combinatory categorial grammar and linguistic...

Combinatory Categorial Grammar

and

Linguistic Diversity

ESSLLI 2005, August 8–12, Edinburgh

Cem Bozsahin ∗ & Mark McConville †

∗Middle East Technical University, Ankara†The University of Edinburgh

Outline

CCG as a theory: Substantive aspects (CB)

Formal aspects (MM)

Limits on NL variation (MM)

Constructions (CB)

Linguistic Diversity: Ergative and accusative languages (CB)

(and OpenCCG implementations)

Morphology and organisation of the lexicon (MM)

Substantive aspects:

Syntactic types, semantic types

and principles

Cem Bozsahin

1

Categorial Grammar (CG) is unique in its treatment of the notion of “pos-sible category”, and how syntax can be made completely type-dependent,rather than structure-dependent, with transparent semantics.

Combinatory Categorial Grammar (CCG) is unique among categorial gram-mars in its treatment of constituency, (un)bounded dependencies, and bind-ing.

There have been no syntactic variables in CCG ever since its inception (Adesand Steedman, 1982, written 1979), and no other grammar than a lexi-calised grammar, therefore there can be no locus for movement-like oper-ations to arise.

With the Minimalist Program’s (MP) recent use of one binary combina-tor (Merge), and attempts to subsume Move under Merge (Epstein et al.,1998), theories seem to converge, differences narrowing down to the no-tion of “possible category.” (see Steedman 2005a for more discussion).

2

.. ’units’ and ’grammatical facts’ are only different names for dif-ferent aspects of the same general fact: the operation of linguisticoppositions. So much so that it would be perfectly possible totackle the problem of units by beginning with grammatical facts.(F. de Saussure , Cours de Linguistique Générale, 1916:168).

Four criteria for tests for grammatical constituents and constituentboundaries: (Noam Chomsky , LSLT, 1956/1979:210)

a. The rule for conjunction

b. Intrusion of parenthetical expressions

c. Ability to enter transformations

d. Certain intonational features

3

• Constituency is something to be accounted for and explained,

• rather than defined by a theory.

• Only (c) above is a theory-specific definition of constituency; othersare empirical criteria.

to-infinitive n. A conventional label for an infinitival verb phrase preceded by the for-

mative to, as in Lisa wants to buy a BMW. In traditional grammar, such a sequence as

to buy was regarded as a single form, the so-called ‘infinitive’ of the verb buy, but this

analysis is rejected by all contemporary theories of grammar: all possible tests point to

the conclusion that the sequence buy a BMW is a constituent (a verb phrase), and

hence to the conclusion that to buy is not a constituent of any kind. (R.L. Trask,

Dictionary of Grammatical Terms in Linguistics, 1993).

4

Mary wants to read and discuss the stories of Edgar Allan Poe

Mary wants to read and to discuss the stories of Edgar Allan Poe

Mary wants to try to read and to be able to discuss the stories of Edgar Allan Poe

Mary wants—I think—to read the stories of Edgar Allan Poe

Mary wants to discuss—or read?can’t remember—the stories of Edgar Allan Poe

Mary wants—I think—to discuss—or read? can’t remember—the stories of EdgarAllan Poe

5

Clearly we may say that if presentations, expressible thoughtsof any sort whatever, are to have their faithful reflections in thesphere of meaning-intentions, then there must be a semantic formwhich corresponds to each presentational form. [...] And if the ver-bal resources of language are to be a faithful mirror of all mean-ings possible a priori, then language must have grammatical formsat its disposal which give distinct expression, i.e. sensibly distinctsymbolisation, to all distinguishable meaning-forms. (EdmundHusserl , Logical Investigations, 1890)

The lexicon of a given language is a finite subset of the set of allcategories, subject to quite narrow restrictions that ultimately stemfrom limitations on the variety of semantic types with which thesyntactic categories are paired in the lexicon. Mark Steedman ,The Syntactic Process, 2000:32

6

• Category : a label (symbol, feature bundle etc.) for the purpose ofcapturing a linguistic distinction.

• If they were to be lexicalised, there should be countably infinitely manysyntactic categories.

• Every syntactic category determines a set of structured meanings (log-ical form—LF), because it is made of substantive categories (nouns,verbs etc.).

• Every LF (and some language specific settings, such as directionality)determine a set of syntactic categories.

