abstracts.They suggestthatfor examplecanbe

associatedwith theMC-TAG in Figure4.68,consistingof two auxiliary trees.They arguethat the

secondauxiliary treeshown in thefigure(thepredicateargumentpart)adjoinsto therootof theS it

modifies,while thefirst auxiliary treeshown in thefigure(thescopepart)adjoinsto theroot of the

higherdiscourseunit. For example,this treewouldadjointo therootof thesotreein thederivation

of example(4.42).

D 5 Ì D 5� � ����for example D 5 Ì

Figure4.68:ElementaryDLTAG Treesfor Examplefor example

In fact, it seemsnaturalconsiderhow [WJSK03]’s analysisof scopeeffects in for example

extendsto the interpretationof all discourseadverbialswhoseinternalargumentsaremodifiedby

quantifiersor aregeneric,includingan example, a result,everycasein asan example, asa result,

in every case, etc. For example,in (4.44), the scopeeffects of for an examplecanbe analysed

identicallyto for examplein (4.42)above.

(4.44)Johnjustbroke his arm.So,asanexample, hecan’t cycle to work now.

We leave thedetailsof this extensionfor futurework. Notebriefly however that if a top-down

traversalof thederivationtreeis used,then[WJSK03]’sanalysisof for examplewouldrequirenon-

tree-localMC-TAGs,sincethesecondauxiliary treeshown in theabovefigureadjoinsto therootof

oneelementarytree(theS it modifies),andthefirst auxiliary treeadjoinsto theroot of a different

elementarytree(the higherstructuralconnective). On the otherhand,if a bottom-uptraversalof

215

Page 233: Thesis about discourse

thederivation treeis used,thantree-localMC-TAGsresult. As illustration, thederivation treefor

[WJSK03]’s analysisof for exampleis shown asthe first treein Figure4.69,where� ����<�!���� � ¿ �! M

signifiesthepredicate-argumentpartof theMC-TAG, and� ����<�!���� � ¿ �! g signifiesthescopepartof

the MC-TAG. In a bottom-uptraversal,we canview ��329834�&! ascomposinginto� ���#<�!���� � ¿ �&! M . The

resultis aderivedtreewith ascopepart,bothof whichcomposeinto � #� .Thesecondderivation treeshown in Figure4.69representsananalysisof for examplethat in-

corporatesits internalpredicateargumentstructure.In thiscase,thescope(� !���� � ¿ �! ) andpredicate

argument( ��!���� � ¿ �&! ) partsof anMC-TAG areassociatedwith thegenericNPexample. In this case,

we cannotsimultaneouslyemploy bothtree-localMC-TAGsandbottom-uptraversal.In a bottom-

uptraversal,�B!���� � ¿ �&! and� ���#< compose;theresultingderivedtreemustcomposeinto ��329834�&! . When

theresultcomposesinto � #� , thescopepart,� !���� � ¿ �&! , adjoinsat theroot; this is anon-tree-localuse

of MC-TAGsakin to theanalysisof quantifiersin [JKR03,10], whichdoesnotaffect thegenerative

capacityof the grammar. On the otherhand,flexible compositionallows us to startat any node.

If we first compose��329834�&! into� ����< , thenthe resultcancomposeinto �B!���� � ¿ �&! . Whenthe result

composesinto � #� , thescopepart,� !���� � ¿ �&! adjoinsat theroot; this is a tree-localuseof MC-TAGs.� #�ð ð ð ð ð ð ð ð��ññññññññ�B1E<�����! (1) �B329834�&! (3)� ����<�!���� � ¿ �&! M (0)

� ����<�!���� � ¿ �! g (0)

� '�� � � � � � ������������1E<�����! (1) �B329%3'�&! (3)� ����< (0)��!���� � ¿ �&! (2)

� !���� � ¿ �! (0)

Figure4.69:DerivationTreesfor PPDiscourseAdverbialswith QuantifiedInternalArguments

We endthis sectionwith the commentthat the analysespresentedabove extendto larger dis-

courses,suchasin (4.45),but alsointroduceadditionalconsiderations.We illustratethis with two

possiblederivationtreesthatcanbeproducedfor (4.45),shown in Figure4.70,alongwith apossible

derivedtree.

(4.45)Mary founda job. ThenMike got a raise.Consequentlythey hadenoughmoney to buy

ahouse.

216

Page 234: Thesis about discourse

In the first derivation tree in Figure 4.70, we show multiple adjunctionat � F�!# . In [Kal02]

multiple adjunctionat a singlenodeis avoidedwhenpossible,ascombiningMC-TAGs with the

unrestricteduseof multiple adjunctionat a singlenodegoesbeyond the power of LTAG. It may

be that in DLTAG too, multiple adjunctionat a singlenodeshouldbe avoided,dependingon the

scopingpossibilitiesof the relationsconveyed by discourseconnectives in discoursescontaining

multiplediscourseconnectives;this is anissuefor futurestudy. In thesecondderivationtreein Fig-

ure4.70,we haveshown adjunctionof� "��!#� 27 to � F�!# andadjunctionof

�;:to� "�C!'� . Notehowever

thatoneinterpretationproducedby boththesederivationtreesis thatMikeandMary having enough

money to buy a houseis a consequenceof Mike gettinga raise.However, anotherinterpretationof

this discoursecouldbethatMike andMary having enoughmoney to buy a houseis a consequence

of both Mike gettinga raiseandMary finding a job. This interpretationhighlightsan additional

possiblederivation treefor this discourse,namelyonein which thesecond� :

, adjoinsto the root

of thefirst�;:

. Thepotentialinteractionbetweentheresolutionof theleft argumentof consequently

andtheadjunctionsiteof thesecond� :

is not yet fully understood.Althoughasdiscussedabove,

[WJSK03] presentthreecasesdescribinghow the relationssuppliedby discourseadverbialsand

structuralconnectivescaninteract,themoregeneralquestionof how theresolutionof theanaphoric

argumentsof discourseconnectivescanberestrictedby thestructuraldescriptionof thediscourse

hasnot yet beenaddressed.In fact,it maywell bethatthis interactionis bestaddressedvia consid-

erationof discoursedeixisresearch,whichhasalreadyshown thattheresolutionof discoursedeixis

interactswith the structuraldescriptionof the discourse(seeChapter2). We will addresssimilar

remainingquestionsin Chapter6.

Thediscussionsabove indicatethatwe maywantto allow someambiguityin theinterpretation

of the substitutionargumentof both the higherand the lower� :

, andwe may want to allow for

someambiguityin theinterpretationsof theleft argumentsof thediscourseadverbialsthen,conse-

quently, but possiblyonly in somecases,andpossiblyonly with respectto their interactionwith the

interpretationof prior structuralrelations.

27Thediscourseadverbial thencanberepresentedstructurallyakin to consequently

217

Page 235: Thesis about discourse

D 5� � � � � � ����� �������D 5�B���6�7� 0 D 5� � � � � �ÚÚ ������

D 5Ï ÏÐÐADVP

then

D 5� F�!' 0 D 5� � ����

ADVP

consequently

D 5�B����=�!�����6���� :(0)� F�!# (3)� � � ������ "��!'� (0)

�;:(0)���C��=�! (3)� 34���7#!�?#5*!'�� "�@9 (0)

�����6���� :(0)� F�!# (3)� "��!#� (0)�A:(0)�B����=�! (3)� 34���>'!�?�5C!'�� "�@9 (0)

Figure4.70:DLTAG DerivedandDerivationTreesfor (4.32)

4.3.4 Comparison of Approaches

Thegreatpotentialfor ambiguitythatarisesin discourseis duein partto thefactthat thesemantic

typeof all discourseunitsis thesame.Thus,while wehaveshown that[Kal02]’senrichedderivation

structurecan be usedat the discourselevel to yield compositionalsemanticson the DLTAG . -derivationgraph,we’ve alsoshown thatflexible compositionin thederivation treeis likely a more

parsimoniousapproach.However, theconsiderationof morecomplex discoursesis neededbeforea

compositionalsemanticsfor theDLTAG derivationtreeis complete.

All of the approachesto building a syntax-semanticsinterface for discoursethat have been

discussedin this sectionareexploratory, ratherthanconclusive. Nevertheless,theessentialtheory

behindeachis well-defined,enablingtheir similaritiesanddifferencesto becompared.

218

Page 236: Thesis about discourse

Apart from the definitional inconsistenciesdiscussedabove, [Gar97b]’s approachto building

a syntaxsemanticinterfaceis a viable onethat builds on feature-basedapproachesto aspectsof

compositionalsemanticsthathave alreadybeenproposedat theclauselevel. Gardent’s discourse

grammar(DTAG) is actuallyquitesimilar in many respectsto DLTAG, in thatit usestreesandthe

operationsof substitutionandadjunctionfor combiningthem.

In fact,it appearsthatall of theelementaryR treesthatGardentexemplifiescouldberepresented

asshown in Figure4.71,e.g. with oneexplicit adjunctionandoneexplicit substitutionsite,where

eachnodeis associatedwith thefeaturestructureshown in Figure4.26.

R[]Ö Ö××A[] Ì B[] Ë

Figure4.71:AnotherRepresentationof theR Treein Figure4.26

Of course,this is nota“recursive” structure,becausethefoot androotnodesarenotof thesame

category. Moreover, therearethusno “anaphoric”discourserelationsin Gardent’s grammar, which

[WJSK03] have shown to be necessaryto limit the computationalpower requiredby discourse

structureto thatof a tree-basedgrammar, asdiscussedabove.

Furthermore,[Gar97b]’s approach(DTAG) relies on the definition of discourserelationsas

featurestructures.ThemajordifferencebetweenDTAG andDLTAG, andindeedbetweenDLTAG

andall otherapproachesto discoursestructureand interpretationdiscussedin Chapter2, is that

DLTAG views cuephrasesthemselvesastheanchors of theelementary“relation” treesinvolvedin

theconstructionandinterpretationof thediscoursemodel.Theseanchorshaveapredicateargument

structureandameaningwhichconveys a relationbetweentheirarguments.

While it is likely thatsomeaspectsof discourserelationsarebestmodeledasfeatures,DLTAG

arguesthat basicpredicateargumentrelationsarealso involved. Becauseat its mostbasiclevel,

discourseconnectives arepredicatesin DLTAG, so in DLTAG we would like to make useof the

structuresthatretainpredicate-argumentinformation,whenbuilding compositionalsemantics.Al-

thoughDLTAGhasalreadyarguedthatthestructuresneededatthediscourselevel aremuchsimpler

thanthoseneededat the clause-level [WJ98], our exploratoryextensionof [Kal02]’s approachto

219

Page 237: Thesis about discourse

DLTAG alreadyappearsto provide morepossibilitiesthanareneededat thediscourselevel; aswe

saw above, for example,many . -edgesmaybevacuous.The[JKR03] approach,basedon traversal

of thederivationtree,appearsto bemoreviablefor DLTAG.

Featuresrecordedin the derived tree could also be employed in DLTAG, however, perhaps

akin to the way they areemployed at the clauselevel in FTAG, e.g. as indicatingconstraintson

propertiesof argumentsandpredicatesthatmustbepresentin orderfor their treesto combinewith

eachother. For example,in FTAG thepresenceof a Û modet feature(presentin elementaryverb

trees)is requiredin theargumentof thethink tree,indicatingthatthisargumentmustbeaclause,not

a subclausalconstituent.Moreover, while predicateargumentaspectsof compositionalsemantics

canbedefinedwith respectto thederivation tree(or . -derivation graph),it might benecessaryto

usea combinationof thederivation treeandthederived treefor anaphorresolution,if thederived

treedefinessomenotionof locality or distancefor anaphorresolution(e.g. the right frontier) that

is not asobvious with thederivation tree. A wholly feature-basedapproach,however, ignoresthe

predicateargumentstructureof connectives,andmoreover might requireanextremelylargesetof

featuresto representtheir idiosyncraticmeanings.In contrast,theassociationof lexical itemswith

treeanchorsmaintainsconsistency with the clauselevel andthe lexicon. However, [Kno96] has

alreadyshown that therelationsimpartedby a wide varietyof cuephrasescanbeviewed in terms

of a limited setof features(althoughthe idiosyncraticmeaningof thesecuesis still lost). DLTAG

semanticscould make useof thesefeatureswhenbuilding compositionalsemanticsbasedon the

derivedtree. Knott’s approachalsolendsitself well to thepossibilityof lexicalizing inferenceand

representingit, oncecomputed,bothstructurallyandsemanticallyasfeaturestructureswithin or in

additionto thefeaturestructuresof the 0 (or otherstructural)connective28.

Otherdifferencesbetweenthe two approachesarelesssignificant. For example,Gardentmo-

tivatesherapproachby thedesirabilityof theincrementalconstructionof discourseinterpretation,

but asshenotes,the ( - -)substitutionoperationalreadypermitssome“lookahead”with respectto

thebuilding of intermediatetrees;fully incrementaltreeconstructionis not possiblein LTAG, nor

is it necessarilydesirable. In addition,both approachespreserve monotonicityin the semantics.

28This suggestionwasoriginally madeby Aravind Joshiin theDLTAG meetingsat theUniversityof Pennsylvania.

220

Page 238: Thesis about discourse

The derivation treeprovides is oneway of preservingthe monotonicityof compositionalseman-

tics evenwhile allowing nonmonotonicityin thesyntacticstructure,andtheuseof top andbottom

featurestructuresis another. Thus,thetwo approachesarelargely complementary, anda complete

syntax-semanticsinterfacefor DLTAG will likely combineaspectsof bothapproaches.

4.3.5 Summary

In this sectionwe have discussedhow theDLTAG grammarbuilds discoursedirectly on top of the

LTAG clausegrammar. We thendiscusseda syntax-semanticinterfacethathasbeenproposedfor

a discoursegrammar(DTAG) similar to DLTAG, andwe alsodiscussedextensionsto DLTAG of

LTAG interfacespresentedin Section4.2. We thencomparedtheapproaches,andconcludedthat

aspectsof bothwill likely playa role in acompletesyntax-semanticinterfacefor DLTAG.

4.4 DLTAG Annotation Project

Becausein both thediscourselevel andclauselevel parse,only oneof theargumentsof discourse

adverbialscomescompositionally, the othermustbe retrieved from thediscourse.DLTAG views

thisasaproblemof anaphoraresolution.As with otheranaphora,developingalgorithmscapableof

resolvingthemin way that reflectstheir actualdistribution in discourserequiresdevelopinganan-

notatedcorpus.In thissection,wepresentanoverview of theDLTAGannotationprojectanddiscuss

two preliminarystudiesthathave alreadybeenperformedin anticipationof thelargerproject.

4.4.1 Overview of Project

Themainobjective of theDLTAG annotationprojectis to build acorpuswith discourseannotation.

While not a completerepresentationof discoursestructure,this projectaddressesa rich interme-

diate level betweenhigh level discoursestructureandclausestructurethat canbe reliably anno-

tated,namely, the syntaxandsemanticsassociatedwith discourseconnectives. We usethe Penn

Treebankannotatedcorpus,which containsnaturallyoccurringdatafrom a varietyof sources,has

alreadybeenannotatedfor clausestructureandpart-of-speech,andis currentlybeingannotatedfor

221

Page 239: Thesis about discourse

predicate-argumentstructure.Our annotationschemawill bedesignedto build anadditionallayer

of discourseannotationinto thePennTreebankcorpus,with links to thisclauselevel information.

Discourseannotationwill occurin two stages.First, the DLTAG parser[FMP�

01] is usedto

parsethediscourse,i.e. thestructuralargumentsof eachdiscourseconnective. Thisparsewill incor-

porateany clauselevel annotationthathasalreadybracketedtheS-internalargumentsof structural

connectives.Second,humanannotatorscorrectany errorsin theparse,andaddannotationtagsfor

the anaphoricargumentsof the discourseadverbials. This two-stagestrategy hasalreadyproved

successfulin the clause-level annotationof the PennTreebankcorpuswith respectto minimizing

humaneffort. Oncecomplete,this annotationcanbe used,alongwith thesyntacticandpredicate

argumentannotationalreadyin Treebank,to developanaphoraresolutionalgorithmsfor adverbial

discourseconnectives.It canalsobeusedto trainastatisticalversionof theDLTAG parserto select

themostlikely parsefrom amongthemany possiblestructuralconnections,andit canbeusedfor

furtherresearchanddevelopmentof NLP applications.

In general,preliminarydevelopmentof theDLTAGannotationprojectinvolvesfirst determining

an initial setof discourseconnectivesto beannotated.While it is hopedthateventuallyall lexical

itemsfunctioningaslinks betweenclausalunits will be annotated,initially the projectwill likely

focuson high-frequency connectives. A reliableannotationschemaandannotationguidelineswill

alsobe developed,consistingof a setof annotationtagsandproceduresfor their use. Moreover,

a “semanticframe” will be built for eachconnective, detailing the semanticpropertiesof each

connective andits arguments.In the remainderof this sectionwe discussin moredetail thekinds

of preliminary studiesaddressingtheseissuesthat will be performedfor eachconnective to be

annotatedin theDLTAG annotationproject.

4.4.2 Preliminary Study 1

Thereareat leasttwo typesof preliminarystudiesinvolved in theannotationof discourseconnec-

tivesandtheir arguments.Thefirst typeof studyhasrecentlybeenillustratedin [CFM�

02]. There

arethreegoalsfor this type of study. First, onesimply wantsto investigatethe syntacticproper-

tiesof discourseconnectivesandtheir arguments.In [CFM�

02], ninediscourseconnectiveswere

222

Page 240: Thesis about discourse

selectedfor study, shown in Table4.1. As shown, threeof the selectedconnectivesconvey a re-

sultativerelation,threeconvey anadditiverelation,andthreeconvey a concessiverelation. Three

differentannotatorseachannotatedseventy-five tokensof oneconnective from eachset.

Table4.1: NineConnectivesStudiedin [CFM�

02]

Resultative Additi ve Concessiveasa result also neverthelessso in addition whereastherefore moreover yet

The annotationschemaconsistedof four annotationtags. First, eachannotatorannotatedthe

boundariesof theleft andrightargumentof eachconnective,usingtheXML tagsÛ ARG t ...Û /ARG tto annotatetheleft argument,and Û CONNt ... Û /CONNt to annotatetheright argumentthatalso

containedthe connective. The remainingannotationtags, illustratedin Table 4.2, were usedto

capturefeaturesthatwould beautomatically-derivable from a parsedcorpus.As shown in thefirst

column,eachannotatorannotatedthesyntactictypeof theleft argumentwith aTYPEtagAs shown

in the secondcolumn,eachannotatoralsoannotatedthe presenceof otherdiscourseconnectives

andpunctuationthatco-occurwith theconnective beingannotated.As shown in thethird column,

eachannotatoralsoannotatedthepositionof thediscourseconnective in thecontainingsentence.

An examplefrom the corpusstudy is shown in (4.46). As shown, the discourseconnective

beingannotatedis asa result. Its left argumentis theprior sentence,thereareno otherdiscourse

connectiveswithin its left or right argumentexcepttheemptyconnective (signaledby theperiod),

andit occurssentence-initially.

(4.46) Û ARG TYPE=MAIN t YourJuly26editorialregardingthepositionof Attorney General

RobertF. Kennedyonprospective taxrelief for DuPontstockholdersis basedonanerroneousstate-

mentof fact Û /ARG t . Û CONNCOMB=PERIODPOS=INITIAL t As a result, yourcriticism of

Attorney GeneralRobertF. KennedyandtheDepartmentof Justicewasinaccurate, unwarranted

andunfair Û /CONNt .

It wasfoundthat featurepercentagesvariedacrossthediscourseconnectives. For example,so

alwaysoccurredsentence-initially, neverthelessoftentook a sub-clausalconstituent(XP) asits left

223

Page 241: Thesis about discourse

Table4.2: AnnotationTagsfor theNineConnectivesStudiedin [CFM�

02]

TYPE COMB POSITIONMAIN (sentence) PERIOD INITIALMULT (multiplesentences) COMMA MEDIALSUB (subordinateclause) (SEMI-)COLON FINALXP (sub-clausalconstituent) AND/BUT

argument,andtherefore oftentookasubordinateclause(SUB)asits left argument.

Eachof theannotatorsthenannotatedanadditionaltwenty-five tokensof onediscourseconnec-

tive from eachsemanticset: asa result,in addition,andnevertheless. It wasfoundthat the initial

patternsof featurespercentagesremainedstable,indicatingthat theseconnectivesdisplaypatterns

with respectto thesefeaturesthataresystematicenoughto aid in automaticargumentdetection.

Thesecondgoal in this type of studyis to testinter-annotatorreliability with respectto argu-

mentannotation.For [CFM�

02], threeadditionalannotatorseachannotatedthe left argumentof

thosetwenty-fiveadditionaltokensof asa result,in addition,andnevertheless. For thisannotation,

theTYPEtagwasreplacedwith aslightly moregeneralLOC tag,whosepossiblevaluesareshown

in Table4.3.TheLOC tagdefinesthesentence,consistingof amainclauseandany attachedsubor-

dinateor adjoinedclauses,astheminimalatomicunit from which theleft argumentis derived.The

valuesof theLOC tagdistinguishargumentsderived from thesentencecontainingtheconnective

(SS),thesingleprior adjacentsentence(PS),any sequenceof adjacentsentences(PP),or asentence

or sequenceof sentencesnot contiguousto theclausecontainingtheconnective. This tag is more

generalthantheTYPEtagin thatit doesnotaskannotatorsto distinguishsub-clausalor subordinate

clauseconstituents;ontheotherhand,it addstheinformationof whethertheargumentis contiguous

with thesentencecontainingtheconnective. All otherfeatureswereautomaticallyderivablesothe

additionalannotatorsdid notannotatethem.

Table4.4shows theinter-annotatorresults.Thefirst columnindicatesfour-way agreement,e.g.

whereall annotatorslabeledlabeledthe left argumentwith thesameLOC value. As shown, four-

224

Page 242: Thesis about discourse

Table4.3: LOC TagValues

SS (samesentence) PP (previousadjacentparagraph)PS (previoussentence) NC (non-contiguous)

wayagreementis greaterthan50%in all cases.Thesecondcolumnindicatesthree-way agreement.

Thethird columnindicatesthecasewheretwo annotatorsagreedon oneLOC value,andtheother

two annotatorsagreedon anotherLOC value;thefourth columnindicatesthatonly two annotators

agreedon a LOC value. Adding the first two columnsshows that annotationof the arguments

of theseconnectivescanbedonereliably; majority agreement(three-way or better)is 92%for in

addition, 96%for asa result. and88%for nevertheless.

Table4.4: Inter-AnnotatorAgreement

Connective 4� 4 3� 4 [2,2] � 4 2� 4in addition 76%(19) 16%(4) 4%(1) 4%(1)asa result 84%(21) 12%(3) 4%(1) 0nevertheless 52%(13) 36%(9) 12%(3) 0

The third goal in this type of studyis to seewhat thesourcesof disagreementteachus about

the annotationguidelinesthat will be neededand the sortsof resolutionalgorithmsthat can be

constructedbasedon syntacticpatterns.Oneguidelinewasdevelopedbasedin part on an initial

“exact match” comparisonbetweenthe left argumentboundariesannotatedby eachof the four

annotatorsfor thethreeconnectivesmentionedabove. It wasfoundthat theannotatorswereusing

differentunderlyingassumptionswhendecidingonthesizeandsyntacticform of theleft argument.