7

man := N: man′

the := (S/(S\NP))/N: λpλq.the′(px)(qx)

some := (S/(S\NP))/N: λpλq.exists′(px)(qx)

sleeps := S\NP: λx.sleep′ x

hits := (S\NP)/NP: λxλy.hit′ xy

reads := (S\NP)/NP: λxλy.read′ xy

that := Scomp/Sfin: λx.x

that := (N\N)/(S\NP): λpλq.and′(px)(qx)

that := (N\N)/(S/NP): λpλq.and′(px)(qx)

NP↑ = S/(S\NP)

IV = S\NP TV = IV/NP

GQ(g) = λpλq.g(px)(qx)

iv(p) = λx.px tv(p) = λxλy.pxy

some := NP↑/N: GQ(exists′)

sleeps := IV: iv(sleep′)

hits := TV: tv(hit′)

reads := TV: tv(read′)

that := (N\N)/IV: GQ(and′)

that := (N\N)/(S/NP): GQ(and′)

Systematicity is not only vertical (cross-categorial), but horizontal as well:categories and semantic types correspond in non arbitrary ways.

8

The lexicon is really an appendix of the grammar, a list of basic ir-regularities. This is all the more evident if meanings are taken intoconsideration, since the meaning of each morpheme belongs to itby an arbitrary tradition. (L. Bloomfield , Language, 1933:274)

• Modern incarnation of grammar-lexicon dichotomy appears to origi-nate from Bloomfield—and adopted by Chomsky (1995:130)

“We distinguish the lexicon from the computational system of a lan-guage, the syntax in a broad sense (including phonology).”

• For Chomsky, any systematicity in the lexicon is a missed generalisa-tion, i.e. it belongs to grammar, cf. (1965:167) through (1995:130).

9

Even when lexicon is considered to be a part of the computational system,the dichotomy prevails (Reinhart and Siloni, 2004, p.160):

“Unlike approaches that decrease the role of the lexicon from anoperative component to a list of items (for example, Borer, in thisvolume; Embick, in this volume; Marantz 1997), we assume thatthe lexicon is a computational component, where derivational op-erations can apply.”

“Further, we attribute the somewhat different nature of reflexiveverbs in Hebrew, Dutch, and English vs. Romance to the distinctcomponent of grammar in which the operation applies: lexicon vs.syntax.”

10

• CCG argues that this dichotomy gets in the way of our understandingof how syntax (in narrow—combinatory—sense) can shape possiblehuman lexicons.

• i.e. The substantive categories of the lexicon are instantiations of theformal categories of Universal Grammar (UG), therefore represent acontinuum.

• Languages differ only in their lexicons (called Radical Lexicalism byKarttunen 1989).∗

• By implication, any combinatory difference must be lexically specifi-able.

∗We might call grammar-lexicon ‘grammaticon ’ to avoid the orthographic dichotomy aswell.

11

Lexicalising a Surface Grammar:

A Game of Algebra

Phrase Structure Grammars and corresponding machinery as models ofsurface syntax encode constituency (and sometimes, order), but some in-formation will be redundantly specified:

12

S → NP VP

VP → Viv

VP → Vtv NP

Viv → slept e 7→ t

Vtv → read e 7→ (e 7→ t)

read ’s subcategorisation for an object NP is defined twice: in its lexicalcategory (Vtv) and in the VP rule.

NB. Its (normalised) semantic type is non-redundantly specified.

VP is the functor, which, applied to a leftward NP, yields an S; i.e. VP = (S\NP)

Viv = VP = (S\NP)

Vtv = VP/NP = (S\NP)/NP

13

Therefore, the only ineliminable parts of the grammar above are slept andread :

slept def= Viv = (S\NP)

read def= Vtv = (S\NP)/NP

Assuming S to be of semantic type t, and NP to be e, a fully interpretableequivalent of the grammar above is

slept := S\NP e 7→ t

read := (S\NP)/NP e 7→ (e 7→ t)

14

Predicate-Argument Structure:Syntacticised argument structure (LF)

This is a point of departure for CCG.

Dowty, Jacobson, Szabolcsi follow the strict Montagovian tradition of notpredicating anything on LF (LF is dispensible).

Steedman, Baldridge, Bozsahin, White follow Montague in the sense thatsyntax is purely type-dependent and entirely blind to both derivation andLF (as opposed to structure-dependent on either derivation or LF), but LFis involved in formulating eg. binding and NL generation.