For example, (4.47) containsa discourseof the form cause-result-result. One annotatormight

annotateboththecauseandthefirst resultastheleft argumentof asa result, while anotherannotator

might annotateonly thefirst resultastheleft argumentof asa result.

(4.47)[Lee won thelottery. [Sohewashappy]]. As a result, hisbloodpressurewentdown.

This typeof disagreementcanbereducedby theuseof aminimal unit guideline.If annotators

225

Page 243: Thesis about discourse

areinstructedto annotatetheminimal unit which couldserve astheleft argumentof a connective,

thenmostannotatorswould annotatethe first resultonly in (4.47). The relationbetweenthe first

andsecondsentenceis not lost, for sowill take asits left argumentthefirst sentence,andtake the

secondsentenceasits right argument. Similarly, in (4.48)oneannotatormight annotatejust the

adjective overworkedastheleft argumentof asa result, while anotherannotatormightannotatethe

entireprecedingclause(Leeis overworked).

(4.48)[Johnis [overworked],andasa result, tired.

Whetherthe minimal unit guidelineappliesin this casedependson how “minimal unit” is

defined.If it is definedastheminimal clause, thentheonly optionfor theannotatoris to annotate

the entireprecedingclause. If it is definedasany phrasalconstituent,then the annotatorwould

annotatetheadjective.

(4.48)alsoillustratesa possiblesyntacticresolutionheuristicfor theleft argumentof discourse

adverbials: in all but onecasein this corpusstudy, whena coordinatingconjunctionlinked two

clauses,and the secondclausecontaineda discourseadverbial, the discourseadverbial and the

coordinatingconjunctionbothtook thesameLOC valuefor their left argument.

4.4.3 Preliminary Study 2

Thesecondtypeof preliminarystudyinvolvedin theDLTAG annotationprojectconcernstheiden-

tificationof lexico-syntacticfeaturesthatdistinguishcontextualargumentsof discourseconnectives

from otherdiscourseunitsin thecontext. Suchastudyis currentlybeingpreparedfor thediscourse

adverbial insteadfor [CFM�

03]. In this study, four annotatorseachannotatethe left argumentof

twenty-five different tokensof insteaddrawn from the PennTreebankcorpus,yielding a total of

onehundredannotatedtokens.Again, theXML tags Û ARG t ...Û /ARG t areused.

To retrieve the left (contextual) argument,eachannotatoris instructedto instantiateit as an

explicit nominalizedAO. For example,in (4.49), the bracketedargumentcould be instantiatedas

Insteadof offenders beingreleasedlocally. To ensureinter-annotatorreliability, eachannotatoris

additionallyannotatingonesetof twenty-five tokensthathasbeenannotatedby anotherannotator,

suchthateachsetis annotatedby two annotators.

226

Page 244: Thesis about discourse

(4.49)Theprisonis abig employer, andpeoplewerereassuredthat Û ARG t no offenderwould

bereleasedlocally Û /ARG t . Instead, they would be releasedfrom theprisonthey hadfirst been

sentto.

Moreover, it hasbeennotedthatthecontextual clauseargumentof insteadoftenhasanegative

subject,object,or verb,or the verb may be modal,asshown in bold-facein (4.49),or, asshown

bold-facedin example(4.50), the left (contextual) clauseargumentmay be embeddedbeneatha

non-factive verb29. What all of thesefeaturesappearto have in commonis that in one way or

anotherthey allow their clauseto be interpretedwith respectto an alternative setof propositions.

Non-factives,for example,do not presupposethetruth of their complementclause;it canbeeither

trueor false.In (4.50),becausetheeventof themtakinga monthin Europein (4.50)is notasserted

to occur, insteadcan,anddoes,assertstheoccurrenceof analternative event(e.g. thembuilding a

dreamhouse).

(4.50)Their broker encouragedthem Û ARG t to take a monthin EuropeÛ /ARG t ; +4W�(Cb�.�jq-they movedto SouthCarolina,wherethey beganbuilding adreamhouseon thebeach.

Eachannotatoris thusalsoannotatingthepresenceof thesefeaturesfor eachof their twenty-five

tokens.In (4.49)-(4.50),notethatthecontextualalternativeclausesdonothave thesefeatures.Each

annotatoris thusalsoannotatingcompetingclausesin the context for thesefeatures,to ascertain

whethertheseandadditionalfeaturesdo in factdistinguishtheleft (contextual)argumentof instead

from othercontextual clauses30.

4.4.4 Future Work

Essentially, thepreliminarystudiesdescribedabove arethesamekinds of studiesalreadydoneat

theclauselevel to determinethepredicateargumentstructuresof verbs.Futurework will include

doing thesetypesof studiesfor all connectives in thecorpus,therebyenablingfurther refinement

of theDLTAG annotationtagsandguidelines.They will alsoenable“semanticframes”,modeled

afterthoseusedin clause-level predicate-argumenttagging(see[KP02]) to beconstructedfor each

29Thesefeatureswereoriginally notedby BonnieWebberin aDLTAG meeting.30Resultsandfurtherdetailsconcerningthis studyis reportedin [CFM Ñ 03].

227

Page 245: Thesis about discourse

discourseconnective that incorporatetheir predicate-argumentstructure,meaning,any constraints

on theAO interpretationof their arguments,andany lexico-syntacticfeaturesthatdistinguishtheir

argumentfrom alternatives.Thesesemanticframeswill serveasabasicsemanticsfor eachconnec-

tive, to helptheannotatorsdeterminetherelevantsemanticrolesplayedby thecontext aroundthem

in thecorpusdiscourses.

4.5 Conclusion

In this chapter, we focusedon theconstructionof LTAG andDLTAG trees,andthecomputationof

compositionaldiscoursesemanticsfrom thesestructures.Two syntax-semanticinterfacefor LTAG

thathave beenproposedwerepresentedandcompared.We alsodiscusseda syntax-semanticinter-

facethathasbeenproposedfor DTAG, a discoursegrammarsimilar in somerespectsto DLTAG.

Drawing on theseinterfaces,wediscussedhow Chapter3’s discussionof thesemanticmechanisms

underlyingthe predicate-argumentstructureand interpretationof discourseadverbialscanbe in-

corporatedinto a syntax-semanticinterfacefor DLTAG. We also discussedthe resolutionof the

anaphoricargumentsof discourseadverbials,framing this discussionin termsof a large DLTAG

annotationprojectcurrentlyunderway.

228

Page 246: Thesis about discourse

Chapter 5

Other WaysAdverbials Contrib ute to

DiscourseCoherence

5.1 Intr oduction

Weconcludeour investigationof S-modifyingADVP/PPadverbialsby emphasizingthatwearenot

claiming that it is only dueto their argumentstructureandinterpretationthat adverbialsestablish

or contribute to discoursecoherence.For example,theargumentstructureof theADVP adverbials

actuallyandreallycannotexplainwhy they have beentreatedasdiscourseconnectives(see[KP02,

Kno96]); asdiscussedin Chapter3, theseadverbialstake only oneAO argument:theinterpretation

of themodifiedclause,whosetruth or fact is assertedto beactualor real. Similarly, theargument

resolutionof thePPadverbialsin anycaseand in fact cannotexplain why they have beentreated

asdiscourseconnectives(see[KP02]); asdiscussedin Chapter3, their internalindefiniteor generic

NP argumentsdenoteunspecified(setsof) entities,andarenot in andof themselvesreferential.

In this chapterwe explore other explanationsfor why suchadverbialscan requirediscourse

context for their interpretation;in particular, thosethat involve the interactionof their semantics

with other aspectsof discoursecoherence.In Section5.2 we introduceprosodyas a semantic

mechanismof discoursecoherenceanddiscussprior analysesof (topic) focus. In Section5.3 we

investigatewhat prior researchhascalled the focussensitivityof certainmodifiers,and discuss

229

Page 247: Thesis about discourse

how focuseffects in both clausalanddiscourse(S-modifying)adverbialscontribute to discourse

coherence.In Section5.4we introduceGriceanimplicatureasanadditionalaspectof meaningthat

arisesfrom theassumptionof discoursecoherenceanddiscusshow prior analyseshave accounted

for it and distinguishedit from the relatednotion of presupposition.In Section5.5 we suggest

how clausalanddiscourse(S-modifying)adverbialscanbeusedto convey implicatures.In Section

5.6 we point to additionalmechanismsthat mustalsobe consideredboth aloneandin relationto

“discourseconnectives”, in orderto constructacompletemodelof discourse.

5.2 Focus

The useof prosodyto convey meaningis a very active areaof currentresearch.Thereareissues

involved in theseanalysesthatarebeyondthescopeof this thesis;in this sectionwe illustratetwo

semanticanalysesof focusthatassumetree-basedgrammarsakin to DLTAG; we alsohighlight an

approachusingcategorial grammarsthatincorporatestheinsightsof boththeseanalyses.[Gar97a]

citesotherapproachesto theanalysisof focus,in particular, [Pul97].

5.2.1 The Phenomena

Theterm focus1 is usedto referto theprosodicemphasizingof partsof utterancesfor communica-

tive purposes.Focusis typically expressedin spoken languageby pitch movement,duration,or

intensityon a syllable(see[Kri , Lad66, Ste00a]). In addition,certainspecificsyntacticconstruc-

tions,suchastheEnglishcleft sentenceshown in (5.1),makeuseof focusto achievecommunicative

effects(see[Pri86]). Hereandbelow weusecapitallettersto markthefocusedphrase,unlessapar-

ticularanalysisemploys adifferentrepresentation.

(5.1) It wasBILL thatsheinvited for dinner.

Therearealsolanguagesthatmake useof specificsyntacticpositions(e.g. thepreverbalfocus

positionin Hungarian),dedicatedparticles(e.g. Quechua),or syntacticmovement(e.g. Catalan),

to achieve thecommunicative effectsof focus([Kri ]).

1Thereis anotheruseof the term focus,which refersto discoursereferentsthat are salientat the currentpoint ofdiscourseandarepotentialantecedentsfor pronouns(see[GS86]).

230

Page 248: Thesis about discourse

To illustratethediscourseeffectsof focus,considertheexamplesdiscussedby [Roo95a, Gar97a].H QuestionAnswer Congruence

Thequestionsin (5.2) canbeansweredby theanswersin (5.3) non-contrastively (a) or con-

trastively (b), but notby theanswersin (5.4):

(5.2a)Who did Mary invite for dinner?

(5.2b) Did Mary invite Bill or Johnfor dinner?

(5.3a)Mary invited BILL for dinner.

(5.3b) Mary didn’t invite BILL for dinner, but JOHN.

(5.4a)Mary invited Bill for DINNER.

(5.4b) Mary didn’t invite Bill for DINNER, but JOHN.

Theanswersin (5.3)-(5.4)areidenticalexceptfor thepositionof focus.Thepositionof focus

thuscorrelateswith theWH-phraseor disjoinedalternativesin questions.

However, (5.4 b) hasanadditionalcontrastive readingthatMary invited Bill for something

other than dinner; in this casethe contrastive focus on dinner doesnot coincidewith the

positionof theWH-phraseor disjoinedalternatives. See[Kri92] for ananalysisof multiple

focusconstructions.H Reasonsand Counterfactuals

Imaginethat JohnandMary arefriends. Johnfinds out that he will inherit a fortuneif he

marrieswithin the year. He arrangesto marry Mary becausegoing throughthe processof

finding someoneelseto marryis, in his opinion,too time-consuming.

Underthesecircumstances,thesetof sentencesin (5.5)areeasilyacceptedastrue.

(5.5a)ThereasonJohnMARRIED Mary wasto qualify for theinheritance.

(5.5b) ThereasonJohnmarriedMARY wasto avoid a time-consumingprocess.

(5.5c)If Johnhadn’t MARRIED Mary, hemightnothave gottentheinheritance.

231

Page 249: Thesis about discourse

Thesetof sentences(5.6)aremorelikely perceivedto befalse.

(5.6a)ThereasonJohnmarriedMARY wasto qualify for theinheritance.

(5.6b) ThereasonJohnMARRIED Mary wasto avoid a time-consumingprocess.

(5.6c) If Johnhadn’t marriedMARY, hemight nothave gottentheinheritance.

Again, thesentencesin (5.5)-(5.6)areidenticalexceptfor thepositionof focus.Theposition

of focuscanthusinfluencethetruth-conditionsof reasoningandcounterfactualstatements.H Conversational Implicatur e

ImaginethatMary andJohnjust received their reportcards.Their motherasksMary about

their Economicsgrades. If Mary answerswith (5.7 a), shegives the impressionthat John

barelypassed.If Mary answerswith (5.7b), shegivestheimpressionthatshedid notpass.

(5.7a)Well, JohnPASSED.

(5.7b) Well, JOHNpassed.

Again, the sentencesin (5.7) areidenticalexceptfor the positionof focus. The positionof

focuscanthuseffect theimplicatureswe draw from whatwe say.

5.2.2 Inf ormation-Structur e and Theoriesof Structured Meanings

[Cho76] observedthatfocusedconstituentspatternsyntacticallylikequantifiersandWH-constituents

with respectto crossover phenomena.For example,supposeheM refersto John. Thenwhile (5.8)

hasonly acoreferentialreading,(5.9)hasbothacoreferentialandaboundvariablereading.

(5.8)Weonly expectedthewomanheM lovesto betrayHIM M .(5.9)Weonly expectedHIM M to bebetrayedby thewomanheM loves.

Thecoreferentialreadingis distinguishedasfollows: We expectedJohnto have theproperty:ß x.x is betrayedby thewomanJohnloves.

Theboundvariablereadingis distinguishedasfollows: WeexpectedJohnto have theproperty:ß x.x is betrayedby thewomanx loves.

232

Page 250: Thesis about discourse

[Cho76] thusproposesthat focusedconstituentsbe analyzedlike WH-constituentsandquan-

tifiers. In otherwords,they mustmove andthey areassignedscopeat the interpretationlevel of

Logical Form (LF), suchthat an emptyvariablein the surfacepositionof the scopingelementis

boundby a lambdaoperatorin thesemanticinterpretation.As notedby [Roo95a], in Chomsky’s

view focushastheforceof anequality, expressedin termsof adefinitedescriptionas“thex is John”.

TheLF for (5.8) is representedasshown in (5.10).

(5.10)We( ÕUÒ only (HIM) ( ÕUÒ expectedthewomanheM lovesto betrayeM ))Chomsky’sboundvariableprinciplethenaccountsfor thecrossovereffectsin (5.9),by allowing

two LFs,shown in (5.11a)-(5.11b). Chomsky’s boundvariableprinciple is asfollows: At LF, the

phonologicalcontentof a pronounmaybeoptionallydeletedif it is c-commandedby a co-indexed

emptyvariable.

(5.11a) LF: We( Õ�Ò only (HIM) ( ÕUÒ expectedeM to bebetrayedby thewomanheM loves))

(5.11b) LF: We( Õ�Ò only (HIM) ( ÕUÒ expectedeM to bebetrayedby thewomaneM loves))

A problemfor thisanalysisis thatfocusmovementin Englishis notgovernedby generalmove-

mentconstraints.In particular, it canviolate islandconstraints,asshown (5.12),which cannotbe

violatedby WH-constituentsor quantifiermovement,asshown (5.13)-(5.14)[Gar97a, Roo85].

(5.12)They investigated[thequestionof whetheryouknow thewomanwhochairedTHE ZON-

ING BOARD]

(5.13)*(Which board)M did they investigate[thequestionof whetheryouknow thewomanwho

chairedeM ?]

(5.14) *They investigated[the questionof whetheryou know the womanwho chairedevery

boardin town] (whereeveryboard in townscopesover thewoman)

In otherresearchcircles,focusis analyzedin termsof anadditionalpartitioningof theclause,al-

thoughthetheoreticalbasisfor thisvariesaccordingto thebackgroundtheoryof theresearcher. For

example,assummarizedby [vH00], [vdG69] introducesthe pair psychological subject-predicate,

viewing psychologyis theultimatebasisfor languagestructure,while ThePragueSchool[Fir64]

usesthetermstheme-rhemeandtopic-comment, which arebothborrowedfrom traditionalrhetoric

233

Page 251: Thesis about discourse

andphilology andarere-envisionedby the AmericanStructuralists,who analyzeof focusasex-

pressingwhat is new in an utterance(see[Hal67], [SHP73]). In this view, the questionWhodid

Mary invite for dinner? canbeansweredby Mary invited BILL for dinner becausetheanswerre-

tainsthegiven informationthatMary invited someonefor dinnerandsuppliesthenew information

thatthis personis Bill. Bill is thusaccentedandotherconstituentsarede-accented.[Pri81] argues,

however, thata binarydistinctionbetweenthefamiliarity of informationinadequatelyaccountsfor

numerousdiscoursephenomenaandproposesa five-way taxonomy. [Cho71] and[Jac72] rephrase

thegiven-new distinctionin termsof presupposition-focus, stressingtheir semantic-pragmaticna-

ture. [Jac72] introducesthenotionof focusasa semanticfeature.

Theoriesof “structuredmeanings”[Kri92, vS82] combinethefocusmovementapproachwith

contemporarytheoriesof informationstructure,reformulatingthedistinctionin termsof background-

focus, andproducingasemanticaccountof focussuchthatphrasesdiffering in thelocationof focus

havedifferentsemanticvalues.They assumethatthepresenceof a focusfeaturecausesthefocused

expression(F) to bemovedoutof its originalposition,leaving a traceatLF, which is interpretedas

a variable. Thebackground (B) consistsof the remainderof theclausewith a lambdaabstraction

over thevariableleft by the focus. This backgroundcorrespondsto presupposedinformationthat

is givenor canbeaccommodated,in thecontext, andis alsorelatedto what [Pri86] calls theopen

propositionof anutterance[Ste00a]. Thefocusconstitutesthenew information.

In structuredmeaningtheories,theresultof aclausewith asinglefocusis astructuredmeaning:

a pair consistingof 1) aproperty(B) obtainedby abstractingthefocusedposition,2) thesemantics

of thefocusedphrase.Thepropertyin (5.15a), for example,is thepropertyof beingintroducedby

Johnto Sue,and � is theindividual denotedby Bill , yielding thestructuredmeaningin (5.15b).

(5.15a) JohnintroducedBILL to Sue.

(5.15b) ( ß x.[introduce(j,x,s)],b)

Thepropertyin (5.16a) is thepropertyof beinganx suchthatJohnintroducedBill to x, and (is theindividual denotedby Sue, yielding thestructuredmeaningshown in (5.16b).

(5.16a) JohnintroducedBill to SUE.

(5.16b) ( ß x.[introduce(j,b,x)],s)

234

Page 252: Thesis about discourse

5.2.3 Alter nativeSemantics

Alternativesemantics[Roo85, Roo92, Roo95a] alsoseeksto producea semanticaccountof focus

suchthat phrasesdiffering in the locationof focushave differentsemanticvalues. However this

theoryreliesonaslightly differentinterpretationof focusthanthepartitioningeffect representedin

structuredmeaningtheories,andconsequentlyproducesawholly semanticanalysisof thephenom-

ena,which is performedin situwithoutany notionof focusmovement.

Accordingto evoking alternativepropositionsis the generalfunction of focus. For example,

a questionlike Who did Mary invite for dinner? asksfor answersof the form Mary invited X

for dinner, whereX variesover persons.The answer, Mary invited BILL for dinner, identifiesa

particularanswerof thisform. Focusonanexpressionis viewedasmarkingthefactthatalternatives

to this expressionareunderconsideration2.

Thustheuseof focusin alternative semanticsis viewednot asdistinguishingtwo partsof the

clause(focusandbackground),but ratherasa semanticfeaturethat triggersthecomputationof an

additionalsemanticvaluefor theentireclause:analternativeset, of which theordinarysemantic

value is a subset.During the interpretationprocess,the focus is left in situ, andthe alternatives

that are generatedfrom the focusedexpressionto yield additionalsemanticvalue arecomputed

wholly semantically. Thenotionof alternativesetshasbeenemployedin avarietyof otherresearch,

including[Bie01, KKW01b, Ste00a].

As shown in(5.17)-(5.18),the focus featureon eachfocusedexpression� , indicatedas � � ,

triggerstheassignmentof bothanordinarysemanticvalue[[.]] � , anda focussemanticvalue[[.]] � :thesetof semanticobjectsobtainablefrom theordinarysemanticvalueby makingasubstitutionin

thepositioncorrespondingto � � .

(5.17)[[mary� likessue]]� = like(m,s)0 D , whereD is thedomainof truth values.

(5.18)[[mary� likessue]]� = like(x,s)for all x 0 D � , whereD � is thedomainof individuals.

More generally, focussemanticvalueis definedrecursively asshown in (5.19),where � rep-

resentsthemeaningof a lexical item, � is thesemantictype of � , and ± ( � M ,...,� � ) representsthe

meaningof acomplex phrasesuchasaclauseor complex verbphrase.

2[SCTÑ 94] useeye-trackingtechniquesto observe theconstructionof alternative setsduringsentenceprocessing.

235

Page 253: Thesis about discourse

(5.19)

a. [[ � � ]] � = D � ,

b. [[ � ]] � = � [[ � ]] � �c. [[ ± ( � M ,...,� � )]] � = ��± (x M ,...,x� ) ¬ x � 0 [[ � � ]] � �

As shown in (5.19a), thefocussemanticvalueof a focusedlexical item � is thesetof semantic

objectsof samesemantictypeas � . As shown in (5.19b), thefocussemanticvalueof anon-focused

lexical item is the singletonset of its ordinarysemanticvalue. As shown in (5.19 c), the focus

semanticvalueof a complex phraseis computedcompositionallyfrom the focussemanticvalues

of its componentlexical items.Thusin (5.18)above, ± ( � M ,...,� � ) would representtheproposition

Mary� likesSue, andits focussemanticvaluewould bethesetof propositionsof the form x likes

Sue, wherex rangesover individuals.

Restrictivealternative semantics[Roo92, Roo95a] hasbeenintroducedto handlecaseswhere

thealternative setthatis usedin theinterpretationof focusis asubsetof thefocussemanticvalueof

aproposition.In restrictivealternativesemantics,� is usedto representthisalternativeset,denoting

a syntacticallycovert (or Vy+#-`-/.�W ) freesemanticvariable,which focusevokesin a presuppositional

way. Whenfocus is used,the focusfeatureis interpretedby the focus interpretationoperator, ˜,

whichconstrainsC andtherebyhandlestheinterpretationof focusasdefinedin (5.20).

(5.20)Where � is asyntacticphraseandC is asyntacticallycovert semanticvariable,� ˜ C introducesthepresuppositionthatC is asubsetof [[ � ]] � containing[[ � ]] � andat

leastoneotherelement.

Like otherfreevariables,C mustfind a referent.Focusinterpretationcontributesa constraint,

but doesnot fix this referentuniquely. In eachspecificcase,C is identifiedwith somesemanticor

pragmaticobjectthatis presentfor independentreasons.Identifying thevariablewith theappropri-

ateobjectis amatterof anaphoraresolution[Roo95a].