Type-logical Grammar maintains Lambek calculus for base logic (i.e. nocombinators in ‘base’) and a strict Montagovian regime, but pure type-dependence is the unifying theme (see Morrill 1994 and Oehrle 2000 forTLG).

15

The semantic type e 7→ t corresponds to a predicate over a single argument,e.g. sleep′ x

The x argument is associated with the syntactic type \NP, via the lambdabinding

slept := S\NP: λx.sleep′ x

Binding asymmetries of these syntacticised arguments are encoded in thepredicate-argument structure.

16

read := (S\NP)/NP: λxλy.read′ xy

©©©

HH

H

©©©

HHH

read′ x

y

pred′ xnxn−1 · · ·x1 reflects the primacy of n arguments where xi−1 immedi-ately dominates xi, yielding the following hierarchy for the linearised nota-tion:

©©

©©©

HH

HHH

©©

©©©

HH

HHH

©©

©©

HH

HH

©©©

HHH

pred′ xn

xn−1

· · ·

x1©

©©

HH

H

©©©

HHH

pred′ 2

1

17

(The terms 1, 2, 3 are borrowed from Relational Grammar (Blake, 1990),but there can be no promotion or demotion of them in CCG’s monostratalarchitecture.)

18

The semantics of read, λx︸︷︷︸

eλy

︸︷︷︸

e

.read′︸︷︷︸

txy, is the same as the lexical entry

in the PSG.

Actually, lambda bindings are a matter of convenience to associate syntaxand semantics; they can all be eliminated.

Eta-conversion:

λx.Fxη

←→ F (if x does not occur free in F)

e.g. λx.1+ xη= 1+

λx1λx2λx3.show′ x2x1x3 = λx1λx2λx3.C(show′BCI)x3x2x1

19

B, C and I are combinators (Curry and Feys, 1958), whose operation

can be defined in lambda-calculus, e.g. Ide f= λx.x

Commutator Composer Identity

C f ab = f ba B f ga = f (ga) Ia = a

λx1λx2λx3.C(show′BCI)x3x2

︸︷︷︸

F

x1η= λx2λx3.C(show′

BCI)x3︸︷︷︸

F

x2

η= λx3.C(show′

BCI)︸︷︷︸

F

x3η= C(show′BCI)

20

show′ x2x1x3?= C(show′BCI)x3x2x1

C(show′BCI)x3x2x1 =

show′BCIx2x3x1 =

show′BC(Ix2)x3x1 =

show′C(Ix2)x3x1 =

show′ (Ix2)x1x3 = CI = T (aka. C∗)

show′ x2x1x3 CIa f = Ta f = f a

21

Don’t try this at home; this is just to show that CCG is a combinatorytheory of syntax-semantics because combinators do show up in LF; notbecause derivations are decorated with the names of the combinators.

Categorial Grammar is a theory of grammar in which the form-meaning re-lation is conceived as a transparent correspondence between the surface-syntactic and semantic combinatorics (Jacobson, 1996).

The other place combinators show up is the universal grammar of syntacticprojection (more on this later).

Universal Grammar of CCG can be conceived as a specialisation of combi-nators for language; combinators might be at work at other symbol-drivencognitive activity, eg. planned sequence of actions (see Steedman 2002for more).

22

Why (S\NP)/NPacc/NPdat: λx1λx2λx3.show′ x2x1x3 for show ?

John showed Mary herself *John showed herself Mary

John showed Mary to herself *John showed herself to Mary

The order in which syntactic arguments are taken is different than theirobliqueness order.

This is another point of departure in CCG. Bach (1979) and Dowty (1996)allow “wrap” in syntax, and Szabolcsi (1989) simulates wrap so that oblique-ness is preserved in the syntactic types (therefore, no use of LF).

However, the following example (from Szabolcsi) is problematic withoutextra categories for reflexives, and binding in VSO languages remains un-accounted for:

23

John introduced [[Mary to himself]and [Susan to herself]]

Wrap in syntax: (S\NP)/NP/w PP: introduce′ cf. (S\NP)/PP/ NP

‘Introduced Mary’ must compose before the reflexive (allowed by wrap) sothat subject or object can bind the reflexive without the use of LF.

Argument cluster coordination does not follow as a theorem of coordinationworking on same types (more on this later).

The version we assume is LF (and only lexical) use of “wrap” to captureobliqueness order in LF.