In casesof question-answercongruence,an antecedentfor C is introducedby the semantics

and/orpragmaticsof questions:if we view the ordinarysemanticvalueof a questionasa setof

possibleanswers,i.e. a setof propositionscorrespondingto potentialanswers,bothtrueandfalse,

thentheantecedentfor C canbetheordinarysemanticvalueof thequestionitself.

236

Page 254: Thesis about discourse

For example,theordinarysemanticvalueof thequestionDoesEvawant teaor coffee?, is the

setcontainingthe propositionsEva wantstea andEva wantscoffee. If the answeris Eva wants

coffee� , thentheconstraintintroducedby ˜ is thatC bea setof propositionsof theform Evawants

y containingat leastEvawantscoffeeandsomethingelse.If theanswerinsteadhadfocuson Eva,

theconstraintwould bethatC is a setof propositionsof theform x wantscoffee. However, in this

lattercase,C wouldbeinconsistentwith theinformationindependentlycontributedby thequestion.

5.2.4 Backgroundsor Alter natives?

Conceptually, a major differencebetweenthe structuredmeaningand alternative semanticsap-

proachesis the distinctionbetweena background(or themeor presupposition)anda setof alter-

natives. The dialoguein (5.21) motivatesthe view that focus usepartitionsthe clauseinto the

background,which expressesgiven information(expressedin thequestion),andthe focus,which

givesnew information(Fred).(5.22)and(5.23)motivatetheview thatfocusyieldstheconstruction

of alternative sets.In (5.22)thereis no prior context at all; thefocusedexpressionis understoodto

expresssomecontrastto otherpossiblereferents.In (5.23),thesecontrastingreferentsareexplicit.

(5.21)Speaker A: Who did Samtalk to?

Speaker B: Samtalkedto FRED.(5.22)Samtalkedto FRED.

(5.23)Speaker A: DoesEdawantteaor coffee?

Speaker B: EdawantsCOFFEE.

Both thestructuredmeaningapproachandthealternative semanticsapproachprovide a single

analysisof the above examples. In truth, we would like both analyses,focus-backgroundandal-

ternative sets,to play a role in a completetheoryof focusinterpretation.As evidenceof this [Kri ]

presentsthedialoguein (5.24).

(5.24)Speaker A: My carbroke down.

Speaker B: Whatdid youdo?�

cananswerwith (1) I calleda mechanicor with (2) I fixedit. If focusexpressesnewness,(1)

shouldhave focuson calleda mechanic, and(2) shouldhave focusjuston fixed, asthecar (i.e. the

referentof it) is given. The lack of accenton it shows thatgivennessplaysa role in accentuation;

237

Page 255: Thesis about discourse

pronounsaregenerallynotaccentable[Kri ]. But thesameuseof focuscanalsoindicatethepresence

of alternative actions�

couldhave takenwhenheanswerswith either(1) or (2). Onewould like to

representall of this information.

As [Kri] notes,theprocessingpower requiredfor theconstructionof (context-dependent)alter-

native setsmaymake (restrictive) alternative semanticsapproacheslesseconomicalthanstructured

meaningapproaches,especiallyfor casesin whichmultiple foci areinvolved,asin (5.25).

(5.25)Mary only invited BILL for dinner. Shealsoonly invited BILL g for LUNCH M .[Roo95a] however givesa detailedcomparisonof structuredmeaningandalternative semantic

approaches,concludingthattheprosandconsof eachtheoryarenearlyequallybalanced.

For our purposes,oneaspectof this differencebetweenthesetwo theoriesis mostsignificant.

Structuredmeaningapproachesform the backgroundby treatingthe focus featureasa semantic

operator, which mustscopeover a boundvariable,forming anabstractedproperty, or background

(or presuppositionor openproposition),which is thenappliedto the focusedphrase.However, in

this semantics,relevantvariationsin therangeof entitesotherthanthefocusedentitiy to which the

propertycanapplyarenot distinguished.For example,in (5.16a), thebackgroundis theproperty

of JohnintroducingBill to someone. Therelevant rangeof peopleotherthanSuethatJohncould

have introducedBill to is notdistinguishedfrom thesetof all individuals.

The equivalent of this abstractedpropertyin alternative semanticapproachesis the focusse-

manticvalueof aclause,i.e. analternative setof propositionsproducedby substitutingalternatives

in thepositioncorrespondingto the focusedphrase.Again, the relevant alternativesthat couldbe

substitutedfor the focusedphrasearenot distinguishedfrom thesetof all individuals. Restrictive

alternative semantics,however, employs a freevariableC whosereferenceto a relevantalternative

setin thecontext is fixedvia anaphoraresolution,andis constrainedonly to beasubsetof thefocus

semanticvalueof theclausecontainingthe focusedphrase.Thus,restrictive alternative semantic

approachesallow therangeof alternativesto thefocusedphraseto bedependentonthecontext. We

will returnto this dependency anddemonstrateits effectsonS-modifiersin Section5.3.

238

Page 256: Thesis about discourse

5.2.5 ContrastiveThemes

[Ste00a] makesuseof bothpartitionsandalternative setsin hiscombinatorialgrammar-basedanal-

ysisof focus,theaimof which is to produceaviableaccountof thesyntax-phonologyinterface.In

particular, [Ste00a] retainsboththeme-rhemeandfocus-backgroundpartitions,while incorporating

thenotionof “contextually-relevant” alternative sets.

In Steedman’s analysis,the intonationpatternsof an utteranceestablisha themeanda rheme.

Essentially, the L+H* LH% tune (amongothers)is associatedwith a theme,and the H* LL%

tune(amongothers)is associatedwith a rheme3. His theme-rhemedistinctioncorrespondsroughly

to the focus-backgrounddistinctionsin structuredmeaningtheories:themesconvey presupposed

information that is given in the prior context or canbe accomodated,while rhemesconvey new

information.Steedman,however, representstheinformationpresupposedby a themein termsof an

alternative setof propositionsinstantiatedby differentpossiblerhemesdependingon the context,

whichhecallsa “rhemealternative set”.

Moreover, in his analysis,the intonationpatternsusedto establishthemesand rhemesmay

additionallyconvey that thereis givenandnew informationwithin both the themeandtherheme,

which hecallsbackgroundandfocus, respectively. In otherwords,within both themeandrheme,

focusedphrasescanbeusedto distinguishthemfrom otheralternativesin thecontext. (5.26)from

[Ste00a]) providesanexample.

(5.26)

Q: I know thatMarcellikesthemanwhowrotethemusical.But whodoesheADMIRE?

A:

Marcel ADMIRES thewomanwho DIRECTED themusical.

L+H* LH% H* LL%

background focus background focus background

�.............theme..............

� �..............................rheme................................

Boundariesof thethemeandtherhemeof theanswerin (5.26)areindicatedby arrows; within

3Steedmanemploys [PH90]’s tunenotations;anexpositionon prosodictunerepresentationanddistinctionwill takeus too far afieldandwill not beundertaken in this thesis;see[Ste00a]for a comprehensive discussionof theparticularintonationpatternsthatcanestablishthemeandrheme.

239

Page 257: Thesis about discourse

both this themeand this rhemetherearebackgroundand focusconsituentsas indicated. These

partitionsareconveyedby theprosodictunesdistinguishingthecapitalized(focused)phrases.

Thethemeof theanswerin (5.26)presupposesa rhemealternative set,a setof propositionsof

theform MarceladmiredX, whereX is instantiatedby avarietyof individualsmadeavailablein the

context, including themanwho wrote themusicalandotherpeoplerelevant to theperfomanceof

a musical.Therhemebackgroundin (5.26)includesthegiven informationthata musicalis under

discussion,while therhemefocusrestrictstherhemealternative setto oneproposition.

The themein (5.26) is intonationallymarked asalsocontaininga focusandbackground.The

themebackgroundestablishesthe theme,while the themefocuspresupposesa set of alternative

themesof the form Marcel Y’d X, whereX is instantiatedby the rhemeandY is instantiatedby

differentthemefoci in thecontext, including likesandadmires.

Steedman’s analysisenablesanaccountof a varietyof differentwaysin which a themecanbe

established.In particular, a themecanbe contrastive. For example,supposetheanswerin (5.26)

wasMarcel HATESthewomanwhoDIRECTEDthemusical, accompaniedby thesameintonation

patternasin (5.26). In this case,thesamethemealternative set is constructed,but thespeaker at

oneandthesametime recognizedthethemerequiredby thequestionandestablishesanew theme.

5.2.6 Summary

In this sectionwe have presentedthreeanalysesof focusasa semanticmechanismof discourse

coherence.In particular, we have advocatedtheview that theprosodichighlightingof a syntactic

constituentcan causethe sentencein which it is containedto be interpretedwith respectto the

discourse,by invoking analternative setwhosemembersareinterpretedwith repectto thecontext.

In thenext sections,wewill discusshow adverbialsemanticscaninteractwith focussemantics.

5.3 FocusSensitivity of Modifiers

In thissectionweaddresshow analysesof thefocussensitivityof sub-clausalmodifiersextendto one

issuethathasnotbeenwidely addressedin thefocusliterature:theeffectsof focusonS-modifying

ADVP andPPadverbials.

240

Page 258: Thesis about discourse

5.3.1 FocusParticles

Certainsyntacticexpressionsare “sensitive to focus”, in that their interpretationdependson the

placementof focus,andaresaidto “associatewith focus” [Jac72]. Onegroupof adverbswhose

sensitivity to focushasbeenwidely investigatedarecalled“focusingadverbs”or, morecommonly,

focusparticles. To illustratetheir focussensitivity, we draw on examplesfrom [Roo95a].

Considera situationin whichJohnhasbothreadthebookandsaw themovie “War andPeace”

but hasreadnothingelse.In thesecircumstances,(5.27a) below is false,but (5.27b) is true. Con-

trastthissituationwith onein whichJohnreadboth“WarandPeace”and“Crime andPunishment”,

but saw nomoviesof either. Then(5.27b) is false,but (5.27a) is true.

(5.27a) Johnonly READ WarandPeace.

(5.27b) Johnonly readWAR AND PEACE.

Sincein eachsituation,thevariantsdiffer only in thelocationof focus,focusis viewedashaving

a truth conditionaleffect in thecontext of theadverbonly [Roo95a].

With many otherfocusparticles,theeffect is saidto bepresuppositional[Roo95a]. In (5.28a),

for example,theuseof alsowith focusintroducesa presuppositionthata propositionof the form

John X’d War and Peace, whereX is not read, is true. In (5.28b), the positionof the focushas

changed,yielding a presuppositionthata propositionof theform JohnreadX, whereX is not War

andPeace, is true.

(5.28a) JohnalsoREAD WarandPeace.

(5.28b) JohnalsoreadWAR AND PEACE.

Similarly, in (5.29a), theuseof evenwith focusintroducesa presuppositionthata proposition

of the form John X’d War and Peace, whereX is not read, is true, while (5.29 b) introducesa

presuppositionthata propositionof the form JohnreadX, whereX is not War andPeace, is true.

The useof evenalsoconveys an additionalpresuppositionalongthe lines that thereis something

unexpectedaboutJohn’s readingWar and Peace,wherewhat is unexpectedcorrespondsto the

focusedphrase.

(5.29a) JohnevenREAD WarandPeace.

241

Page 259: Thesis about discourse

(5.29b) JohnevenreadWAR AND PEACE.

Generally, the differencebetweena presuppositionaland a truth-conditionaleffect of focus

is shown empirically by constructingtwo otherwiseidentical sentenceswhich differ only in the

locationof focus.If a situation(i.e. a prior context) canbeconstructedsuchthatthetwo sentences

have differenttruth valuesgiventhis situation,thenfocushasa truth-conditionaleffect. If no such

situationcanbeconstructed,but insteadthecontext thetwo sentencesseemto requirediffers,then

focushasapresuppositionaleffect ([Roo95a]).

All of the focusparticlesmentionedabove arefurtherclaimedto presupposethe truth of their

containingclause.Whetheror not all of these“presuppositions” aresemanticpresuppositionsis a

matterof somedebate.As notedby [Bie01], for example,[KP79] arguethatthepresuppositionsof

focusparticlesaredueto conventionalimplicature[Gri75] ratherthansemanticpresupposition.We

will discussthedifferencebetweenpresuppositionandimplicaturein Section5.4.

In bothstructuredmeaningandalternative semanticsapproaches,however, thetruthof thecon-

tainingclauseis representedasasemanticpresupposition.In StructuredMeaningapproaches,focus

particlesareviewedasoperatorswhich take thefocusandthebackground,i.e.,a structuredmean-

ing, astheirargument.Themeaningof only, for example,is definedby [Hor69] asin (5.30):

(5.30)only combiningwith thestructuredmeaning(R,� M ...�B� ) yields:

theassertion:e x M ...e x � [R(x M ...x� ) K (x M ...x� ) = ( � M ...��� )]thepresupposition:R(� M ...�B� ) is true.

In (5.31 a), for example,thereis only one focusedphrase:BILL, andso the backgroundof

its structuredmeaningin (5.31 b) is a one-placerelation. Thus the semanticvalueof the clause

shown in (5.31c)- (5.31d) is anassertionthatJohnintroducednobodyotherthanBill to Sue,anda

presuppositionthatJohnintroducedBill to Sue.

(5.31a) Johnonly introducedBILL to Sue.

(5.31b) structuredmeaning:( ß x.[introduce(j,x,s)],b)

(5.31c) assertion:e x[introduce(j,x,s)K (j,x,s)=(j,b,s)]

(5.31d) presupposition:[introduce(j,b,s)presupposed]

242

Page 260: Thesis about discourse

In (5.32a), therearetwo focusedphrases,andsoR hasa correspondingnumberof arguments.

(5.32a) Johnonly introducedBILL to SUE.

(5.32b) structuredmeaning:( ß x ß y.[introduce(j,x,y)], b,s)

(5.32c) assertion:e xe y[introduce(j,x,y) K (j,x,y)=(j,b,s)]

(5.32d) presupposition:[introduce(j,b,s)]

[Kri92] furtherproposessemanticvalueswithin theStructuredMeaningframework for avariety

of complex focusstructures,includingmultiple focusparticleswith nestedor independentfoci.

In Alternative Semantics,focusparticlesareassigneda lexical semanticvaluewhichquantifies

over propositions.For example,thefocusparticleonly is definedby [Roo85] asin (5.33),wherep

is auniversallyquantifiedpropositionvariable4.

(5.33)only combiningwith aclause± yields:

theassertion:e p[p0 [[ ± ]] � & = p � p = [[ ± ]] � ]In Montague’s intensionallogic,= p is understoodasmeaningthatp is true; = evaluatesa propo-

sition at thecurrentindex, andthereforecombinesthepresuppositionandassertion.

This rule is differentfrom its StructuredMeaningcounterpartin thatthequantificationis at the

level of propositions:no alternative to [[ ± ]] � is bothdistinctfrom [[ ± ]] � andtrue.

As notedin Section5.2, restrictivealternative semanticshasbeenintroducedto handlecases

wherethealternative setthat is usedin theinterpretationof focusis a subsetof thefocussemantic

valueof a proposition.For example,in thesentencebelow, thealternative setconsistsof just three

propositions,ratherthanthefull setof propositionsof the form JohnintroducedX to Sue(i.e. the

full focussemanticvalue).

JohnbroughtTom,Bill, andHarry to theparty, but heonly introducedBill � to Sue.

In restrictive alternative semantics,focus particlesarestill assigneda lexical semanticvalue

which quantifiespropositions;however, their domainof quantificationis no longer the focusse-

manticvalueof the proposition,but insteadis the covert free variable,C introducedby the focus

4The differencebetweenë �"!$# , whosefocussensitivity yields a truth-conditionaleffect, and %&!('�ë and )+*�)+� , whosesensitivity yieldsa presuppostionaleffect,canberepresentedsemanticallyusing , and - .

243

Page 261: Thesis about discourse

interpretationoperator, whosereferentset of alternative phrasesis fixed by context, constrained

only to beasubsetof thefocussemanticvalueof theproposition.

Thefocusparticleonly is now defined[Roo85] asin (5.34),wherep is a universallyquantified

propositionvariable.

(5.34)only combiningwith aclause± yields:

theassertion:e p[p0 C & = p � p = [[ ± ]] � ]5.3.2 Other FocusSensitiveSub-ClausalModifiers

Although also, only, evenare the mostcommonlystudiedfocusparticles,the literatureyields a

varietyof additionalphrasescategorizedasfocusparticles(see[web, Kon91, QGLS85]). [Kon91]

basestheseinclusionson thefact thatall phrasescategorizedasfocusparticlessharecertainprop-

erties.Syntactically, for example,focusparticlescanoccurin avarietyof positionsin thesentence.

Someexamplesof this positionalvariationareshown in (5.35). As shown, focusparticlesoften

immediatelyprecede(or follow) the focusedphrase,and their felicituous usein other positions

appearsto be dependenton the positionof the focusedphrase.However, the exact natureof this

dependency is still anopenquestion;it varieswith respectto eachlexical itemandis to someextent

dialect-specific.For example,many readerswill find thelast threeexamples(j-l) awkwardor even

infelicituousif evenis replacedwith only, andsomereadersmaypreferevenin theseexamplesto

bereplacedby stressedalso.

(5.35a) EvenJOHNshowedthepaintingto Mary.

(5.35b) JOHNevenshowedthepaintingto Mary.

(5.35c) JohnevenSHOWED thepaintingto Mary.

(5.35d) JohnevenshowedTHE PAINTING to Mary.

(5.35e) Johnevenshowedthepaintingto MARY.

(5.35f) JohnevenSHOWED THE PAINTING to Mary.

(5.35g) JohnshowedevenTHE PAINTING to Mary.

(5.35h) Johnshowedthepaintingevento MARY.

(5.35i) Johnshowedthepaintingto MARY, even.

244

Page 262: Thesis about discourse

(5.35j) JohnshowedTHE PAINTING to Mary, even.

(5.35k) JohnSHOWED thepaintingto Mary, even.

(5.35l) JOHNshowedthepaintingto Mary, even.

Semantically, focusparticlessharethe propertyof beingeitheradditiveor restrictive(or the

relatedtermsinclusiveor exclusive) with respectto thewaythey associatethepropositioncontaining

the focusedphraseto the other propositionsin the alternative set (or alternatively, the way they

associatethe focusedphrasewith the background).For example,only is restrictive, becauseas

notedin theprior sectionit assertsthat thepropositioncontainingthe focusphraseis theonetrue

propositionin thealternative set. In contrast,evenandalsoareadditive, presupposingat leastone

othertruepropositionin thealternative set.

In addition,focusparticlesmay order(or scale)the presupposedalternatives,andwithin this

ordering,evaluatethepositionof thealternative contianingthefocusedphrase.For example,even

conveys that thealternative propositioncontainingthe focusedphraseis lesslikely thantheother

alternative propositionswithin thealternative set.

On thebasisof suchproperties,[web, Kon91, QGLS85]includea varietyof additionalADVP

andPPin their listsof focusparticles.While many of theselexical itemshavemultiplemeanings,at

leastonecanbereadasadditiveor restrictive with respectto its interactionwith analternative setin

a focusconstruction5. For example,thelexical itemsin (5.36),thoughmorerestrictedpositionally,

arevery similar in meaningto also, andareoftencategorizedasfocusparticles.

(5.36)aswell, in addition,too

It is not surprisingthat theselexical itemsdisplaythesamepresuppositionalsensitivity to fo-

cusasdoesalso, asshown by theexamplesin (5.37). (5.37a) introducesa presuppositionthat a

propositionof theform JohnX’d War andPeace, whereX is not read, is true. (5.37b) introducesa

presuppositionthatapropositionof theform JohnreadX, whereX is notWar andPeace, is true.

(5.37a) JohnREAD WarandPeace,in addition/aswell/too.

(5.37b) JohnreadWAR AND PEACE, in addition/aswell/too.

5See[Kon91] for a discussionof theidiosyncraticmeaningandpositionalvariationsof otherfocusparticles.

245

Page 263: Thesis about discourse

[Kon91] alsocategorizesthe lexical itemsshown in (5.38)asadditive focusparticles,though

they tooaremorerestrictedpositionally.

(5.38)likewise, similarly, somuch as

Konigarguesthatthefirst two itemsin (5.38)donotinduceanorderingof thesetof alternatives,

asexemplifiedin (5.39),andconvey a meaningakin to, albeit slightly richer than,also. The last

item in (5.38) he treatsakin to even, in that it inducesan orderingof the setof alternatives and

evaluatesthealternative containingthefocuswith respectto thisordering,asexemplifiedin (5.40).

(5.39)JOHN likewise/similarlysaw themovie.

(5.40)I doubtJohnwill somuch asGREETMary.

[QGLS85] distinguishestwo typesof restrictive focus particles. The first type of restrictive

focusparticlesaresimilar in meaningto only, asshown in (5.41).

(5.41)but, exclusively, just,merely, purely, simply, solely

It is thusnot surprisingthattheselexical itemsdisplaythesametruth-conditionalsensitivity to

focusthatonly does,asexemplifiedin (5.42). If Johnboth readthebookandsaw themovie War

andPeacebut hasreadnothingelse,(5.42a) is false,but (5.42b) is true. If JohnreadbothWar and

PeaceandCrimeandPunishment, but saw no moviesof either, thesetruthvaluesarereversed.

(5.42a) Johnbut/just/solelyREAD WarandPeace.

(5.42b) Johnbut/just/solelyreadWAR AND PEACE.

Thesecondset[QGLS85]categorizesasrestrictive focusparticlesareshown in (5.43).

(5.43)at least,chiefly, especially, in particular, largely, mainly, mostly, notably, particularly,

primarily, principally, specifically

[QGLS85, 604] view theselexical items as restrictivebecausethey “restrict the application

of the utterancepredominantlyto the part focused”. In contrast,however, [Kon91] arguesthat

this predominancedoesnot make theselexical items (which he calls particularisers) restrictive,

but ratherindicatesthat they inducean orderingon the alternatives andevaluatethe proposition

containingthefocusedphrasewith respectto thatordering,akin to theadditive focusparticle,even.

246

Page 264: Thesis about discourse

[Kon91, 96-97] thus arguesthat theselexical items are likely additive particles,and statesthat

they “clearly” presupposethetruth of otherpropositionsin thealternative set.For example,(5.44)

appearsto imply thatotherpeoplebesidesyoungpeoplearealsosusceptibleto peerpressure.

(5.44)YOUNG PEOPLEin particular aresusceptibleto peerpressure.

However, wefind thatthis “additive quality” is notclearlyapparentin all of thelexical itemsin

(5.43).For example,at least, exemplifiedin (5.45),doesnotclearlypresupposethetruthof another

propositionin thealternative set(thatJohnlikesX whereX is not Mary), nor doesit assertthatno

otherpropositionin thesetis true,althoughit doesinduceanorderingof theelementsin thesetand

evaluatethepropositioncontainingthefocusedphrasewith respectto thatordering.

(5.45)Johnat leastlikesMARY.

In Chapter3 (Table72) we saw that many of theseadverbialscanalsomodify S. In their S-

internaluse,they have beencalled“focusparticles”becausetheinterpretationof thealternative set

they presupposeappearsto dependuponthe presenceof focus. Semantically, the alternative set

theseparticlespresupposeis equatedwith thealternative setintroducedby theinterpretationof fo-

cus,whichcontainsthepropositioncontainingthefocusedphraseaswell asalternativepropositions

in which thefocusedphraseis replacedby otherentitiesof like semantictype.

Indeed,S-internalusesof focus particlesin constructionswithout focus have not even been

consideredin much of the literature,which implies that they are infelicituous or at leasthighly

restricted.If suchcasesdoexist, it is notclearhow thealternativesetthey presupposeis interpreted,

becausethereis no alternative setintroducedby focusinterpretationfor it to beequatedwith. Of

course,if the focusparticleis itself focused,it canrefer to a differentsetthanthesetinvoked by

focuson thephraseit modifies,asshown in (5.46).

(5.46i) Speaker A: DoesEvawantteaor coffeeat lunch?

(5.46ii) Speaker B: Sheonly wantsCOFFEEat lunch.

(5.46iii) Speaker B: SheONLY wantsCOFFEEat lunch.

If B answerswith (ii), heis likely answeringthequestionposedby Speaker A. As notedabove,

the questionprovides an antecedentfor the alternative set that the interpretationof the focused

247

Page 265: Thesis about discourse

elementCOFFEEintroduces,namely, asubsetof thefocussemanticvalueof theclausecontaining

thefocus.This setis equatedwith thealternative setthatonly presupposes.If B answerswith (iii),

however, hemaybothanswerthequestionandsupplyadditional information.For example,B may

indicateby answeringwith (iii) boththatEvawantscoffeeandthatcoffeeis all shewantsfor lunch.

The latter alternative set,introducedby the focusedon only, appearsto includepropositionssuch

asEvaalsowantscoffeeat lunch, which itself presupposes(via also) thatEva is having something

elsefor lunch.

Interestingly, however, not all lexical itemsthatpresupposeanalternative setrequirethepres-

enceof focus to interpretthat alternative set. For example,the alternative phrasesdiscussedin

Chapter3, includingother, such aredefinedin [Bie01, WJSK03] in termsof alternative sets.The

form otherX, for example,refersto the resultof excludinganentity or setof entitiesfrom a con-

textually relevantalternative set.Thus,otherdogs in (5.47)refersto thesetof dogsresultingfrom

excluding oneor moredogsin the discoursecontext from a larger presupposedsetof alternative

dogs.In contrast,such dogsin (5.48)refersto thesetof dogsresultingfrom usingoneor moredogs

in thediscoursecontext asanexampleof apresupposedsetof dogs.

(5.47)Johnlikesotherdogs.

(5.48)Johnlikessuchdogs.

In otherwords,other is a restrictivephrase,akin to only6, andsuch is anadditiveor inclusive

phrase,akinto also, with respectto thewaythey associatethepropositioncontainingthemodifierto

otherpropositionsin thealternative set.However, thenotionof focusis not invoked; thereneednot

bea focusedphrasein theseconstructionsto determinethesetof alternativesunderconsideration.

Onewonderswhy thesephrasesarenot focusparticles,suchthat their interpretationis depen-

denton or at leastsensitive to thepresenceof focus. Focusandfocusparticlescanbepresent,as

exemplifiedby thediscoursein (5.49).

(5.49)BothJohnandMary hateSally’scatsanddogs.But while Mary likesothercatsanddogs,

Johnonly likesotherDOGS.

6As BonnieWebberpointsout,however, other is not exclusive; it canbeadditive, asshown by thefelicity of theuseof aswell in thefollowing discourse:“Mik e only likespoodles,but Johnlikesotherdogsaswell.”

248

Page 266: Thesis about discourse

In this example,theuseof focuson dogs invokesa setof alternative propositionsof the form

John likesotherX. Theprior context suppliesoneparticularalternative instantiationof X, namely

thecatswhichMary likes.In contrast,theuseof other invokesanalternativesetwhosepropositions

assertthe existenceof alternative dogs7; againthe context suppliessomeparticularalternatives,

namelySally’s dogs.

Onereason,therefore,thatBierner’s alternative phrasesarenot focus-sensitive in thesameway

asthefocusparticlesis thatthesetof alternativesinvokedby theuseof analternative phrasedonot

containany of thesameelementsthatarecontainedin thesetof alternativesthat is invokedby the

useof focusor a focusparticle.

Of course,asshown in (5.50),we canalsofocusother itself.

(5.50)Mike likespoodles.Johnonly likesOTHERdogs.

In this case,the alternative set associatedwith the semanticsof other containspropositions

assertingtheexistenceof contextually relevantdogs(e.g. thepoodlesthatMike likes),andthis set

alsocontainsthe“otherdogs”themselves,e.g.thesubsetof dogsthatresultfrom excludingpoodles

from thesetof all (salient)dogs.In contrast,thealternative setassociatedwith thefocusonother is

thesetof propositionsof theform JohnlikesX dogs, wherevaluesfor X mayincludedeterminers

andquantiferssuchasthis,all, a, some, etc, aswell ascomparativessuchasbigger, smaller, faster,

etc. Thesemanticsof only requiresthattheseotherpropositionsarefalse.

5.3.3 S-Modifying “F ocusParticles”

As notedabove,thelabel“focusparticle”hasbeenappliedto only S-internalusesof focussensitive

modifiers;their sensitivity to focusin their useasS-modifiershasnot beenaddressed.Moreover,

bothstructuredmeaningandalternative semanticsapproacheshave definedthesemanticsof focus

particleswholly in relationto a focusedphrase.However, by definingthesemanticsof focuspar-

ticles in relation to the free variableC (introducedby the focus interpretationoperator)that can

resolve to a context-dependentalternative set,restrictive alternative semanticscanmoreeasilybe

7In Bierner’s semantics,thealternativesarerepresentedaspropositionswhich asserttheexistenceof anentity.

249

Page 267: Thesis about discourse

extendedto accountfor S-modifying,discourseadverbialusesof theseadverbs8.

Regardlessof the syntacticconstituentto which they adjoin, all of the adverbialsdiscussed

abovepresupposeanalternative setof propositions.Wethustake thepositionadvocatedby [Ern84]

anddiscussedin Chapter3, that ideally thesamesemanticsshouldberetainedfor a givenmodifier

whenever its meaningdoesnot changedespiteits positionalvariability. In other words, instead

of postulating“homonyms”, thesyntacticconsitituentcorrespondingto theexternalargumentin a

givensemanticsshouldsimply beallowedto vary. Notehowever thatmuchlike we saw in Chapter

3 in clause-level adverbialresearch,it is oftenthecasethatfocusparticleresearchers(see[Kon91])

call S-modifyingusesof focusparticles“conjunctive” or “discourseconnective” andexcludethem

from their analysis.

Our positionis neverthelessapplicableto mostof thefocusparticlesdiscussedabove, because,

again,regardlessof the syntacticconstituentthey modify, they presupposean alternative set of

propositions.Somerequirethatat leasttwo membersof thisalternative setbetrue(e.g.even,also),

while othersrequirethat only onememberof this setbe true (e.g. only, solely). Moreover, they

candisplay the samefocussensitivity asS-modifiersthat they do as(S-internal)focusparticles;

however, asS-modifiers,their interpretationis not dependenton thepresenceof a focusedphrase.

Considerexamples(5.51a)-(5.52a),whichcontainno singlefocusedphrase.

(5.51a) TheMetswon theworld series.It evenrainedin Lima.

(5.52a) TheMetswon theworld series.It alsorainedin Lima.

[Roo95b, p.17] cites example(5.51 a) as a “direction for investigation”,suggestingthat its

interpretationindicatesthatit rainingin Lima is moreimprobablethantheMetswinning theworld

series.In Rooth’sview, thisexampleshowsthatthealternativepresupposedbyevenmustbeallowed

to resolve to theinterpretationof theclauseTheMetswontheworld seriesevenin theabsenceof a

focusedphrase.In otherwords,bothpropositionsarecontainedandorderedin somepresupposed

alternative set.

Althoughthefocusparticlesin examples(5.51a)-(5.52a)aresyntacticallyS-internal,semanti-

8This is not to saythat this approachis sufficient, or thatstructuredmeaningapproachescouldnot alsobeextendedto take theseobservationsinto account.

250

Page 268: Thesis about discourse

cally thesubjectit in theseexamplesis anexpletive,usedin Englishbecauseasubjectmustalways

be presentin declaratives9. Semantically, moreoever, thesefocusparticlesarefunctioningasdis-

courseadverbials,relatingtheAO interpretationof theclausethey modify to theAO interpretation

of a prior clause,by presupposinganalternative setthatcontainsthemboth. In (5.52a), for exam-

ple, thesetpresupposedby alsocanbeinterpretedascontainingtheeventof theMetswinning the

world seriesandthe event of it raining in Lima; the resultinginterpretationof the discoursemay

simply bethata setof eventsoccurred.This sameinterpretationis achieved if also is S-initial and

S-modifying,asshown in (5.52b).

(5.52b) TheMetswon theworld series.Also, it rainedin Lima.

Moreover, as noted in Chapter3, there are 48 naturally occurring instancesof S-initial S-

modifying also foundin our corpus;oneexampleis shown in (5.52c). Again, thereis no apparent

focusedphrase.The setpresupposedby also containsthe performancesof a variety of different

stocks,foundastheabstractobjectinterpretationsof thefour clausesconstitutingthefirst sentence.

Thetwo instancesof S-initial S-modifyingevenin ourcorpusareactuallymis-parsedNP-modifiers,

however, andasearchof theraw datain ourcorpusproducedno examplesof S-finaleven.

(5.52c) Pfizergained1 7/8 to 67 5/8, Schering-Ploughadded2 1/4 to 75 3/4, Eli Lilly rose1

3/8 to 62 1/8andUpjohnfirmed3/4 to 38. Also, SmithKlineBeechamrose1 3/8 to 39 1/2. (WSJ)

Similarly, in (5.51 a), the set presupposedby even can be interpretedas containingordered

improbableevents.AlthoughS-initial S-modifyingusesof evenappearto beinfelicituous,we can

achieve thesameinterpretationin (5.51b) usingS-finalS-modifyingusesof even.

(5.51b) TheMetswon theworld series.It rainedin Lima, even.

If we replaceeven in (5.51 a) with only, as in (5.53 a), we find that we cannotinterpretthe

setpresupposedby only ascontainingboththeeventof theMetswinning theworld seriesandthe

event of it raining in Lima, becausethe semanticsof only assertsthat only onepropositionin its

presupposedalternative setis true.

(5.53a) TheMetswon theworld series.It only rainedin Lima.

9Althoughin speechthis “rule” doesnotalwayshold.

251

Page 269: Thesis about discourse

The alternative setpresupposedby only in this casecontainsa singleabstractobject. Unless

thespeaker is implicitly denying thetruth of thefirst sentence,its AO interpretationcannotthusbe

includedin thisalternative set.Wemight thereforesimply interpretthisdiscourseasasetof uncon-

nectedevents. Only on suchan interpretationwould simply presupposethat in no otherlocations

did it rain.

On theotherhand,it appearsthat,onanotherinterpretation,only in (5.53a)canmakeuseof an

inferreddiscourserelationbetweenthe two sentenceswhenselectingits contextually relevant set

of alternatives. For example,we might infer a causalrelationbetweenthe two clauses(which we

canalsomake explicit: TheMetswontheworld seriesbecauseit only rainedin Lima.). Thenthe

interpretationof it only rainedin Lima producesthealternative setcontainingall contextually rele-

vant locationswhereit rained,namelyNew York, andassertsthat it did not rain in theselocations,

yielding theinterpretationthattheMetswon theworld seriesbecauseit did not rain in New York.

Notice however thatalthoughonly canappearfelicituously asan S-initial S-modifier, we find

thatadifferent(anddifficult to understand)interpretationresultsin (5.53b).

(5.53b) TheMetswon theworld series.Only, it rainedin Lima.

This peculiarity of S-initial, S-modifying, discourseadverbial usesof only hasnot beenre-

markedonin theliteraturethatwehave investigated.A numberof suchinstancesthataren’t difficult

to understandareneverthelessfoundin ourcorpus,exemplifiedin (5.54).

(5.54a) His realname’s DiMaggio,only we call him Maggiebecausehehasto take tranquiliz-

ers.(Brown)

(5.54c) Theremight have beena pool of cool waterbehindany of thesetree-clumps:only –

therewasnot. It might have rained,any time; only– it did not. (Brown)

Only in all of thesecasesis paraphrasableby exceptthat. Thealternative phraseexceptfor X is

discussedin [Bie01]; exceptis treatedasexcludingX from a setmadeavailablein thecontaining

sentence(e.g. animalsin (5.55a).Thisanalysisextendsnaturallyto thesubordinatingconjunction

exceptthat shown in (5.55b) andto its adverbialcounterpartexceptshown in (5.55c), althoughas

discussedbelow, thequestionof how AOs(and/orrelationsbetweenthem)make availablealterna-

252

Page 270: Thesis about discourse

tivesetsof AOsis still anopenquestion.

(5.55a) Exceptfor dogs,Mary hatesanimals.

(5.55b) Overall,Mary is ananimal-hater, exceptthatshelovesdogs.

(5.55c) Overall,Mary is ananimal-hater. Exceptshelovesdogs.

It maybepossibleto usethesemanticsof thefocusparticleonly to accountfor its homonymous

usesin (5.54)10. The clearestcasethat supportsthis analysisis (5.54 c), wherethe modalmight

makesavailablea setof alternatives, e.g. it rained, it didn’t rain, andonly assertsthat only one

of thesealternatives is true. As notedin [WJSK03, KKW01b], however, all the ways in which

an alternative setof AOs canbe derived from a sentenceor discourseunit arenot yet known; we

are just beginning to considerAOs as units of interpretationat all. Moreover, it might turn out

that theseusesof only are bestrepresentedas signalling a causalrelation akin to but, with the

additionalfeaturethat the interpretationof themodifiedpropositionis thesinglefactblocking the

normalconsequenceof theinterpretationof thefirst sentence11. To illustratethisanalysis,consider

a modalvariantof (5.53 b), shown below in (5.53c). The interpretationof this discourseis that

the single (relevant) fact blocking the Mets winning the world seriesis the fact that it rainedin

Lima. A completeunderstandingtheidiosyncraticlexical semanticsandresolutionpropertiesof all

discourseadverbials,includingonly, requiresanannotationstudysuchasdescribedin Chapter4.

(5.53c) TheMetsmight/couldhave theworld series.Only, it rainedin Lima.

So far we have notedonly that focus neednot be presenton a particularphrasein clauses

containingdiscourseadverbial usesof “focus particles”,in orderfor their presupposedalternative

setto beinterpreted.Importantly, however, Roothnotesthat thereis a questionof whetherthereis

no focusor whethertheentire clauseis focused(in (5.51a)). Certainly, if oneadvocatesthetheory

of informationstructure,thenthereis aninformationstructureto every sentence,althoughasnoted

in Section5.2ourunderstandingof whichprosodictunesindicatethemeandrhemeis notcomplete,

andmoreover, in many casesinformationstructureis unmarkedprosodically[Ste00a]. For example,

10The analysiswould procedesimilarly to DLTAG’s analysisof otherwise ([WJSK03, KKW01b]) and instead([CFM Ñ 03]).

11This possibilitywassuggestedby BonnieWebber, personalcommunication.

253

Page 271: Thesis about discourse

asdiscussedin [Ste00a], asentencemaybeall theme,if it repeatsaprior sentence,or all rheme,as

in theanswerto thequestionin (5.56)(from [Ste00a]). In sucha case,therhemealternative setis

thesetof all (contextually relevant)propositions.

(5.56)Guesswhat?Marcell provedCOMPLETENESS!

Whetheror not they aremarkedprosodically, theclausesmodifiedby discourseadverbialuses

of focusparticlesexemplifiedabove couldbeanalysedas“all-rheme”. This in fact is theimpactof

whatRoothconsiders[Roo95a] whenhesaysthattheentireclausemight befocused.

5.3.4 FocusSensivity of S-Modifying Adverbials

However, S-modifyingadverbialscanalsobe sensitive to the presenceof focuson a sub-clausal

phrase. For example,“adverbsof quantification”have beendiscussedin [Roo85] as displaying

sensitivity to focus,asexemplified in (5.57). As Roothnotes,a bankclerk escortinga ballerina

wouldmake (5.57a) false,but not (5.57b). An officerescortingabankclerkwould runmake (5.57

b) false,but not (5.57a).

(5.57a) OFFICERSalwaysescortedballerinas.

(5.57b) OfficersalwaysescortedBALLERINAS.

Althoughthespecificsemanticsfor eachadverbdependsontheadverb(e.g.always,sometimes,

usually, etc.),[Roo85] andsubsequentRoothpapersgenerallyarguethefollowing:H Adverbsof Quantificationdenotea relationbetweensetsof events,andcombinecomposi-

tionally with anS.H TheSwith whichanAdverbof Quantificationcombinesdenotesa temporalabstract,asetof

eventswhichfills thesecondargument(scope)of theadverb.H Thefirst argument(restriction)is a freecontext variableC over setsof events.H Focuscontributesto fixing thevalueof C.

In Rooth’s analysis,the focusinterpretationoperatorconstrainsthealternative set(C) presup-

posedby adverbsof quantificationto be a subsetof the focussemanticvalueof S. In the caseof

254

Page 272: Thesis about discourse

(5.58a), this valuewill be(5.58b), which Roothcallsa ‘focus closure’,thesetof eventsof Mary

takingsomeoneto themovies12.

(5.58a) Usually, Mary takesJOHNto themovies.

(5.58b) � t ¬¯h y[AT(t, [ ¾ Mary take y to themovies])]�The interpretationof (5.58a) is then: mosteventsof Mary taking someoneto themovies are

eventsof Mary takingJohnto themovies.

Essentially, we expect that all S-modifying adverbials that presupposean alternative set of

propositionswill besensitive to thepresenceof focus,becausethesetpresupposedby theadverbial

andthesetpresupposedby focuswill containat leastoneidenticalobject,namely, theinterpretation

of thesentencebeingmodified. For example,thealternative setsof bothadverbsof quantification

andfocuscontainthe modifiedclause. In contrast,the alternative setpresupposedby alternative

phrasesdoesnot containat leastoneobjectidenticalto anobjectin thesetpresupposedby focus,

andsotheir two setscannotbeequated.

This expectationthatwhentwo setscontainsimilar objectsit will bedifficult to interpretthem

astwo differentsetsis capturedin theAI planningheuristic“useexisting objects”[Sac77], i.e. if

focus interpretationmakes availablean alternative set,we will equateothersimilar presupposed

alternative setsto this setif possible.In fact, this authorhasnot comeup with a casein which, in

thepresenceof focusandanalternative-setpresupposingS-modifier, thetwo setsthey presuppose

arenot in somesortof set-subsetrelation. It appearsthat thesetevoked by themodifier is always

relatedto thesetevokedby focus.For example,consider(5.57c).

(5.58c) Mary is a greatgrandmother. Every Tuesdayshedoessomethingfun with eachgrand-

child. Usually, shetakesJOHNto themovies.

It appearsthatweinterpretthepresupposedsetof eventsassociatedwith usuallynotasthesetof

eventsof Mary takingX to themovies, but rather, eitherasthesetof eventsof Mary doingsomething

fun with hergrandchildJohn(e.g.Mary takesJohnto therecord store, Mary takesJohnto thepark,

etc.), or to thesetof eventsof Mary doingsomethingfun with eachof hergrandchildren(e.g.Mary

12The formulausestheAT operatorfrom tenselogic (see[Dow79]), whosesyntaxis consistentwith theevent logicdiscussedin [Kam79] ([Roo85]).

255

Page 273: Thesis about discourse

takesSusieto therecord store, Mary takesEvato thepark,etc.). However, it alsoappearsthatwhen

we produceeitherof theseinterpretations,we arealso prosodicallyfocusingmovies, so that this

focusis alsoproducinganalternative setin whichtheentireverbphrasevaries.Furtherstudymight

indicatethat thereis sometailoredcontext in which it is possible,however, to distinguishthe set

presupposedby anadverbialfrom thesetpresupposedby a focusedphrase,without thepresenceof

additionalfocus.

S-modifyingadverbialsthatdon’t presupposealternative setscanalsobesensitive to thepres-

enceof focus. For example,[KKW01a] show that the interpretationof the discourseconnective

althoughis sensitive to thepresenceof focusbecausetheconsequentof thedefeasiblerule it pre-

supposesanddenies(see[Lag98]) will bemadeavailableby the rhemealternative set. For exam-

ple,(5.59a)presupposesanddeniestheexpectationthat in normalcircumstances,if Clydemarries

Bertha,he will inherit somemoney, representedas in (5.59 b). The rhemealternative setof the

mainclauseincludesavarietyof alternatives,includingClydeinheritingall themoney, someof the

money, andnomoney, asexemplifiedin (5.59c).

(5.59a) AlthoughClydemarriedBertha,hedid not inherit aPENNY.

(5.59b) marry(c,b) t inherit(c,money)

(5.59c) ��N inherit(c,money), inherit(c,money), ...�Evaluative S-modifying adverbshave also beenargued([Kri, Kon91]) to be sensitive to the

presenceof focus.As Krifka statesit, (5.60a)presupposesthatthereis somealternativeX suchthat

it would have beenmorefortunatefor Mary to invite X for dinner. (5.60b) presupposesthat there

is somealternative X suchthatit wouldhave beenmorefortunatefor Mary to invite Bill to X.

(5.60a) UnfortunatelyMary invited BILL to dinner.

(5.60b) UnfortunatelyMary invited Bill to DINNER.

This presuppositionshouldnot beattributedto thesemanticsof unfortunately, however; rather,

it arisesdueto the presenceof alternativescreatedby the useof focus. Notice that Krifka states

thepresuppositionas“it wouldhavebeenmorefortunate”if anotheralternativeshadoccurred.In

otherwords,thepresuppositionisn’t presupposedtrue; we have no senseof whetheror not Mary

256

Page 274: Thesis about discourse

invitedanyoneelseto any othermeal.Thefocussimplymakesthesealternativesavailable.Thefact

thatthesealternativesareperceivedasbeing“more fortunate”arisessimplybecausethealternative

setis interpretednotascontainingpropositionsof theform UnfortunatelyMary invitedX to dinner

andUnfortunatelyMary invited Bill to X, respectively, but rathertheir unmodifiedversions,Mary

invitedX to dinnerandMary invitedBill to X. Thatthealternativesarenotunfortunatefollows; that

they aremorefortunatethansomethingunfortunatealsofollows.

It is not alwaysthecasethatanS-modifieris excludedfrom thepropositionsin analternative

setinvokedby a focusedphrase,however. For example,thesetof alternativescreatedby theuseof

focusin (5.61)appearsto bethesetof alternative actionsJohnmight have takenaftergettingup; in

otherwords,therelationsuppliedby thenbetweenall alternativesandtheprior clauseis maintained.

(5.61)Johngotup. ThenheLEFT.