(S\NP)/NPacc/NPdat : λx1λx2λx3.show′ x2x1x3

Binding restrictions of VSO languages now hold at LF, as in all others in-cluding OVS and OSV languages, with one category for reflexives (for En-glish).

24

Besides theoretical implications, wrapped categories no longer representsurface order of constituents.

25

Type determinism significantly constrains the possible syntax-semanticscorrespondence in the lexicon (and predicts some LFs depending on theavailability of some categories).

The principle of Categorial Type Transparency: (Steedman 2000:36)

For a given language, the semantic type of the interpretation to-gether with a number of language-specific directional parametersettings uniquely determines the syntactic category of a category.

This principle works both ways: The semantic type of an interpretation isentirely determined by the syntactic type.

26

e.g. S is t NP is e N is e 7→ t (property)

If A is of type α and B β, then A\B and A/B have the semantic type β 7→ α

see := (S\NP)/NP: λxλy.(see′ x)y e 7→ (e 7→ t)

The LF λxλy.(see′ y)x for English is ruled out by the following observation:John saw himself (subject must LF-command object)

It is the right LF for a VSO language such as Irish (and for the same rea-son): (S/NP)/NP: λxλy.(see′ y)x

λx.see′ xx is universally disallowed: *Heself saw John (only pro-terms—eg.x in (ana′ x) or (pro′ x)—can be LF-commanded by themselves, and the firstoccurence of x is not a pro-term).

27

Predicates such as help avail some LFs that can materialise different waysunder PCTT:

John helped him to fix the car

((S\NP)/(Sinf\NP))/NP

e 7→ ((e 7→ t) 7→ (e 7→ t))

: λy︸︷︷︸

e

λP︸︷︷︸

e7→tλx

︸︷︷︸

e.help′︸︷︷︸

t

Pyx

“Exceptional” case marking verbs pair naturally with help-like verbs:

Both : λyλPλx.pred′Pyx and : λyλPλx.pred′ (Py)x are possible by PCTT:

28

John expects him to fix the car

((S\NP)/(Sinf\NP))/NP

e 7→ ((e 7→ t) 7→ (e 7→ t))

: λyλPλx.expect′ (Py)x

Him (y) gets its case from expect as the /NP-argument of the verb.

It is not a semantic argument of expect, but fix the car (P).

Can we eliminate all rules? Syncategorematic rules appear to be problem

X → X andX

where X is any category in the grammar (N,V,S,NP,VP,PP...)

and = (X/X)\X

In a lexicalised grammar, “any category” includes all the lexical categories,

thus S, N, V={(S\NP), (S\NP)/NP, (S\NP)/NP/NP}, VP, NP, PP coordina-tion is predicted.

Also predicted is the conjunction of infinitely many categories that are de-rived from the lexical categories (i.e. the closure of the lexicon with respectto all combinatory possibilities)

29

The simplest combinatory possibility is function application :

X/Y : f Y : a ⇒ X : f a (>)

Y : a X\Y : f ⇒ X : f a (<)

which allows lexical and derived VPs to coordinate:

Mary read the book and studied

(S\NP)/NP NP/N N S\NP>

NP>

S\NP

30

There is no record of history of derivations in CCG.

Therefore, no way to ‘peek’ inside a constituent to extract some informa-tion.

Example: a step-by-step derivation

31


NP (S\NP)/NP NP/N N (X\X)/X S\NP

32


NP (S\NP)/NP NP/N N (X\X)/X S\NP>

NP

33


NP (S\NP)/NP (X\X)/X S\NP

NP

34


NP (S\NP)/NP (X\X)/X S\NP

NP>

S\NP

35


NP (X\X)/X S\NP

S\NP

36


NP (X\X)/X S\NP

S\NP>

(S\NP)\(S\NP)

37


NP

S\NP

(S\NP)\(S\NP)

38


NP

S\NP

(S\NP)\(S\NP)<

S\NP

39


NP

S\NP

40


NP

S\NP<

S

41


S

42


NP (S\NP)/NP NP/N N (X\X)/X S\NP>

NP>

S\NP>

(S\NP)\(S\NP)<

S\NP<

S

43

Same example in OpenCCG:cem-english> tccgLoading grammar from URL: file:/home/bozsahin/openccg/grammars/cem-english/grammar.xml

Enter strings to parse.Type ’:r’ to realize selected reading of previous parse.Type ’:h’ for help on display options and ’:q’ to quit.

tccg> Mary read the book and studied1 parse found.