We seethesameeffect in otherclausaladverbials,including theepistemics,asexemplifiedin

(5.62). Again thesetof alternativescreatedby theuseof focusappearsto bethesetof alternative

propositionsof theform ProbablyI sawX, ratherthanI sawX. In otherwords,therelationsupplied

by probablyto all thealternativesis maintained.

(5.62)ProbablyI saw MARY.

Moreover, it is not thecasethatall evaluative adverbsareintuitively excludedfrom thepropo-

sitionsin thesetof alternatives. For example,it is easyto imaginea situationin which thesetof

alternative propositionsintroducedby focusin (5.63)wouldhave beensimilarly not surprising.

(5.63)Not surprisingly, I saw MARY.

Ontheotherhand,theevaluativeadverbialbymistakeappearsto patternlikeunfortunately. The

alternativesof (5.64)have theform Mary invitedX to dinner, ratherthanBy mistake, Mary invited

X to dinner.

(5.64)By mistake,Mary invited BILL to dinner.

It would beinterestingto studywhetheror not it is predictablefrom themeaningof theadver-

bial whetherthealternativesintroducedby focuswill includetheadverbialor not.

257

Page 275: Thesis about discourse

More generally, thereis a significantamountof researchon the interactionof focusanddis-

courseconnectives; [KKW01b], for example,discussthe effect of focuson the resolutionof the

anaphoricargumentof otherwise, [Umb02] discussestheeffect of focusin conjunctedclauses,and

[Joh95] discussestheuseof focusin when-clauses.

5.3.5 FocusingS-Modifying Adverbials to Evoke Context

Focuseffectssuchasthosediscussedabovecanalsointeractwith thesemanticsof clausaladverbials

to evoketheprior context. An exampleis shown in (5.65).Following [Ste00a], theclausaladverbial

on December31 is therhemeandrhemefocus,andhis policy will expire is thetheme.This theme

presupposesarhemealternative setof theform X, hispolicyexpires, whereX rangesover temporal

coordinates.Onealternative temporalcoordinatecanbederivedfrom theS Michael’s coursesare

justaboutto end.

(5.65)Q: Michael’s coursesarejustaboutto end.Whenwill his insuranceend?

A: ON DECEMBER31,his policy will expire.

A clausaladverbialcanalsobeathemefocus,asexemplifiedin (5.66).ProsodyonOnFebruary

14 indicatesthe presenceof a themealternative setof the form Y, I’ll be at location X, whereY

rangesover temporalcoordinates.Alternativesarefound in thecontext, including tonight andthe

setof timesquantifiedby oftenin theclauseoftenyoutravel.

(5.66)

Q: Oftenyou travel, andI don’t know if you’ll behomefor dinner. For

example,I know thatyou’ll behometonight,but wherewill yoube

onFEBRUARY 14?

A: ON FEBRUARY 14, I’ ll bein HOUSTON.

A clausaladverbial canalsobea contrastive theme,asexemplifiedin (5.67)-(5.68).In (5.67),

the reply doesnot answerthequestionfully, ratherit assertswhat theansweris likely to be. The

rhemefocus, the movies, suppliesthe answerto what, restrictingthe rhemealternative setof the

form wewill X tomorrow. Thethemefocus,probably, indicatesthepresenceof alternative themes,

namelyanalternative setof propositionsof the form Y, wewill X tomorrow, whereY rangesover

possibleepistemicvalues(e.g.probably, definitely, possibly, etc).

258

Page 276: Thesis about discourse

(5.67)Q: Whatwill wedo tomorrow?

A: PROBABLY, wewill go to THE MOVIES tomorrow.

Similarly, in (5.68),thereply doesnot answerthequestionfully, ratheranopinionof what the

answershouldbe is asserted.The rhemefocus,flat, restrictsthe presupposedset of alternative

rhemesof theform Congressshoulddecideon a X tax. Thethemefocus,myopinion, indicatesthe

presenceof alternative themesof theform In Y’sopinion,Congressshoulddecideona X tax, where

Y rangesover contextually relevant individuals.

(5.68)

Q: I know whatkind of taxFredthinksCongressshouldimplement,but whatkind

of tax do you think Congressshouldimplement?

A: In MY VIEW, CongressshouldimplementaFLAT tax.

DLTAG (see[FMP�

01, FW02]) hasalreadyarguedthat nowadaysfunctionsasa contrastive

themein examplessuchas(5.69),wherewhat thehusbanddoesnowadayscontrastswith whathe

usedto do,or will do in thefuture.

(5.69)Q: Whatdoesyourhusbanddo?

A: NOWADAYS, hetakescareof thekids..

Furthermore,we find corpusexamples,suchas(5.70),which show thatnowadaysasa theme

cancontrastwith analternative derivedfrom theVP usedto.

(5.70) To write a play, the dramatistusedto draw on his imaginationandknowledgeof life.

Nowadays,all hedraws onsomeoneelse’s book. (simplifiedBROWN example)

[Ste00a]’s analysisof thesyntax-phonologyinterfacemakesuseof a rangeof specificprosodic

tunesthat have beenassociatedwith sub-clausalconstituentswhen they function as themeand

rheme. In orderto studyhow prosodyeffectsthediscoursecoherenceof S-modifyingadverbials,

we mustfirst understandtherangeof intonationpatternsthatcanbeassociatedwith them.[AC74],

for example,studytheappropriateintonationpatternsfor British EnglishS-modifyingadverbs. It

maybethecasethatthereis somespecifictuneassociatedwith all S-initial,S-modifyingadverbials,

dueto their beinglocatedin topic position. Unfortunately, the variety of prosodictheoriesin the

literature(see[PH90])havenotyetyieldedreliableannotationof speechcorpora,makingit difficult

to studyprosodiceffectsbasedon somethingother thanintuition. Moreover, in text corpora,the

259

Page 277: Thesis about discourse

writer’s prosodicintentionsarelost; interpretationof prosodyrelieswholly on intuition.

In additionto beingfocusedor containingfocusedelements,anotherway which focuseffects

canbeemployedto causea clausaladverbial to evoke context is throughtheuseof a focusparticle

modifier. In our ADVP corpus,approximatelysixty tokensare found in which a focus particle

is modifying an ADVP adverbial, andin our PPcorpus,approximatelyforty tokensarefound in

which a focusparticle is modifying a PPadverbial. In the majority of casesin both corpora,the

focusparticlemodifieris even. Someexamplesareshown in Table5.1.

As shown, someof theseadverbialsarealreadyanaphoricor comparative andsomayalready

dependon abstractobjectsin the discoursecontext for their interpretation. Although, as stated

above, theseexamplesaredrawn from text corporaandit is thusnot possibleto determinetheir

intonationpatterns,that focusparticlesandfocuscancausea clausaladverbial to requirethedis-

coursecontext for its interpretationis exemplifiedin (5.71),usingoneof thetokensfrom Table5.1.

In (5.71),thereis rhemefocusongrandparent, andthemefocusonnormalexperience. Alternatives

to therhemefocusarefoundin thecontext asfriendsandneighbors, while alternativesto thetheme

focusarederivedfrom therelative clausewhoselivesare surroundedbycrime.

(5.71)Childrenwhoselivesaresurroundedby crime may frequentlyfacethedeathof friends

andneighbors.However, even in NORMAL EXPERIENCE,many childrenarelikely to facethe

deathof aGRANDPARENT.

Table5.1: ADVP/PPAdverbialswith FocusParticleModifier

# ADVP Adverbial # PP Adverbial

4 evennow 1 alsounlike Mr.Ruder18 evenso 1 evenatacar’s length3 eventhen 1 evenin normalexperience1 only adecadeago 1 evenin that1 only hoursearlier 1 evenonhis toughconstitution1 evenmoreremarkably 1 evenwithoutdeals

Interestingly, while our corpusrevealsthatfocusparticlesarefrequentlyfoundmodifying sub-

ordinatingconjunctions(e.g.even/onlyif/when), they donotappearto befoundasfrequentlyin our

corpusmodifying themostcommondiscourseadverbialsthattake hiddensemanticarguments.For

260

Page 278: Thesis about discourse

example,even/onlyasa result/inadditionall soundodd,anddonotappearin ourcorpus.Moreover,

wefind thatthenin (5.72a)cannotbeinterpretedasasequencerelationbetweentheeventinterpre-

tationsof thetwo sentences,akin to after that in (5.72b). It canonly beinterpretedasa discourse

deicticor anaphoricreferenceto the temporalcoordinateof thefirst sentence,akin to at that time

in (5.72c)13. However, evenafter that andevenat that time in (5.72b)-(5.72c) respectively, are

felicitous. A full understandingof how focusparticlescombinewith discourseadverbialsrequires

furtherstudy, andmayturnout to bebestanalyzedusingagradientnotionof processingdifficulty.

(5.72a) Johnwon’t wake up until latein theafternoon.Eventhenhewill eatbreakfast.

(5.72b) Johnwon’t wake up until latein theafternoon.Evenafter that hewill eatbreakfast.

(5.72c) Johnwon’t wake up until latein theafternoon.Evenat that timehewill eatbreakfast.

5.3.6 Summary

In the prior sectionwe presentedanalysesthat addresshow the prosodicfocus on a sub-clausal

constituentcaneffect the interpretationof the sentencecontainingit, causingthat sentenceto be

interpretedwith respectto thediscourse.In thissectionweshowedhow theseanalysesalsoaccount

for “focussensitive” S-internalmodifiers.In theseaccounts,thesemanticsof such”focusparticles”

aredefinedin termsof (e.g. dependenton) thealternative setevoked by thepresenceof prosodic

focus on or in the sub-clausalconstituentmodified by the focus particle. We then investigated

the usesof thesefocus particlesas S-modifiers. We first addressedcaseswherealthoughthere

wasno obviously focusedsub-clausalelement,the S-modifierwasneverthelessinterpretable.It

appearsthatin suchcasesthediscoursecontext suppliesareferentfor theS-modifier’spresupposed

alternative set.We thenaddressedcaseswherebotha ”focus sensitive” S-modifierandsub-clausal

prosodicfocuswerepresentin a clause. Basedon the existenceof thesetwo cases,we suggest

that focus particlesarenot unconditionallydependentfor their interpretationon the presenceof

(sub-clausal)focus. Rather, the focus“sensitivity” of certainmodifiersarisesdueto thesemantic

similarity betweenthe alternative set they presupposeand the alternative set that prosodicfocus

presupposes.Dueto this similarity, whenbothelementsarepresentin a clause,their presupposed

13Thedifferencebetweenthesetwo interpretationof thenis discussedin moredetail in Chapter3.

261

Page 279: Thesis about discourse

alternative setsdisplaya strongtendency to resolve to thesamecontextual or accommodatedset.

Finally, we investigatedthe effect of focuson S-modifiersthat have not previouly beenclassified

as”focus sensitive”, e.g. that don’t presupposealternative sets. We first presentedprior analyses

of how focuscaneffect the interpretationof somecommondiscourseandclausaladverbials. We

thenillustratedsomewaysthatfocusonor in otherclausalanddiscourseadverbialscaneffect their

interpretedwith respectto thediscoursecontext.

5.4 Implicatur es

As [Hir91] notes,a variety of “meanings”over andabove the literal contentof an utteranceare

conveyed whena speaker uttersa sentence.Classesof suchmeaningshave beendistinguishedin

thelinguisticsliteratureasshown in (5.73).

(5.73)H Entailments:meaningswhichmustalsobetruewhenthesentenceis true(see[ODA93])H Presuppositions:meaningsentailedby boththesentenceandits negation(see[Bea97])H Implicatures:non-truth-functional meanings(see[Gri89])H IllocutionaryForce:speaker’s actvia theutterance(e.g.asserting, promising)(see[Sea69])H PerlocutionaryEffect: effect of the utteranceon hearer(e.g. convincing, inspiring)(see

[Sea69])

As discussedin Chapter3, entailmentis generallyconsideredpartof thetruth-functionalcom-

ponentof utteranceinterpretation.Presuppositionis sometimesequatedwith implicature; thelatter

is the focusof this section;the distinctionwill alsobe discussedin this section. We will return

briefly to illocutionary forceandperlocutionaryeffect in Section5.6.

5.4.1 Gricean Implicatur e

[Gri89] early proposedan influential pragmaticapproachto accountfor the non-truthfunctional

meaningsconveyed by an utterance. He distinguisheswhat is said from what is implicated, as

262

Page 280: Thesis about discourse

shown in Figure5.1.

Whatis said representscontext-independentmeaningsthatdeterminethetruthconditionsof an

utteranceandcanbeaccountedfor with a truth-functionalsemantics.Whatis implicatedrepresents

non-truth-conditional meanings.

Conventionalimplicaturesaredefinedasmeaningsthatarebothnon-truth-functional andcontext-

independent.Many researchersview themasidenticalto pragmaticpresuppositions;bothwill be

discussedlaterin thissection.

Non-Conventional implicatures, in contrast,are definedas non-truth-functional and context-

dependentmeaningsthatarisein a givencontext dueto thespeaker’s andhearer’s mutualrecogni-

tion of rulesgoverningconversation.While suchmeaningscanbe linkedto non-linguisticcontext

(Non-Linguistic), includingaestheticandculturalknowledge,in thissectionwe focuson thosethat

arelinkedto language(Conversational).

UtteranceInterpretationð ð ð ð ð ð ðñññññññWhatis Said Whatis Implicated� � � � � � ��������

Conventional Non-Conventional� � � � � �������Non-Linguistic Conversational� � � �����

Generalized Particularized

Figure5.1: GriceanFramework

[Gri75, 45] arguesthat a singleCooperative Principle (CP) is known to conversationpartici-

pants:Makeyourconversationalcontribution suchasis required,at thestageatwhich it occurs,by

theacceptedpurposeor directionof thetalk exchangein whichyouareengaged.

[Gri75, 46-47]assertsfour maximswhich furtherspecifyhow theCPis observed:H Maxim of Quantity: Make your contribution as informative as is required(for the current

purposesof theexchange.Do notmake yourcontribution moreinformative thanis required.

263

Page 281: Thesis about discourse

H Maxim of Quality: Try to make your contribution one that is true. Do not saywhat you

believe to befalse.Do not saythatfor whichyou lackadequateevidence.H Maxim of Manner: Be perspicuous.Avoid obscurityof expression.Avoid ambiguity. Be

brief (avoid unnecessaryprolixity). Be orderly.H Maxim of Relevance:Berelevant.

Gricearguesthatbecausetheserulesarestandardlyfollowedby conversationparticipants,pur-

poselyexploiting themcanconvey additionalmeanings,or conversational implicatures. An oft-

citedexampleof exploiting theMaxim of Quantityis shown by the“letter of recommendation”for

astudentof philosophy, shown in (5.74).

(5.74)DearSir,

Thisstudent’s Englishis grammatical,andhishandwritingis legible.

Yours,...

As [Hir91] notes,it might seemthat themaxim is violated,becausea philosophyrecommen-

dationis generallyexpectedto containa significantnumberof favorablestatementsrelevant to the

student’s skills in philosophy. This letter containsonly two statements,which, thoughfavorable,

arenot very relevant to philosophy. Griceargueshowever that thewriter is in factobeying theCP

andits maxims;the writer conveys by this letter that s/hehassaidasmuchass/hetruthfully can

abouthis/herstudent’s skills in philosophy, i.e. that s/hehasnothingfavorableto sayspecifically

pertainingto theseskills.

Griceviews this asa particularizedconversationalimplicature. Like all conversationalimpli-

catures,it is context dependentbecausechangingthecontext canchangeor remove it; for example,

if thepersonrequestingthe letter hadstatedthat thewriter wasto commentonly on thestudent’s

Englishandhandwriting,theimplicaturewouldnolongerarise.It is particularizedbecausethecon-

text that licensesit is particular or special. In contrast,a generalizedconversationalimplicature,

suchasthatshown (5.75b) whichariseswhenaspeaker asserts(5.75a) (examplefrom [Hor96]), is

still context dependentin thatchangingthecontext canremove or alter it, but is generallylicensed

in the absenceof particularor marked contexts; a particularizedcontext canremove it, suchasa

264

Page 282: Thesis about discourse

gameof hideandseekin which thespeaker knows thelocationof his wife but doesnotwantto tell

theseeker. In eachcase,however, it is not thepropositionitself thatlicensestheimplicature,but the

CP, its maxims,andthecontext.

(5.75a) My wife is eitherin thekitchenor in thebathroom.

(5.75b) I don’t know for a factthatmy wife is in thekitchen.

In general,Grice argues,speakers obey the CP and its maxims,andrely on the fact that the

hearerknows this in orderto convey conventionalimplicatures.Of course,a speaker maychoose

to lie, therebyviolating theCP andthemaximsby deliberatelymisleadingthe hearer. Moreover,

a speaker may chooseto opt-out of the CP and the maxims,for example,by invoking the Fifth

Amendmentin acourtof law.

It is often difficult to decidepreciselywhich maxim a speaker intendsto invoke to convey a

conversationalimplicature.In (5.76),for example,all of themaximscanberelatedto theinference

of theimplicature.Whenaspeakerutters(5.76b) in answerto thequestionin (5.76a),s/helicenses

the conversationalimplicaturein (5.76 c) by the sharedassumptionbetweenspeaker andhearer

that the speaker is sayingasmuchas(Quantity)s/hecantruthfully (Quality) saythat is relevant

(Relevance)andsheis sayingit in a way thatis not ambiguousor obscure(Manner).For example,

although(5.76b) is entailedif thespeaker hasfour dollars,saying(5.76b) in thatcaseviolatesthe

maximsof QuantityandManner.

(5.76a) Do youhave any money?

(5.76b) I have threedollars.

(5.76c) I don’t have morethanthreedollars.

Therehave beena varietyof attemptsto formalizetheCPandits maxims,aswell asattempts

to categorize conversationalimplicaturesand formalize how they are inferred (see[Hir91] and

referencestherein). Quantity-basedimplicatures,that is, generalizedconversationalimplicatures

which arisedue to the maxim of Quantity, have received a lot of attentionin the literature(see

[Hor96, Gaz79a, Hir91]). [Hir91] calls themscalar implicatures. Shearguesthat the successful

conveyanceof ascalarimplicaturereliesonthespeaker’sandhearer’smutualperceptionof therank-

265

Page 283: Thesis about discourse

ing of thespeaker’s utterancewith respectto theotherutterancess/hemight have utteredinstead.

Building on earlierwork, her theoryspecifiestheconditionsunderwhich a speaker may licensea

scalarimplicatureandthatahearermusthaveaccessto in orderto interpretthis implicature.In par-

ticular, shecitesa rangeof research(includingherown) which hasshown thatquantifiers,modals,

conjunctions,numerals,definitesand indefinites,spatio-temporalorderings,epistemicverbsand

verbsof incompletion,set/subsetandentity/attribute andgeneralization/specialization relationsas

well asahostof otherformsin naturallanguageentailor otherwiseevoke orderings(linearaswell

ashierarchical)thataspeaker canuseto convey scalarimplicature.

In (5.76)wesaw anexampleof ascalarimplicaturearisingfrom anorderinginducedby theuse

of a cardinal,the numberthree. In the context of the questionposedin In (5.76a),by answering

with (5.76b) theimplicatureconveyedis that threeis anupperboundon theamountof money the

speaker has. The lower valuesarein fact entailed,e.g. if you have threeof something,you also

have two. If thequestionhadinsteadbeen“Can you afford themagazine?”(whereit is mutually

known that themagazinein questioncoststhreedollars),the implicatureconveyed in (5.76b) due

to the maxim of Quantity is that the speaker hasmore thanthreedollars, i.e. that he canafford

themagazinewithout breakinghis wallet. While valuesabove threearenot entailedby theuseof

three, it appearsto evoke theordering;it is generallyacceptedthatmentionof a cardinalW maybe

ambiguousbetweenthereadingexactlyn, at mostn, andat leastn[Hir91].

In (5.77)we seeanexampleof a scalarimplicatureof notall arisingfrom theorderinginduced

by speaker A’s useof the quantifiersome, which speaker B thenquestions,andspeaker A must

correct(examplefrom [Hir91, 84]). Theorderinginducedby quantifierscannothoweverbedefined

aslogical entailment;in logical terms,e x P(x) doesnot entail h xP(x). However, as[Hir91] notes,

while this meansthatuniversallyquantifiedstatementssuchasAll X are Y doesnot entailSomeX

is Y, it canbeassumedthatuniversallyquantifiedstatementssuchasAll of theX likeY arelogically

representedas h xP(x) J¼e x P(x) andsodo entailtheirsomeX counterparts.

(5.77)

A: Well, someof it youcanchargeto yourgrant.

B: Some?

A: Oh,all.

266

Page 284: Thesis about discourse

In thesameway, however, modifiersthatquantifyoverpersons,placesor times,or thingsinduce

orderingsthataspeaker canuseto implicatethatuppervaluescannotbeassertedtruthfully. For ex-

ample,aspeaker’s useof somebodymayimplicatenoteverybody, andaspeaker’s useof sometimes

mayimplicatenotoften,notusually, notalways.

[Hir91] notesthatunlike quantifierorderings,modalorderingsareentailedin modalsystems,

where,for example,if a propositionis certain,thenit is alsopossible.In (5.78)we seeanexample

of a scalarimplicatureof not certainarisingfrom theorderinginducedby speaker B’s useof the

modalmay, which indicatesthatthepropositionis possible(examplefrom [Hir91, 84]).

(5.78)A: Youwerein theneighbourhoodof thepantryat onetime,wereyounot?

B: I mayhave been.

Hirschberg arguesthatconjunctionsalsoinduceorderingsthat licensescalarimplicatures.We

saw oneexamplein (5.75),hereandin (5.79),[Hir91] arguesthator includesanalternative setof

propositions.By assertingonly oneof thesepropositions,speakerB impliesthatthealternativesare

false(or unknown). This implicaturedoesnot follow logically, (i.e. asan exclusive disjunction),

becausethe speaker could cancelthe implicature(seebelow), e.g. by addingand dinner sounds

goodtoo.

(5.79)A: Do youwantto goout to dinneror seeamovie?

B: A movie soundsgood.

5.4.2 Pragmatic and SemanticPresupposition

As statedabove, [Gri75] definesconventional implicaturesas meaningsthat areboth non-truth-

functionalandcontext-independent; he views them asarisingby virtue of the meaningof some

word or phrasethespeaker hasused.He distinguishedthemfrom conversationalimplicaturesac-

cording to their cancelabilityand detachability. Conversationalimplicaturesare cancelablebut

non-detachable.For example,theimplicatureof theletterof recommendationin (5.74)canbecan-

celedby altering the context or appendingadditionalmaterial(e.g. but I don’t meanto suggest

that...). However it is non-detachable,in that expressingthe literal contentof what is saidusing

differentlexical items(e.g.penmanshipinsteadof handwriting) doesnot remove theimplicature.

267

Page 285: Thesis about discourse

In contrast,conventional implicaturesare non-cancelableand detachable.For example, the

conventionalimplicaturethat arisesfrom the speaker’s useof but in (5.80 a) is (5.80 b). That it

is non-cancelableis shown by the infelicitousnessof (5.80 c). That it is detachableis shown in

(5.80d); usingand insteadof but suppliesthesametruth-conditionalcontentwhile detachingthe

implicature. Moreover, the truth of what is said in (5.80 a) is not dependenton the truth of the

implicaturein (5.80 b); (5.80 a) is true if and only if Mary is both poor and honest,and false

otherwise.

(5.80a) Mary is poorbut honest.

(5.80b) Thereis somecontrastbetween(Mary’s) poverty and(her)honesty.

(5.80c) *Mary is poor but honest,althoughthere’s no connectionbetween(her) poverty and

(her)honesty.

(5.80d) Mary is poorandhonest.

Griceusesthesamenotionsto distinguishwhathecallssemanticpresupposition.In contrastto

implicature,presuppositionsareneithercancelablenordetachable.For example,thepresupposition

thatarisesin (5.81a) is (5.81b). Thatit is non-cancelableis shown by theinfelicitousnessof (5.81

c). Thatit is non-detachableis shown in (5.81d); usingceasedinsteadof stoppedsuppliesthesame

truth-conditionalcontentandretainsthepresupposition.Moreover, thetruthof whatis saidin (5.81

a) is dependenton the truth of the presuppositionin (5.81b); (5.81a) is true if andonly if (5.81

b) is true(whetherit is truth valueis falseor unknown if (5.81b) is falsevariesdependingon the

semantictheory.

(5.81a) Michaelhasstoppedbeatinghis wife.

(5.81b) Michaelhasbeenbeatinghis wife.

(5.81c) *Michael hasstoppedbeatinghis wife, althoughhenever beatherin thefirst place.

(5.81d) Michaelhasstoppedbeatinghis wife.

Similarly, [Kar73] distinguishsemanticandpragmaticpresupposition.In this view, if the se-

manticpresupposition(s)of asentenceis true,thenit is truewhetherthatsentenceis trueor false.A

pragmaticpresuppositionof asentence,in contrast,mustbeentailedin contextswherethatsentence

268

Page 286: Thesis about discourse

is felicitously uttered.In [KP79], pragmaticpresuppositionis equatedwith conventionalimplica-

ture,andit is notedthatmany of thelinguistic investigationsin theliteraturethathave invoked the

notionof presuppositionarein factinvoking pragmaticpresuppositions.

Theconfusionover whethersomethingis a semanticor pragmaticpresuppositionhasbeenfre-

quentlydiscussedanddebatedin theliterature,eithergenerally, or with respectto specificlinguistic

expressionsthat trigger them. Somedoubtthat semanticpresuppositionsexist at all, arguing that

presuppositionsdon’t assertrequirementson truth-conditions,but ratheron theappropriatenessof

utterancesin context (see[Bea97, Hor96, Gaz79b, Sta74, KP79,Str59]). [vdS92], ontheotherhand,

hasarguedthat presuppositionalexpressionsarein fact anaphoricexpressions,while [Sim00] ar-

guesthatthepresuppositionsassociatedwith change-of-stateverbs(e.g.startandstop) areactually

conversationalimplicatures.

Someparticularexpressionsto which presuppositionhasbeenattributed areshown in(5.82);

many of thesehave beendiscussedin this thesis.

(5.82)H definite,quantified,anaphoric,comparativeandfactivedescriptions(see[HK98, Bea97, KK70])H focusparticlesandfocus(see[Roo95a])H discourseconnectives(see[Lag98, KKW01b, Ste00b, KP79])H factive, change-of-state,andjudgmentverbs(see[KK70, Bea97])H cleftsandWH-questions(see[Pri86])

In semantictheories,presuppositionis usually definedas a binary relation betweenpairs of

sentences[Bea97]; onesentencemight presupposeanotherin a semantictheoryif the truth of the

secondis apreconditionfor thefirst to betrueor false,asdiscussedabove. Therepresentationof the

presuppositionhasbeenvariouslyhandledin semantictheories,dependingon the trigger andthe

theory;partial functions,hiddenarguments,assignmentfunctionshave all beenemployed for this

purpose,asdiscussedin Chapter3. Standardtestsfor semanticpresuppositionincludeembedding

269

Page 287: Thesis about discourse

undernegationor modaloperators,anddiscoursecontext tests,althoughthe latter hold for both

typesof presupposition,andthetestsarenotalwaysapplicableor usefulfor agivenexpression.

In pragmatictheories,presuppositionis definedin termsof the attitudesand knowledgeof

languageusers[Bea97], with or without referenceto specificlinguistic forms (e.g. thesentence).

Therepresentationof presuppositionsin pragmatictheoriesmayemploy modaloperatorsexpressing

intention,belief,andmutualbelief,asin [Hir91].

In mosttheoriesof presupposition,thepresupposedmeaningis generallyassumedto beeither

known, foundor inferablefrom thecontext, or accomodatable.

5.4.3 Summary

In this sectionwe have introducedthenotion of conversationalimplicatureasadditionalmeaning

that arisesfrom the speaker’s assumptionthat discourseis coherent.Conversationalimplicatures

areusedby a speaker to convey additionalmeaningover andabove the literal contentof what he

says.We have distinguishedconversationalandconventionalimplicatureaccordingto thenotions

of cancelabilityanddetachability, anddiscussedhow analysesof conventionalimplicatureoverlap

with semanticand pragmatictheoriesof presupposition.In the next section,we discusshow a

speaker canuseS-modifyingadverbialsto convey implicatures.

5.5 UsingS-Modifying Adverbials to Convey Implicatur es

In this sectionwe briefly review how our analysisin this thesishasalreadyinvoked thenotionof

presuppositionwhich existing semantic(or pragmatic)theoriesmustaccountfor. We thenshow

how a speaker’s useof S-modifyingadverbialscanconvey meaningsakin to whatGricehascalled

conversationalimplicatures,whichdiscoursetheorymustalsoaccountfor.

5.5.1 Presupposition

In this thesis,wehaveencounteredavarietyof usesof thetermpresuppositionasasemanticnotion,

anda varietyof semanticrepresentationsfor it, dependingon theenvironmentin which it wasen-

countered.For example,in our ADVP/PPadverbialdataset,we saw that theinternalPPargument

270

Page 288: Thesis about discourse

or theADJ derivative of theADV canbe(or containelementsthatare)anaphoric,quantified,adef-

inite description,comparative, and/orcontaina hiddenargument.We alsosaw that theadverbials

in our datasetcaninteractwith focus. While we focusedour discussionon how semanticrepre-

sentationsof theselinguistic forms could be extendedto their usewithin S-modifyingadverbials,

therebyaccountingfor thefactthatthey maydependfor their interpretationontheAO interpretation

of non-NPconstituentsin theprior context, we saw above that thedependency of theselinguistic

formson their context canbeviewedaseithera semanticor pragmaticpresupposition.

5.5.2 Conversational Implicatur es

It is not always the casethat the dependency of an S-modifying adverbial on discoursecontext

canbeaccountedfor wholly semantically. For example,asnotedin the introduction,S-modifying

adverbialssuchasactuallyor really take only oneAO argumentsemantically:theinterpretationof

the modifiedclause.As discussedin Chapter3, theseadverbssupplyepistemicfeaturesto the S

they modify. While focuslikely playsa role in their interpretation,conversationalimplicaturecan

alsobeinvolved. A varietyof examplesof S-modifierssimilar in meaningto actuallyareshown in

Table5.2alongwith their corpuscounts.

Table5.2: Higher-OrderedEpistemicAdverbialsYielding Implicatures

Count Adverbial Count Adverbial

31 actually 1 in truth105 in fact 6 really2 in reality 16 surely

In Section5.4we discussedhow modalsinducemodalorderingssuchthata speaker’s useof a

modalcanyield a scalarimplicature.In (5.78), for example,thespeaker’s useof maywasargued

to implicatethat the truth of thepropositionis not certain or not knownto him/her. In (5.83),on

theotherhand,[Hir91] arguesthatby denying theassertionof ahighervaluein themodalordering,

thespeaker conveys thatthelessercanis trueor unknown to him/her. In bothcasesthemodalvalue

involvedin theimplicatureis presentin thecontext (i.e. in A’sutterance).

271

Page 289: Thesis about discourse

(5.83)A: I would like to know if I cantake off thebackplate.

B: Youshouldn’t have to.

As notedin Section5.4,modalorderingsareentailedin modalsystems;for example,if apropo-

sition is actual,real, sure, or in thesetof facts, thenit is alsopossible, probable, etc. Thus,while

a speaker cannotassertsomethingis actual, real, sure, or in thesetof factsandat thesametime

implicatethata lessermodalvalueof thepropositionsis not known or certainto him/her, we argue

thatthespeakercanusethesehighermodalvaluedS-modifierswhenhebelievesthata lessermodal

valueof thepropositionsis not mutuallyknown or certain.In otherwords,peopledo not normally

assertthat a propositionis true, to do so would violate (at least)the maximsof Quantity(do not

makeyourcontribution moreinformative thatrequired)andManner(bebrief). Thatis, unlessthere

is evidenceindicatingto thespeakeror writer thattheheareror readermaysupposethemodalvalue

of thepropositionto befalseor unknown, or its truth unexpected.Suchevidencemight comefrom

thecontext.

For example,in (5.84), from our corpus,the article writer citesthe quoteasfrom Cervantes.

Thewriter of thereplyappearsto take thisasevidencethatthearticlewriter did notknow thatKing

Solomonis theoriginal source.Theimplicaturearisingby B’s useof actually, thatthetruth of the

modifiedpropositionwasnot mutuallyknown, is however bothcancelable(e.g. if B appends,“but

I bet you alreadyknew that”) andnon-detachable;B cansubstitute,for example,in fact, andstill

achieve thesameimplicature.

(5.84)Your Oct. 2 articleon DanielYankelovich cited thequote“A goodnameis betterthan

greatriches” asbeing from Cervantes’“Don Quixote.” Actually, Cervantesborrowed that quote

from awriter of some25centuriesearlier:Israel’s King Solomonwrotethosewordsin theBookof

Proverbs(22:1). (WSJ)

Theseadverbialscanalsobeusedwhencontext impliesa propositionthat is “hard to believe”,

asin (5.85),alsofrom ourcorpus.

(5.85) CathrynRice could hardly believe her eyes. While giving the Comprehensive Testof

BasicSkills to ninthgradersatGreenville High SchoollastMarch16,shespottedastudentlooking

at crib sheets.Shehadseencheatingbefore,but thesenoteswereuncanny. “A stockbroker is an

272

Page 290: Thesis about discourse

exampleof ...” Virtually word for word, the notesmatchedquestionsandanswerson the social-

studiessectionof thetestthestudentwastaking. In fact, thestudenthadtheanswersto almostall

of the40 questionsin thatsection.(WSJ)

Furthermore,asin (5.86), theseadverbialscanbe usedto explicitly deny the truth valueof a

propositionthathasbeenassertedby anotherspeaker.

(5.86)A: Youarewrong.

B: Actually/Surely/Really/Infact, I’m notwrong.

Of course,[Hir91]’s analysisof (5.78),wherethespeaker’s useof maycanimplicatethat the

the truth of the propositionis not certain or not known to him/her, can be directly extendedto

S-modifyingadverbialsthat asserta lower valuein the modalordering;thesetoo canbe usedby

a speaker to implicate that a higher value is falseor unknown. Someexamplesof S-modifying

adverbialswhoseinternalargumentis interpretedasa lower-orderedmodalityandcanbeusedfor

thispurposeareshown in Table5.3.

Table5.3: Lower-OrderedEpistemicAdverbialsYielding Implicatures

# Adverbial # Adverbial

4 in asense 1 in onesense4 in away 2 in oneway1 to adegree 2 in certainrespects1 to anextent 5 with few exceptions

The interpretationof theseadverbialsasconveying a modality is due to the interpretationof

their internalargument,asdiscussedin Chapter3. Theirinterpretationasconveying alower-ordered

modality, however is dueto thedeterminersthatmodify theseinternalarguments.[Hir91]’sanalysis

of (5.77)), whereA’s useof the quantifiersomecanyield a scalarimplicaturenot all, thus also

appliesto S-modifyingadverbials,asdoesheranalysisof quantifiersover people,placestimesand

things(e.g.sometimes), andheranalysisof cardinalW yieldingscalarimplicaturesof at mostn or at

leastn or exactlyn. Someexamplesof S-modifyingadverbialscontaininglower-orderedquantifiers

over people,timesandplacesthatcanbeusedto convey suchimplicaturesareshown in Table5.4.

Interestingly, however, it appearsthatPPS-modifyingadverbialscontainingthehigher-ordered

273

Page 291: Thesis about discourse

quantifiersall or anycanalsobeusedby speakersto convey an implicature. As notedin Section

5.4, in contrastto modalandtemporalorderings,e xP(x) doesnot entail h xP(x). Wearguethatthis

however canbetheimplicaturecausedby theuseof theadverbialsin Table5.5. Corpusexamples

containingadverbialsfrom thefirst columnof Table5.5areshown in (5.90)-(5.89).

Table5.4: Lower-OrderedQuantificationalAdverbialsYielding Implicatures

# Adverbial # Adverbial

7 atonepoint 1 for thosefew6 at times 1 on two occasions10 for amoment 1 to some

Table5.5: Higher-OrderedQuantificationalAdverbialsYielding Implicatures

# Adverbial # Adverbial

24 in any case 9 above all13 in any event 49 afterall6 at any rate 10 anyway

(5.87)Creative accountingis a hallmarkof federalcredit. Many agenciesroll over their debt,

payingoff delinquentloansby issuingnew loans,or convertingdefaultedloanguaranteesinto direct

loans.In anycase, they avoid having to write off theloans.(WSJ)

(5.88) His arm had beengiving him sometrouble and Rectorwas not enoughof a medical

expertto determinewhetherit hadhealedimproperlyor whetherHino wassimply rebellingagainst

thetediouswork in theprint shop,usingthestiffnessin his armasanexcuse.In anyeventRector

senthim to thelocalhospitalto have it checked...(BROWN)

(5.89)Manchester’s unusualinterestin telegraphyhasoftenbeenattributedto thefact that the

Rev. J. D. Wickham,headmasterof Burr andBurton Seminary, wasa personalfriend andcorre-

spondentof the inventor, SamuelF. B. Morse. At anyrate, Manchesterdid not lag far behindthe

first commercialsystemwhichwassetup in 1844betweenBaltimoreandWashington.(BROWN)

As discussedin Chapter3, due to their modificationby all or any, the abstractobjectscase,

event,ratedo notneedto beidentifiedwith setsof abstractobjectsin thecontext. They thusdo not

274

Page 292: Thesis about discourse

functionsemanticallyasdiscourseconnectives.Wearguethataspeaker’suseof theseadverbialscan

however implicatethatthereis somerelevantabstractobjector setof abstractobjectsin thecontext

whichshouldbeviewedascontainedin thesetsidentifiedwith theseinternalargumentsandwhich

thusshouldberelatedto theAO interpretationof themodifiedpropositionvia theprepositionhead

of the PPadverbial. Our basisfor this argumentis simply that,unlessthereis a relevant abstract

object(or set)underconsideration,peopledo not normallyassertthe relationof a proposition(or

otherabstractobject) to the setof all cases,events,or rates;to do so would violate (at least)the

maxim of Quantity (do not make your contribution more informative that required)andManner

(be brief). In the above examples,the context doescontainsa relevant case, event, rate. Once

an implicated(setof) abstractobjectsin the context is identified,the particularrelationsupplied

by the prepositioncanbe interpreted;of course,in mostcasesthis relationwill be a metaphoric

interpretationof thepreposition,or evenidiosyncraticto theadverbial.However, theseimplicatures

arecancelable,dependingon context, but non-detachable.For example,a speaker could say: “At

any rate,althoughI’m not consideringanything in particularthat we’ve discussedso far”, anda

speaker canusein anyeventandat anyrate interchangeably.

We argue that above all createsthe sameimplicature; by using an adverbial that relatesthe

modifiedpropositionto someunspecifiedsetof all, the speaker implicatesthat somerelevant set

of abstractobjectsis to be found in the context. An examplewherethis is the caseis shown in

(5.90)(therelevantsetincludesall theAOsthatthepresidentdid or did notwant).Of course,above

all doesnot assertthat themodifiedpropositionis physicallyabove all otherrelevantpropositions;

ratherit assertsthat the modifiedpropositionis the most important(above in status)proposition

amongall relevantothers.

(5.90)ThePresidenthadsetfor himselfthetask,whichhebelievedvital, of awakeningtheU.S.

andits allies to the hardandcomplex effort necessaryto shift that balance.He did not want the

effort weakenedby any illusion that summitmagicmight make it unnecessary. He wantedtime,

too, to review theUnitedStates’globalcommitmentsandto testboththepolicieshehadinherited

andnew oneshewasformulating.Aboveall, hedid notwantto appearto berunninghatin handto

PremierKhrushchev’s doorstep.(WSJ)

275

Page 293: Thesis about discourse

While we arenot attemptingto completelyaccountfor the idiosyncraticmeaningof thesead-

verbials,we do believe theseimplicaturescanbe employed andcanaccountfor onepart of their

interpretation. Note however that it is sometimesdifficult to understandhow the assumptionof

discoursecoherenceyieldsthecontext dependency of a discourseadverbial. For example,while it

may be that thesameimplicaturesarisewhena speaker usestheadverbialsafter all andanyway,

therelationof their linguistic form to their meaningis muchmoreabstract.Considertheexamples

from ourcorpusin (5.91)-(5.92).

(5.91)Thestockmarket’s dizzyinggyrationsduringthepastfew dayshave madea lot of indi-

vidual investorswish they couldbuy somesortof insurance.After all, they won’t soonforget the

stockbargainsthatbecameavailableaftertheOctober1987crash.(WSJ)

(5.92)Therearemany, many thingsto do. Find out what you like to do mostandreally give

it a whirl. If you can’t think of a thing to do, try something– anything. Maybeyou will surprise

yourself.True! Wearenotall greatartists.I, frankly, can’t draw astraightline. Maybeyou arenot

thatgiftedeither, but how aboutputteringaroundwith theold paints?Youmayamazeyourselfand

acquirea realknackfor it. Anyway, I’ ll betyouhave a lot of fun.(BROWN)

As theseexamplesindicate,after all doesn’t assertthatthemodifiedpropositioncomestempo-

rally or textually afterall thatprecededit; rather, it appearsthat in many casesit canbepreceded

by because, andit appearsto indicatethat themodifiedpropositionis amongthemostsignificant,

andalreadyknown, possiblecauses.It may be that after all shouldbe analyzedas taking as its

argument,or implicating,asetof causesarisingfrom aninferredor explicit causalrelationbetween

themodifiedclauseandtheprior discourse.[WJSK03]proposea similar analysisfor for example,

whoseinterpretationappearsto beparasiticonapreviousexplicit or inferredstructuralrelation,but

while for for examplethis relationmaybe oneof cause,result,or elaboration,or suppliedby the

idiosyncraticmeaningof averb,after all seemsto requireacause.

Anywayappearsto indicatesa returnto a discussionthat got off on a tangent. It may be that

anywayshouldbe analyzedas taking as its argument,or implicating, one or more elaborations

arisingfrom an inferredor explicit elaborationrelationbetweenthemodifiedclauseandtheprior

discourse.

276

Page 294: Thesis about discourse

This discussionis exploratory; while it is clear that the modifiedclausesof theseadverbials

are beingrelatedto theprior discourse,eitherthroughfocus,implicature,idiosyncraticlexical se-

mantics,or somecombinationthereof;in orderto determinewhetheror not it is usefulto consider

their linguistic form andits interactionwith implicaturesascontributing to their interpretation,an

annotationprojectsuchasthatdescribedin Chapter4 is required.

5.5.3 Interaction of Focusand Implicatur e

Thus,wecaninvoke thenotionof conversationalimplicatureto explainwhy S-modifiershavebeen

treatedasdiscourseconnectiveseven thoughtheir linguistic form doesnot causethemto refer to

anabstractobjectin theprior discourse:by usingthemthespeaker hascreatedan implicature. In

many of thecaseswhereimplicaturesarise,therealsoappearsto befocus. It wasnotedin Section

5.2 that focuscaneffect the interpretationsof conversationalimplicatures.[Hir91] alsodiscusses

how intonationcaneffect thedisambiguationof variousimplicatures;for ourpurposes,wenotethat

the two mechanismsof discoursecoherenceinteract;it maybethat to induceanawarenessof the

orderinginvokedby linguistic forms,they mustbefocused.This requiresfurtherstudy, however.

5.5.4 Summary

In this sectionwe have addressedhow S-modifyingadverbialscanbe usedby a speaker to create

conversationalimplicatures,causingthesentencecontainingthemto beinterpretedwith respectto

thediscourse.We have shown that [Hir91]’s analysisof how scalarimplicaturescanarisethrough

the useof lexical itemsthat induceorderingscanbe directly appliedto the useof theseforms in

or asS-modifyingadverbials.We have furtherpostulatedtwo additionaltypesof implicaturesthat

canarisewhenthespeaker’s obeyanceof theCPandits maximsinteractswith therequirementthat

discoursebecoherent.

277

Page 295: Thesis about discourse

5.6 Other Contrib utions

5.6.1 DiscourseStructure

Many researchershavestudiedthepragmaticeffectsof discourseconnectivesondiscoursestructure

andinterpretation.As discussedin Chapter2, [GS86, Sch87] view many adverbials,andintonation,

ascuesof discoursestructure. For example,anywaymight signala return (pop) to a preceding

segment,while nowmight signaltheembedding(push)of a segment. As discussedin Chapter3,

viewing the interpretationsthata discourseunit makesavailableasabstractobjectsmayenableat

leastsomeof these“pragmatic”usesto beaccountedfor in thesemanticsof theDLTAG model.

5.6.2 Performatives

As notedin Section5.4, thereareothermeaningsassociatedwith utterancesthat have not been

discussedin this thesis;in particular, their illocutionary forceanda perlocutionaryeffect. [Sea69]

associatestheseadditionalmeaningsconveyedby anutterancewith, amongotherthings,intonation

and“performative verbs”, e.g. I promiseyou..., I baptizeyou...,etc. It may be that that the use

of adverbialscanhave a similar forceandeffect, which shouldeventuallybeincorporatedinto the

DLTAG model.

5.7 Conclusion

In this chapterwe have exploredtwo waysapartfrom their predicateargumentstructureandin-

terpretationthat adverbialscan be usedto contribute to discoursecoherence.Our intent was to

demonstratehow prosodicfocusandimplicaturecancauseanadverbialwhich is not normallyde-

pendenton thediscoursefor its interpretationto beinterpretedwith respectto thediscourse.While

our analysiswaspreliminary, it makesthe importantpoint thatadverbial semanticsis not theonly

factorinfluencingthe interpretationof adverbials,andshouldnot beconsideredin isolationasthe

only mechanismcausingadverbialsto creatediscoursecoherence.It is our expectationthatall of

thesefactorswill eventuallyhave to beincorporatedinto acomprehensive modelof discourse.

278

Page 296: Thesis about discourse

Chapter 6

Conclusion

6.1 Summary

The underlyingthemeof this thesisis that discourseis not a completelyseparatecategory from

syntaxandsemanticsandthatdiscourse-level coherencecanarisefrom thesamesubstrateasclause-

level coherence.

We have overviewedsimilaritiesanddifferencesbetweena varietyof theoriesof discourseco-

herence,which, taken together, distinguishdifferentmodulesrequiredto build a completeinter-

pretationof discourse.We discussedDLTAG ([FMP�

01, CFM�

02, WJSK03, WKJ99, WJSK99,

WJ98]) asa theorythatbridgesthegapbetweenclauseanddiscoursemodules,by usingthesame

syntacticandsemanticmechanismsthatbuild theclauseinterpretationto build anintermediatelevel

of discourseinterpretationon topof theclauseinterpretation.In DLTAG,cuephrases, or discourse

connectives, arepredicates,like verbs,exceptthey cantake interpretationsof clausesasarguments.

For coordinatingandsubordinatingconjunctions,bothargumentscomestructurally. For adverbial

cuephrases,which aremainly adverb (ADVP) andprepositional(PP)phrases,only oneargument

comesstructurally. Basedon considerationof computationaleconomyandbehavioral evidence,

DLTAG arguesthat theotherargumentof theseadverbialsmustberesolved anaphorically. How-

ever, while DLTAG proposesthat certainadverbialsfunctionasdiscourseconnectives,it doesnot

isolatethissubsetfrom thesetof all adverbials.

279

Page 297: Thesis about discourse

Becausethesetof adverbialsis compositional,andthereforeinfinite ([Kno96]), it is notpossible

to list all of the adverbialsthat function asdiscourseconnectives. We have thereforepresenteda

corpus-basedinvestigationof how semanticsandpragmaticscausecertainadverbialsto beclassified

asdiscourseconnectives,while otherconnectivesarenot. Our investigationhasshown thatin many

casesdiscourseconnectives are not an accidentalgroupingof ADVP and PP adverbials; rather,

their discoursepropertiesarisenaturally from their semantics.We have distinguisheddiscourse

adverbialsandclausaladverbialssemanticallyin termsof their predicateargumentstructureand

interpretation. We have arguedthat whetheror not ADVP and PP adverbialsfound in a corpus

are classifiedas discourseadverbialsdependson the interpretationof their semanticarguments.

We have shown thatdiscourseadverbialsarevery similar to discoursedeixis, in thatboth require

for their interpretationan abstractobjectmadeavailable in the prior discourseor spatio-temporal

context.

Our semanticanalysisis presentedindependentlyof any particularsemanticformalism,partly

becausehow informationis representedin a modeldepends,at leastto someextent,on whatcom-

putationalproblemsthe resultingmodelwill be usedto solve, andwhetherthe expenseof a par-

ticular representationproducesworthwhile results. Neverthelessour resultsmust be taken into

accountwhenbuilding a discoursemodel,for in their entiretythenumberof discourseadverbials

far overshadow the few discourseconnectives that have beenso far addressedin the literature.

We have thusoverviewed a variety of semanticformalismsanda variety of clauseanddiscourse

level syntax-semanticinterfaces. Drawing on this research,we presentedoneway in which the

predicate-argumentstructureandinterpretationof discourseadverbialscanbe incorporatedinto a

syntax-semanticinterfacefor DLTAG. We also discussedthe DLTAG annotationproject, whose

goalis to annotatetheargumentsof all discourseconnectivesin thePennTreebankcorpus.

It is notonly dueto theirargumentstructurethatadverbialsappearto requirediscoursefor their

interpretation,however. Wehave encountereda numberof adverbialsthathave beentreatedasdis-

courseconnectivesdespitethe fact that their discoursesemanticsalonedoesnot causethemto be

interpretedwith respectto abstractobjectinterpretationsin thediscourseor spatio-temporalcontext.

Wethushaveexploredotherexplanationsfor why suchadverbialscanrequirediscoursecontext for

280

Page 298: Thesis about discourse

their interpretation;in particular, thosethat involve the interactionof their semanticswith other

aspectsof discoursecoherence.We discussedprosodyasa semanticmechanismof discourseco-

herenceandshowedhow focuseffectsin bothclausalanddiscourseadverbialscontributediscourse

coherence.WediscussedGriceanimplicatureasanadditionalaspectof meaningthatarisesfrom the

assumptionof discoursecoherenceandshowedhow clausalanddiscourse(S-modifying)adverbials

canbeusedto convey implicatures.

In summary, we’veshown thatthediscoursesemanticsof adverbialscangoa longwaytowards

building completemodelof discourseinterpretation,but otheraspectsof discoursecoherencemust

alsobetakeninto account,andcorpusannotationandanalysisis alsorequired,to allow usto better

understandtheempiricalrealizationof discoursesyntaxandsemanticsandthecorrespondencesbe-

tweendiscourseconnectivesanddiscourserelations,suchaswhenandwhichdiscourseconnectives

areused,versuswhenandwhichdiscourserelationsareinferred.

6.2 Futur e Dir ections

Weendby identifying anumberof otherissuesfor futurestudythataresuggestedby theinvestiga-

tionsin this thesis.Eachof theseconstitutesabroaderuseof thesyntactic,semanticandpragmatic

functionsof adverbials.

Thefirst line of researchconcernstherepresentation,derivation,andresolutionof abstractob-

ject interpretations.Althoughasstatedabove, how informationis representedin a modeldepends,

at leastin part,on whatcomputationalproblemstheresultingmodelwill beusedto solve,because

DLTAG is built on topof aclause-level module,ideally thesemanticsemployedfor bothdiscourse

deicticreferenceto abstractobjectsandadverbialmodificationof abstractobjectswouldbesimilar.

However, the issueof how to representandderive abstractobject interpretationsis still an open

question.For example,asnotedin Chapter2, the fact that abstractobject interpretationsarenot

grammaticalizedasnounsprior to discoursedeixis reference,andthe fact that thereappearto be

structuralrestrictionsondiscoursedeicticreferenceto themhasleadsomeresearchersto arguethat

abstractobjectsarenotpresentasentitiesin thediscoursemodelprior to discoursedeixisreference.

Accordingto theseresearchers,their entity readingis addedto thediscoursemodelvia discourse

281

Page 299: Thesis about discourse

deixis reference. [DH95], for example,usetype coercionalongwith other computationaloper-

ationsto accessAO interpretations,while [Web91] usesreferring functions. In contrast,[Sto94]

and[Ash93] argue that AO interpretationsarealreadypresentin the discoursemodelbeforedis-

coursedeixisreferenceto them.[Sto94] usesapossibleworldssemanticsin whichdiscoursedeixis

refersto informationstates,while[Ash93] representseventsashiddenargumentsto verbsandfacts

andpropositionsaspiecesof semanticstructure,both of which canbe the referentsof discourse

deixis. We saw in Chapter3 that [Ern84, Moo93, Ver97, KP79] all make useof the (*j`_a. basic

abstractobjectsin exploring the semanticinterpretationof adverbial modification,althoughonly

[Moo93] formalizesthe representationof theseentities,treatingboth eventsand factsasentities

alreadypresentin thedomainof individuals,andusingpredicatelogic to representsbothashidden

argumentvariablesof theirassociatedpredicates.

Whatall of theseresearchersagreeon is thatAO interpretationsmustbederivablefrom non-NP

constituents.However, thiswork hasonly consideredthequestionin thecontext of discoursedeixis

or a few adverbials,andmoreover only considereda relatively smallnumberof AO interpretations.

In Chapter3 weintroducedadditionalcomplexity into theanalysis,namelythatthepossibleobjects

modifiedby adverbialscorrespondsto a muchwider rangeof objectsthanpreviously considered.

We neverthelessbelieve thatby extendingthe rangeof objectsto beconsidered,thecurrentstudy

may lead to a more comprehensive solution. It may, for example,be the casethat all AOs can

be subsumedwithin [Ash93]’s existing classification;in otherwords, it may be feasibleto treat

all AOs as sub-typesof events, fact-like objects,or proposition-like objects,as [Ash93, Ven67]

begin to do. If this is the case,we could treathiddenargumentsas “place-holders”for abstract

objects,andallow themto becoercedto specificdenotations,suchasconsequencesandreasonsthat

aredependenton anddeterminablefrom thepredicationon them. Theproblemof abstractobject

anaphoraresolutionhasalsobeenfarmorestudiedfor discoursedeixisthanfor discourseadverbials.

Fully understandingboththemechanismsthatdeterminewhichelementsareaccessibleto function

as antecedentsand how AO interpretationsare derived will however requirethe productionand

analysisof anannotatedcorpussuchasthatdescribedin Chapter4, andshould,moreover, consider

theanalysesalreadyproposedfor discoursedeixis.

282

Page 300: Thesis about discourse

Another openquestionconcernsthe incorporationof abstractobjectsinto CenteringTheory

([WJP81]). As discussedin Chapter2, CenteringTheorymodelsdiscourseprocessingfactorsthat

explaindifferencesin theperceivedcoherenceof discourses.Thecentralideais thateachutterance

presentsacenters, e.g.atopicentity; themostcoherentdiscoursesarethosein which,acrossaseries

of utterances,thecenterremainsconstant,andcanbereferredto by a pronoun.To date,however,

referenceto abstractobjectentitiesin CT hasonly beenconsideredin termsof discoursedeixis

reference(see[Eck98]). The effect of adverbial referenceto abstractobjectentitieson discourse

coherencehasnotbeenconsidered.It wouldbevery interestingto studyif andhow CT can“scale-

up” to thediscourselevel andincorporatenon-reifedclausalinterpretationsandrelationsbetween

themsuppliedby discourseconnectivesandinference.

Anotherline of researchconcernsthe extensionof the corpus-basedinvestigationsperformed

in this thesisto adverbials in other languages.For example, in English we were often able to

illustrate semanticargumentsby making them overt (e.g. in addition to this). According to a

Germanspeaker1, in German,similar semanticargumentsarenot optional;they areeitherexplicit

or lexically incorporatedinto theadverbial. For example,asshown in (6.1) theGermanequivalent

of as a result is infelicitous,but theequivalentof asa resultof that is fine, andeven betteris the

casewherethepreposition,demonstrative,andindefinitenounareincorporatedinto asinglelexical

item.

(6.1)

bad “als Folge” “asa result”

fine “als Folgedessen” “asa resultof that”

better “demzufolge” “dem”=“that”

“zu”=“as”

While we find a discussionin [Ale97] of somecross-linguisticpropertiesof adverbials,includ-

ing similaritiesin scope,positionandmeaning,it would bewell worthwhileto studythesemantic

equivalentsacrosslanguagesof the discourseadverbialsfound in our dataset, to seeif the same

semanticmechanismsareemployed to causethem to function as discourseconnectives. In this

way we would alsofurther enlarge the dataset,enablingthe developmentof a widely applicable

adverbialsemantics.1SIGDial reviewer, personalcommunication

283

Page 301: Thesis about discourse

Another importantline of researchconcernsthe practicalapplicationof the investigationsin

this thesisandof DLTAG in general.In Chapter3, for example,we discussedonenaturallanguage

generationsystem(SPUD,[SD97]). It wouldbeinterestingto seeif incorporatingdiscoursesyntax

andsemanticsinto suchsystemsimproves their efficiency andresults. The corpusannotationof

discoursesyntaxandsemantics,moreover, shouldleadto anaphoraresolutionalgorithmsfor the

anaphoricargumentsof discourseadverbials,andmay alsohelp improve question-answeringand

otherinformation-retrieval systems,which needto know what to look for, preferablywith a mini-

mumof humaninteraction.For example,if sucha systemrunsacrossasa resultin sometext, then

it shouldknow immediatelyto look for a cause, even if this object is not explicitly statedin the

samesentenceor prior sentences.

Finally, a significanttaskremainingis to investigatehow theDLTAG structuresandinterpreta-

tions built from thesyntaxandsemanticsof discourseconnectivescanbe incorporatedinto high-

level modulesof both discourseand dialogue. While DLTAG simplifies the constructionof an

intermediatelevel of discourse,the relationshipbetweenhow DLTAG discoursetreescombineto

makeentirediscoursesanddialogues,theconstraints,if any, thesestructuresplaceonanaphorares-

olution, andhow inferenceandotheraspectsof discoursecoherence,including focusandGricean

implicatureareto beincorporated,all remainvery interestingsubjectsfor futurestudy.

284

Page 302: Thesis about discourse

Bibliography

[AC74] D. Allen andA. Cruttenden.Englishsentenceadverbials:Their syntaxandtheir into-

nationin British English.Lingua, 34:1–30,1974.

[Ale97] A. Alexiadou. AdverbPlacement:A CaseStudyin AntisymmetricSyntax. JohnBen-

jaminsPublishingCompany, 1997.

[Ash93] N. Asher. Referenceto Abstract Objects. Kluwer, Dordrecht,1993.

[Aus61] J.Austin. PhilosophicalPapers. 1961.Reprintedin J.L. Austin,PhilosophicalPapers,

ed.by J.O. UrmsonandGeoffrey J.Warnock(Oxford,1990).

[Bea97] D. Beaver. Presupposition.In J. vanBenthemandA. ter Meulen,editors,Handbook

of Logic andLanguage, pages939–1007.Elsevier ScienceB.V., 1997.

[Bie01] G. Bierner. AlternativePhrases:Theoretical AnalysisandPracticalApplication. PhD

dissertation,Universityof Edinburgh,2001.

[Bik00] D. Bikel. A statisticalmodelfor parsingandword-sensedisambiguation,2000.

[Bla87] DianeBlakemore.SemanticConstraintson Relevance. Blackwell,Oxford,1987.

[Bos95] J.Bos. Predicatelogic unplugged.In P. Dekker andM. Stokhof,editors,Proceedings

of the10thAmsterdamColloguium, pages133–142.1995.

[BP83] J. BarwiseandJ. Perry. Situationsand Attitudes. The MIT Press,Cambridge,MA,

1983.

285

Page 303: Thesis about discourse

[BWFP87] SusanBrennan,Marilyn Walker-Friedman,andCarlPollard.A Centeringapproachto

pronouns.In Proceedingsof the25thAnnualMeetingof theAssociationfor Compu-

tational Linguistics, pages155–162,Stanford,CA, 1987.

[Byr00] D. Byron. Semanticallyenhancedpronouns.Proceedingsof theDiscourseAnaphora

andReferenceResolutionConference(DAARC2000), 12,2000.

[CFM�

02] C. Creswell,K. Forbes,E. Miltsakaki, R. Prasad,B. Webber, andA. Joshi. Thedis-

courseanaphoricpropertiesof connectives. Proceedingsof DAARC, 2002.

[CFM�

03] C. Creswell, K. Forbes, E. Miltsakaki, JasonTeeple, B. Webber, and A. Joshi.

Anaphoricargumentsof discourseconnectives: Semanticpropertiesof antecedents

versusnon-antecedents.Proceedingsof EACL, 2003.

[CFS97] A. Copestake, D. Flickinger, andI. Sag.Minimal RecursionSemantics.An Introduc-

tion. Manuscript,Stanford University, 1997.

[Cho71] N. Chomsky. Deep structure,surface structure,and semanticrepresentation. In

D. Steinberg andL. Jakobovitz, editors,Semantics, pages193–217.CambridgeUni-

versityPress,Cambridge,1971.

[Cho76] N. Chomsky. Conditionson rulesof grammar. Linguistic Analysis, pages303–351,

1976.

[CL93] N. Chomsky andH. Lasnik. The theoryof principlesandparameters.In J. Jacobs,

A. von Stechow, W. Sternefeld,andT. Venneman,editors,Syntax:An International

Handookof Contemporary Research, pages506–569.Mouton de Gruyter, Berlin,

1993.

[Coh84] R. Cohen. A computationaltheoryof the functionof cluewordsin argumentunder-

standing. Proceedingsof the 10th InternationalConferenceon ComputationalLin-

guistics, pages251–258,1984.

286

Page 304: Thesis about discourse

[CQ52] A. ChurchandW. Quine. Sometheoremson definabilityanddecidability. Journal of

SymbolicLogic, 17:179–187,1952.

[CW00] StephenClark andDavid Weir. A class-basedprobabilisticapproachto structuraldis-

ambiguation,2000.

[Dav67] D. Davidson.Thelogical form of actionsentences.In N. Rescher, editor, TheLogic of

DecisionandAction. Universityof Pittsburgh Press,Pittsburgh, PA, 1967. Reprinted

in D. Davidson,EssaysonActionsandEvents, Oxford: OxfordUniversityPress,1982,

pp.105-122.

[DH95] O. Dahl andC. Hellman. Whathappenswhenwe useananaphor. In Presentationat

theXVthScandinavianConferenceof Linguistics, Oslo,Norway, 1995.

[DiE89] B. DiEugenio. Clausalreferencein Italian. In Proceedingsof thePennsylvaniaLin-

guisticsColloquium, Philadelphia,PA, 1989.

[Dow79] D. Dowty. Word MeaningandMontagueGrammar. D. Reidel,Dordrecht,1979.

[Eck98] M. Eckert. DiscourseDeixisandNull Anaphora in German. PhDdissertation,Uni-

versityof Edinburgh,1998.

[EM90] MichaelElhadadandKathleenMcKeown. Generatingconnectives. In Proceedingsof

COLING. Helsinki,Finland,1990.

[Ern84] ThomasErnst. Toward an Integrated Theoryof AdverbPosition in English. PhD

dissertation,IndianaUniversity, 1984.

[ES99] M. Eckert andM. Strube.Resolvingdiscoursedeicticanaphorain dialogues.In Pro-

ceedingsof the9th Conferenceof theEuropeanChapterof theAssociationfor Com-

putationalLinguistics, Bergen,Norway, 1999.

[Fir64] JanFirbas. On definingthe themein functionalsentenceanalysis.TravauxLinguis-

tiquesdePrague, 1:229–236,1964.

287

Page 305: Thesis about discourse

[FM02] K. ForbesandE. Miltsakaki. Empiricalstudiesof centeringshiftsandcuephrasesas

embeddedsegmentboundarymarkers. University of PennsylvaniaWorking Papers in

Linguistics, 7(2),2002.

[FMP�

01] K. Forbes,E. Miltsakaki, R. Prasad,A. Sarkar, B. Webber, andA. Joshi. D-LTAG

system:Discourseparsingwith alexicalizedtree-adjoininggrammar. In ESSLLI’2001

Workshopon Information Structure, Discourse Structure and Discourse Semantics.

Helsinki,Finland,2001.

[Fox87] BarbaraFox. DiscourseStructure andAnaphora: writtenandconversationalEnglish.

CambrideUniversityPress,Cambridge,England,1987.

[Fra88] Bruce Fraser. Types of English discoursemarkers. Acta Linguistica Hungaria,

38(1):19–33,1988.

[FvG01] Anette Frank and Josefvan Genabith. Gluetag: Linear logic basedsemanticsfor

LTAG-andwhat it teachesusaboutLFG andLTAG. In Miriam Butt andTracy Hol-

loway King, editors,Proceedingsof theLFG01Conference. CSLI Publications,2001.

[FW02] K. ForbesandB. Webber. A semanticaccountof adverbialsasdiscourseconnectives.

Philadelphia,PA, 2002.

[Gar97a] C. Gardent. Interpretingfocus. PresentationSlides for ESSLI 1997 Conference,

http://www.coli.uni-sb.de/. claire/teaching/essll97/essli97-lect3.ps,1997.

[Gar97b] ClaireGardent.DiscourseTAG. Clausreportnr.89,Universityof theSaarland,Saar-

brucken,1997.

[Gaw86] J. Gawron. Situationsandprepositions.Linguisticsand Philosophy, 9(3):327–382,

1986.

[Gaz79a] G. Gazdar. Pragmatics: Implicature, Presuppositionand Logical Form. Academic

Press,New York, 1979.

288

Page 306: Thesis about discourse

[Gaz79b] G. Gazdar. A solutionto theprojectionproblem. In Oh andDineen,editors,Syntax

andSemantics11: Presupposition, pages57–89.AcademicPress,New York, 1979.

[Gaz99] GeraldGazdar. http://www.cogs.susx.ac.uk/lab/nlp/gazdar/teach/nlp/nlp.html, 1999.

[GC00] S. Gustafson-Capova. The influenceof prosodicprominenceon the interpretationof

ambiguousanaphorsin Swedish.Proceedingsof theDiscourseAnaphora andRefer-

enceResolutionConference(DAARC2000), 12,2000.

[GHZ93] J. Gundel,N. Hedberg, andR. Zacharski.Cognitive statusandthe form of referring

expressionsin discourse.Language, 69,1993.

[gra] http://www.edufind.com/english/grammar/DETERMINERS1.cfm.

[Gre69] S.Greenbaum.Studiesin EnglishAdverbialsUsage. Longmans,London,1969.

[Gri75] H. P. Grice. Logic andconversation. In P. Cole andJ. Morgan,editors,Syntaxand

Semantics,vol. 3, pages41–58.AcademicPress,1975.

[Gri89] H. P. Grice. Studies in the Way of Words. Harvard University Press,

Cambridge, MA, 1989. A summary of Grice’s work can be found at:

http://www.artsci.wustl.edu/. philos/MindDict/grice.html.

[Gro99] The XTAG ResearchGroup. A Lexicalized Tree Adjoining Grammarfor English.

http://www.cis.upenn.edu/. xtag,1999.

[GS86] B. GroszandC. Sidner. Attention,intentionandthestructureof discourse.Journalof

ComputationalLinguistics, 12:175–204,1986.

[Hal67] M. Halliday. Noteson transitivity andthemein English. Journal of Linguistics, 3,

1967.

[Har99] D. Hardt. Dynamicinterpretationof verbphraseellipsis. LinguisticsandPhilisophy,

22:187–221,1999.

289

Page 307: Thesis about discourse

[Hei82] IreneHeimt. Semanticsof Definiteand IndefiniteNounPhrases. PhD dissertation,

Universityof Massachusetts,1982.

[HH76] M. HallidayandR. Hasan.Cohesionin English. Longman,London,1976.

[Hir91] JuliaHirschberg. A Theoryof ScalarImplicature. GarlandPublishingCompany, New

York,1991.Publishedversionof Ph.D.Dissertation,Universityof Pennsylvania,1985.

[HK98] IreneHeimandAngelaKratzer. Semanticsin GenerativeGrammar. Blackwell,1998.

[Hob85] JerryHobbs.Ontologicalpromiscuity. In Proceedingsof the23<�� AnnualMeetingof

theAssociationfor ComputationalLinguistics, pages61–69.PaloAlto, CA, 1985.

[Hob90] JerryHobbs.Literatureandcognition.CSLILecture Notes, 21,1990.

[Hor69] L. Horn. A presuppositionaltheoryof ‘only’ and‘even’. CLS5, Chicago Linguistics

Society, 1969.

[Hor96] L. Horn. Presuppositionand implicature. In S. Lappin, editor, The Handbookof

Contemporary SemanticTheory, pages299–319.Blackwell,Oxford,1996.

[Hov90] EduardHovy. Parsimoniousandprolifigateapproachesto the questionof discourse

structurerelations. In Proceedingsof the Fifth International Workshopon Natural

Language Generation, pages128–136.1990.

[Hov93] EduardHovy. Automateddiscoursegenerationusing discoursestructurerelations.

Artificial Intelligence, 63:69–142,1993.

[Hov95] EduardHovy. The multifunctionality of discoursemarkers. In Proceedingsof the

Workshopon DiscourseMarkers. Holland,January, 1995.

[HSAM93] JerryHobbs,Mark Stickel, DouglasAppelt,andPaulMartin. Interpretationasabduc-

tion. Artificial Intelligence, pages69–142,1993.

[Hum48] David Hume.An Inquiry ConcerningHumanUnderstanding. TheLiberal Arts Press,

New York, 1955edition,1748.

290

Page 308: Thesis about discourse

[Jac72] R. Jackendoff. SemanticInterpretation in Generative Grammar. MIT Press,Cam-

bridge,MA, 1972.

[Jac90] R. Jackendoff. Semanticstructures. In Current Studiesin LinguisticsSeries. Cam-

bridge,MA, 1990.

[JKR03] Aravind Joshi,LauraKallmeyer, andMaribelRomero.Flexible compositionin LTAG:

Quantifierscopeandinverselinking. In Proceedingsof theInternationalWorkshopon

CompositionalSemantics. Tilburg, TheNetherlands,2003.

[Joh95] MichaelJohnston.dxV;.�W -clauses,adverbsof quantification,andfocus.In Proceedings

of WCCFL13,Stanford LinguisticsAssociation, CLSI, StanfordUniversity, 1995.

[Jos87] Aravind Joshi.An introducationto treeadjoininggrammar. In Alexis Manaster-Ramer,

editor, Mathematicsof Language, pages87–114.JohnBenjamins,Amsterdam,1987.

[JR98] J. Jayezand R. Rossari. Discourserelationsversusdiscoursemarker relations. In

Proceedingsof the ACL Workshopon Discourse Relationsand Discourse Markers,

pages72–78.Montreal,Canada,1998.

[JVS99] A. Joshi and K. Vijay-Shanker. Compositionalsemanticswith lexicalized tree-

adjoininggrammar(LTAG): How muchunderspecificationis necessary?In Alexis

Manaster-Ramer, editor, Proceedingsof theThird InternationalWorkshopon Compu-

tational Semantics. Tilburg, Netherlands,January, 1999.

[Kal02] LauraKallmeyer. UsinganenrichedTAG derivationstructureasbasisfor semantics.

In Proceedingsof theSixthInternationalWorkshopon TreeAdjoiningGrammarand

RelatedFrameworks(TAG+6), pages101–110.Universityof Venice,2002.

[Kam79] H. Kamp. Events, instants,and temporalreference. In R. Baeuerleet al., editor,

SemanticsfromDifferentPointsof View. SpringerVerlag,Berlin, 1979.

[Kar73] L. Karttunen.Presuppositionsof compoundsentences.Linguistic Inquiry, 4(2):169–

193,1973.

291

Page 309: Thesis about discourse

[Kas93] R.Kasper. Adjunctsin themittelfeld. In J.Nerbonne,K. Netter, andC.Pollard,editors,

Germanin Head-DrivenPhraseStructure Grammar, Lecture NoteSeries,Chapter4,

pages39–69.Utrecht,TheNetherlands,1993.

[Keh95] Andrew Kehler. Interpreting CohesiveFormsin the Context of DiscourseInference.

PhDdissertation,HarvardUniversity, 1995.

[KJ99] LauraKallmeyer andAravind Joshi. Factoringpredicateargumentandscopeseman-

tics: Underspecifiedsemanticswith LTAG. In Paul Dekker, editor, Proceedingsof the

12thAmsterdamColloquium, pages169–174.Amsterdam,December, 1999.To appear

in theJournalof LanguageandComputation,2002.

[KK70] P. Kiparsky andC.Kiparsky. Fact. In M. BierwischandK. Heidolph,editors,Progress

in Linguistics, pages143–13.Mouton,TheHague,1970. Also in (PetofiandFranck,

1973).

[KKR91] R. Kittredge,T. Korelsky, andO. Rambow. On theneedfor domaincommunication

knowledge.ComputationalIntelligence, 7:305–314,1991.

[KKW01a] I. Kruijf f-Korbayova andB. Webber. Concession,implicature,and alternative sets.

In Proceedingsof theInternationalWorkshoponComputationalSemantics(IWCS-4),

Tilburg, January, 2001.

[KKW01b] I. Kruijf f-Korbayova andB. Webber. Informationstructureandthe interpretationof

‘otherwise’. In ESSLLI2001Worshopon InformationStructure, DiscourseStructure,

andDiscourseSemantics, pages61–78,Helsinki,Finland,2001.

[Kno96] Ali Knott. A Data-DrivenMethodology For MotivatingA Setof CoherenceRelations.

PhDdissertation,Universityof Edinburgh,1996.

[Kon91] EkkehardKonig. TheMeaningof FocusParticles: A ComparativePerspective. Rout-

ledge,London,1991.

292

Page 310: Thesis about discourse

[KOOM01] A. Knott, J. Oberlander, M. O’Donnell, andC. Mellish. Beyond Elaboration: The

interactionof relationsandfocusin coherenttext. In J. SchilperoordT. Sandersand

W. Spooren,editors,Text representation:linguisticandpsycholinguisticaspects, pages

181–196.Benjamins,2001.

[KP79] L. Karttunenand S. Peters. Conventional implicature. In C. Oh and D. Dinneen,

editors,Syntaxand Semantics:Presupposition,volume11, pages1–56.Academic

Press,New York, 1979.

[KP02] Paul Kingsbury andMarthaPalmer. FromTreebankto Propbank.In Proceedingsof

theThird InternationalConferenceonLanguageResourcesandEvaluation,LREC-02.

LasPalmas,CanaryIslands,Spain,May 28-June3, 2002.

[KR93] H. KampandU. Reyle. FromDiscourseto Logic. Kluwer AcademicPublishers,1993.

[Kra89] A. Kratzer. An investigationof thelumpsof thought.LinguisticsandPhilosophy, 12,

1989.

[Kri] M. Krif a. Focus.http://cognet.mit.edu/MITECS/Entry/krifka2.

[Kri92] M. Krifka. A compositionalsemanticsfor multiple focusconstructions.CornellWork-

ing Papers in Linguistics, 10:127–158,1992.

[LA90] D. Litman andJ. Allen. Discourseprocessingandcommonsenseplans. In P. Cohen,

J.Morgan,andM. Pollack,editors,Intentionsin Communication. MIT Press,1990.

[LA93] A. LascaridesandN. Asher. Temporalinterpretation,discourserelationsandcommon-

senseentailment.LinguisticsandPhilosophy, 16:437–493,1993.

[LA99] A. LascaridesandN. Asher. Cognitive states,discoursestructure,andthecontentof

dialogue.In Proceedingsto Amstelogue. 1999.

[Lad66] R.Ladd.IntonationalPhonology. CambridgeUniversityPress,Cambridge,MA, 1966.

[Lag98] Luuk Lagerwerf. Causal ConnectivesHave Presuppositions. Holland Academic

Graphics,TheHague,1998.

293

Page 311: Thesis about discourse

[Lak74] R. Lakoff. Remarksonthisandthat. In Papers fromtheTenthRegionalMeetingof the

Chicago LinguisticSociety. 1974.

[Lit98] K. Litkowski. Analysisof subordinatingconjunctions.CL ResearchTechnicalReport

(Draft) 98-01,CL Research,Gaithersburg, MD, 1998. http://www.clres.com/online-

papers/sc.html.

[LO93] A. Lascaridesand J. Oberlander. Temporalconnectives in a discoursecontext. In

Proceedingsof the 6th InternationalConferenceof the EuropeanChapterof the As-

sociationfor ComputationalLinguistics, pages260–268.Utrecht, The Netherlands,

1993.

[Lon83] RobertLongacre.TheGrammarof Discourse. PlenumPress,New York, 1983.

[Lyo77] J.Lyons.Semantics. CambridgeUniversityPress,Cambridge,MA, 1977.

[Mar92] J.R.Martin. EnglishText: SystemandStructure. Benjamin,Amsterdam,1992.

[Mar97] Daniel Marcu. Instructionsfor manuallyannotatingthe discoursestructureof texts.

http://www.isi.edu/. marcu/software.html, 1997.

[Mar99] DanielMarcu.A formalandcomputationalsynthesisof GroszandSidner’sandMann

and Thompson’s theories. In Workshopon Levels of Representationin Discourse.

Edinburgh,Scotland,1999.

[Mar00] DanielMarcu. Therhetoricalparsingof unrestrictedtexts: A surface-basedapproach.

ComputationalLinguistics, 25(3):395–448,2000.

[McC88] J.D. McCawley. TheSyntacticPhenomenaof English. The University of Chicago

Press,1988.

[McL01] Mark McLauchlan. Maximum entropy modelsandprepositionalphraseambiguity,

2001.

[MG82] Sally McConell-Ginet.Adverbsandlogical form. Language, 58:144–184,1982.

294

Page 312: Thesis about discourse

[Mit82] Anita Mittwoch. On the differencebetweeneatingandeatingsomething:Activities

versusaccomplishments.LinguisticInquiry, 13,1982.

[Mod01] N. Modjeska. Towardsa resolutionof comparative anaphora:A corpusstudy of

‘other’. In PAPACOL. Italy, 2001.

[Mon74] R. Montague.On thenatureof certainphilosophicalentities.In R. Thomason,editor,

Formal Philosophy, pages148–187.YaleUniversityPress,New Haven,CT, 1974.

[Moo93] R. Moore. Events,situationsandadverbs. In R. WeischedelandM. Bates,editors,

Challengesin Natural LanguageProcessing. CambridgeUniversityPress,Cambridge,

1993.

[MP92] JohannaMoore andMarthaPollack. A problemfor RST: The needfor multi-level

discourseanalysis.ComputationalLinguistics, 18(4):537–544,1992.

[MP93] JohannaMooreandCecileParis. Planningtext for advisorydialogues:Capturingin-

tentionalandrhetoricalinformation.ComputationalLinguistics, 19(4):651–694,1993.

[MS88] MarcMoensandMark Steedman.Temporalontologyandtemporalreference.Journal

of ComputationalLinguistics, 14(2):15–28,1988.

[MT88] William Mann andSandraThompson.RhetoricalStructureTheory: Towarda func-

tional theoryof text organization.Text, 8(3):243–281,1988.

[MTC95] Gail Mauner, Michael Tanenhaus,andGreg Carlson. Implicit argumentsin sentence

processing.Journalof MemoryandLanguage, 34:357–382,1995.

[MW] OnlineMiriam-WebsterDictionary, http://www.m-w.com/home.htm.

[Nun79] G. Nunberg. The non-uniquenessof semanticsolutions:Polysemy. Linguisticsand

Philosophy, 3, 1979.

[ODA93] W. O’Grady, M. Dobrovolsky, and M. Aronoff. Linguistics: An Introduction. St.

Martin’s Press,New York, 1993.

295

Page 313: Thesis about discourse

[Par84] BarbaraPartee.Nominalandtemporalanaphora.LinguisticsandPhilosophy, 7:243–

286,1984.

[Pas91] R. Passonneau.Somefactsaboutcenters,indexicals,anddemonstratives.Proceedings

of the29thAnnualMeetingof theACL, 1991.

[Per90] FernandoPereira. Categorial semanticsand scoping. ComputationalLinguistics,

16(1):1–10,1990.

[PH90] J. PierrehumbertandJ. Hirschberg. Themeaningof intonationcontoursin the inter-

pretationof discourse.In P. Cohen,J. Morgan,andM. Pollack,editors,Intentionsin

Communication, pages271–312.MIT Press,Cambridge,MA, 1990.

[Pol96] Livia Polanyi. TheLinguisticStructureof Discourse. CSLI TechnicalReport,Stanford

CA, 1996.

[Pri81] E. Prince. Towarda taxonomyof given/new information. In P. Cole,editor, Radical

pragmatics. AcademicPress,NY, 1981.

[Pri86] E. Prince. On thesyntacticmarkingof presupposedopenpropositions.In A. Farley,

P. Farley, andK. McCullough, editors,Papers from the Parasessionon Pragmatics

andGrammaticalTheory, 22ndRegionalMeeting, pages208–222.ChicagoLinguistic

Society, 1986.

[PS87] C. PollardandI. Sag. Informationbasedsyntaxandsemantics.In Volume1: Funda-

mentals. Stanford,CA: Centerfor theStudyof LanguageandInformation,1987.

[PSvdB94] H. Prust,R.Scha,andM. vandenBerg. Discoursegrammarandverbphraseanaphora.

LinguisticsandPhilosophy, 17:261–327,1994.

[PT] PennTreebank.Seehttp://www.ldc.upenn.edu/ldc/online/treebank/ for documentation.

[Pul97] S. Pulman. Higher orderunificationandthe interpretationof focus. Linguisticsand

Philosophy, 20:73–115,1997.

296

Page 314: Thesis about discourse

[PvdB96] Livia Polanyi andMartin van denBerg. Discoursestructureanddiscourseinterpre-

tation. In P. Dekker andM. Stokhof, editors,Proceedingsof the TenthAmsterdam

Colloquium. ILLC, Amsterdam,1996.

[PvdB99] Livia Polanyi and Martin van den Berg. Logical structureand discourseanaphora

resolution.In D. Cristea,N. Ide,andD. Marcu,editors,Proceedingsof theWorkshop

on therelationof Discourse/DialogueStructure andReference. 37thAnnualMeeting

of theAssociationof ComputationalLinguistics,1999.

[QGLS85] R. Quirk, S.Greenbaum,G. Leech,andJ.Svartvik. A ComprehensiveGrammarof the

EnglishLanguage. Longman,London,1985.

[Qui72] R. Quirk. A Grammarof Contemporary English. Longman,London,1972.

[Red90] GiselaRedeker. Ideationalandpragmaticmarkersof discoursestructure.Journal of

Pragmatics, 14:367–381,1990.

[Roo85] M. Rooth. Associationwith Focus. PhD dissertation,University of Massachusetts,

Amherst,1985.

[Roo92] M. Rooth. A theoryof focusinterpretation.Natural Language Semantics, 1:75–116,

1992.

[Roo95a] M. Rooth. Focus.In S.Lappin,editor, Handbookof Contemporary SemanticTheory,

pages271–298.Blackwell,London,1995.

[Roo95b] M. Rooth. Indefinites,adverbsof quantification,andfocussemantics.In G. Carlson,

editor, Generics. 1995.

[Sac77] E. Sacerdoti.A Structure for PlansandBehavior. Elsevier, Amsterdam,1977.

[Sae96] K. Saeboe.Anaphoricpresuppositionsandzeroanaphora.LinguisticsandPhilosophy,

19:187–209,1996.

297

Page 315: Thesis about discourse

[SBDP00] Matthew Stone,ToniaBleam,ChristineDoran,andMarthaPalmer. Lexicalizedgram-

mar andthe descriptionof motion events. In WorkshopTAG/ 5, Paris, 25-27May,

2000.

[Sch71] P. Schreiber. Someconstraintsontheformationof Englishsentenceadverbs.Linguistic

Inquiry, 2(1),1971.

[Sch85] R. Schiffman. DiscourseConstraints on ‘it’ and ‘that’: A studyof language usein

career-counselinginterviews. PhDdissertation,Universityof Chicago,1985.

[Sch87] D. Schiffrin. DiscourseMarkers. CambridgeUniversityPress,Cambridge,MA, 1987.

[Sch97] FrankSchilder. Treediscoursegrammar, or how to get attachedto a discourse. In

P. Dekker andM. Stokhof,editors,Proceedingsof theSecondInternationalWorkshop

on ComputationalSemantics. Tilburg, Netherlands,1997.

[SCT�

94] J. Sedivy, G. Carlson,M. Tanenhaus,M. Spivey-Knowlton, andK. Eberhard. The

cognitive functionof contrastsetsin processingfocusconstructions.In P. Boschand

R. vanderSandt,editors,FocusandNatural Language Processing. IBM Deutschland

InformationssystemeGmbH,Institutefor Logic andLinguistics,1994.

[SD97] Matthew Stoneand ChristineDoran. Sentenceplanningas descriptionusing tree-

adjoininggrammar. Proceedingsof the 35th AnnualMeetingof the Associationfor

ComputationalLinguistics(ACL), pages198–205,1997.

[SdS90] D. R. ScottandC. S. de Souza. Gettingthe messageacrossin RST-basedtext gen-

eration. In R. Dale, C. Mellish, andM. Zock, editors,Current Research in Natural

Language Generation. AcademicPress,1990.

[Sea69] J.Searle.Speech Acts:AnEssayin thePhilosophyof Language. CambridgeUniversity

Press,Cambridge,MA, 1969.

[Shi86] StuartShieber. An Introductionto Unifcation-BasedApproachesto Grammar. CSLI,

Stanford,CA, 1986.

298

Page 316: Thesis about discourse

[SHP73] P. Sgall,E. Hajicova,andE. Penesova. Topic,FocusandGenerativeSemantics. Scrip-

tor, Kronberg andTaunus,1973.

[Sib92] P. Sibun. Generatingtext without trees. ComputationalIntelligence, 8(1):102–122,

1992.

[SIG02] Proceedingsof the Third SIGdial Workshopon DiscourseandDialogue,11-12July,

Philadelphia,PA, 2002.

[Sil76] Michael Silverstein. Shifters, linguistic categories and cultural descriptions. In

K. BassoandH. Selby, editors,Meaningin Anthropology, pages11–55.University

of New Mexico Press,1976.

[Sim00] MandySimons.Why somepresuppositionsareconversationalimplicatures.Handout

for University of Pennsylvania LinguisticsSpeaker SeriesTalk, November9, 2000.

Dept.of Philosophy, Carnegie Mellon University.

[SP88] R. SchaandL. Polanyi. An augmentedcontext freegrammarfor disocurse.In Pro-

ceedingsof the 12th International Conferenceon ComputationalLinguistics, pages

22–27.COLING, 1988.

[SSN93] TedSanders,Wilbert Spooren,andLeoNoordman.Coherencerelationsin acognitive

theoryof discourserepresentation.CognitiveLinguistics, 4(2):93–133,1993.

[Sta74] R.Stalnaker. Pragmaticpresuppositions.In M. MunitzandP. Unger, editors,Semantics

andPhilosophy, pages197–214.New York UniversityPress,1974.

[Ste96] M. Steedman.SurfaceStructure andInterpretation. TheMIT Press,Cambridge,MA,

1996.

[Ste00a] M. Steedman.Informationstructureandthe syntax-phonologyinterface. Linguistic

Inquiry, 31(4):649–689,2000.

[Ste00b] M. Steedman.Theproductionsof time: temporalityandcausalityin linguistic seman-

tics. Draft, 2000.

299

Page 317: Thesis about discourse

[Ste00c] M. Steedman.TheSyntacticProcess. TheMIT Press,Cambridge,MA, 2000.

[Sto94] M. Stone. Discoursedeixis,discoursestructure,andthesemanticsof subordination.

Ms.,1994.

[Str59] P. Strawson.On referring.Mind, 235:320–344,1959.

[Suz97] H. Suzuki.Noteson thegrammarof -ly adverbsin English.Typological Investigation

of LanguagesandCulturesof theEastandWest, 1997.

[SW86] D. SperberandD. Wilson. Relevance. Harvard University Press,Cambridge,MA,

1986.

[Swa88] T. Swan. SentenceAdverbialsin English: A Synchronic andDiachronic Investigation.

Novus,Tromso-Studieri SprakvitenskapX, Oslo,1988.

[Umb02] CarlaUmbach.Contrastandcontrastive topic. In IvanaKruijf f-Korbayova andMark

Steedman,editors,ESSLLI2001WorkshoponInformationStructure, DiscourseStruc-

ture andDiscourseSemantics, 2002.

[vD79] T. vanDijk. Pragmaticconnectives. Journal of Pragmatics, 3:447–456,1979.

[vdB96] Martin van den Berg. Discoursegrammarand dynamic logic. In P. Dekker and

M. Stokhof,editors,Proceedingsof the TenthAmsterdamColloquium. ILLC, Ams-

terdam,1996.

[vdG69] G. von derGabelentz.IdeenzueinervergleichendenSyntax. Wort undSatzstellung,

Zeitschriftfur VolkerpsychologieundSprachwissenschaft6, 1869.

[vdS92] R. vanderSandt.Presuppositionprojectionasanaphoraresolution.Journalof Seman-

tics, 9:333–377,1992.

[Ven67] Z. Vendler. Linguisticsin Philosophy. CornellUniversityPress,Cornell,NY, 1967.

[Ver97] CorneliaMaria Verspoor. Contextually-Dependent Lexical Semantics. PhD disserta-

tion, Universityof Edinburgh,1997.

300

Page 318: Thesis about discourse

[vH00] K. vonHeusinger. Accessibility, discourseanaphora,anddescriptivecontent.Proceed-

ingsof theDiscourseAnaphora andReferenceResolutionConference(DAARC2000),

12,2000.

[vS82] A. vonStechow. Structuredpropositions.Arbeitspapier59desSonderforschungsbere-

ichs99.UniversitatKonstanz,1982.

[Wal93] Marilyn Walker. InformationalRedundancyandResourceBoundsin Dialogue. PhD

dissertation,Universityof Pennsylvania,1993.

[web] http://cctc.commnet.edu/grammar/adverbs.htm,

http://www.cis.upenn.edu/. xtag/tech-report/node173.html,

http://faculty.washington.edu/ . marynell/grammar/AdverbPl.html.

[Web88] B. Webber. Discoursedeixis: Referenceto discoursesegments. Proceedingsof the

26thAnnualMeetingof theAssociationfor ComputationalLinguistics, 1988.

[Web91] B. Webber. Structureandostensionin theinterpretationof discoursedeixis.Language

andCognitiveProcesses, 6(2),1991.

[WJ98] B. WebberandA. Joshi.Anchoringa lexicalizedtree-adjoinggrammarfor discourse.

In Proceedingsof the Coling/ACL Workshopon Discourse Relationsand Discourse

Markers, pages86–92.Montreal,Canada,1998.

[WJP81] M. Walker, A. Joshi,andE. Prince. Centeringin naturallyoccurringdiscourse:An

overview. In M. Walker, A. Joshi,andE. Prince,editors,CenteringTheoryin Dis-

course. ClarendonPress,Oxford,1981.

[WJSK99] B. Webber, A. Joshi,M. Stone,andA. Knott. Discourserelations:A structuraland

presuppositionalaccountusing lexicalisedTAG. In Proceedingsof the 36th Annual

Meetingof theAssociationfor ComputationalLinguistics, pages41–48.CollegePark,

MD, 1999.

301

Page 319: Thesis about discourse

[WJSK03] Bonnie Webber, Aravind Joshi,Matthew Stone,and Alistair Knott. Anaphoraand

discoursesemantics.To appearin ComputationalLinguistics, 2003.

[WKJ99] B. Webber, A. Knott, andA. Joshi. Multiple discourseconnectives in a lexicalised

grammarfor discourse.In Proceedingsof the3rd InternationalWorkshopon Compu-

tational Semantics, pages309–325.Tilburg, TheNetherlands,1999.

[WN98] FivePapersonWordNet,1998.ftp://ftp.cogsci.princeton.edu/pub/wordnet/5papers.ps.

302