Parse: s------------------------------(lex) Mary :- np(>T) Mary :- s/@i(s\@inp)(lex) read :- s\.np/.np(lex) the :- np/^n(lex) book :- n(>) the book :- np(>) read the book :- s\.np(lex) and :- s$1\*(s$1)/*(s$1)(lex) studied :- s\.np(>) and studied :- s\.np\*(s\.np)(<) read the book and studied :- s\.np(>) Mary read the book and studied :- s

44

Strings like the following need more than application, unless we allowphonologically empty elements in the lexicon:

Mary wants to read and to discuss

(Sinf\NP)/(S\NP) (S\NP)/NP (Sinf\NP)/(S\NP) (S\NP)/NP? ?

(Sinf\NP)/NP (Sinf\NP)/NP

the stories of Edgar Allan Poe

Function composition

X/Y : f Y/Z : g ⇒ X/Z : λx. f (gx) (> B)

Y\Z : g X\Y : f ⇒ X\Z : λx. f (gx) (< B)

45

Type Raising : turning arguments into functions looking for functors lookingfor such arguments.

A : a ⇒ T/(T\A) : λ f . f a (> T)

A : a ⇒ T\(T/A) : λ f . f a (> T)

where A is an argument category in the lexicon, and T is the result in anyfunction category over A that the grammar licenses.

Combinatory UG only uses the formal categories such as X and Y,

Lexical rules use substantive categories such as lexical generalisations Tand A above, which are generalisations over S, NP, PP etc.

46

Type raising and composition engenders so-called “non-constituent coor-dination”.

It is constituent coordination in CCG because the conjuncts are fully inter-pretable surface constituents:

[Johnson admires] and [Monboddo says he detests] a saxophonist> B > B < T

S/NP (S$\?S$)/?S$ S/NP S\(S/NP)>

(S/NP)\?(S/NP)<

(S/NP)<

S

47

Johnson admires

NP (S\NP)/NP: johnson′ : λzλw.admires′ zw

> T

T/(T\NP): λP.Pjohnson′

> B

S/NP: λx.admires′ x johnson′

B f g = λx. f (gx) = λx.admires′ x johnson′

(gx) = (λzλw.admires′ zw)x = λw.admires′ xw

f (gx) = λP.Pjohnson′(λw.admires′ xw) = (λw.admires′ xw)johnson′

= admires′ x johnson′

48

Unbounded compositions of type S/NP are predicted to be conjoinablewith Johnson admires:

Monboddo says he detests

NP (S\NP)/S NP (S\NP)/NP> T > T

T/(T\NP) T/(T\NP)> B

S/S> B

S/(S\NP)> B

S/NP

49

The star modalities (\? and /? ) on the slash allow lexical control of theconstruction, e.g. like-categories for coordination (Baldridge, 2002):

*player that shoots and he misses

(N\N)/(S\NP) S\NP (S\?S)/?S S>

S\?S*** < B

S\NP

50

?

¦ ×

·

The slash modalities

• The ? modality is the most restricted and allows only the most basicapplicative rules;

• ¦ permits order-preserving associativity in derivations;

• × allows limited permutation;

• and · is the most permissive, allowing all rules to apply.

51

Principles constraining universal combinatory syntax:

The Principle of Adjacency (PA): Steedman (2000:54)

Combinatory rules may only apply to finitely many phonologically realisedand string-adjacent entities.

The Principle of Consistency (PC):All syntactic combinatory rules must be consistent with the directionality ofthe principal functor.

The Principle of Inheritance (PI):If the category that results from the application of a combinatory rule is afunction category, then the slash type of a given argument in that categorywill be the same as the one(s) of the corresponding argument(s) in theinput function(s).

52

Adjacency: syntactic projection via UG rules has combinatory basis.

Consistency: certain rules cannot be part of UG eventhough they satisfyadjacency:

X/?Y Y ⇒ X (>)

Y X\?Y ⇒ X (<)

Y X/?Y ⇒ X (disallowed)

X\?Y Y ⇒ X (disallowed)

53

PI eliminates composition rules below, eventhough PA and/or PC hold:

X/¦Y Y/¦ Z ⇒ X\¦ Z (disallowed)

X/¦Y Y/¦ Z ⇒ X/× Z (disallowed)

X/¦Y Y/¦ Z ⇒ X/¦ Z (> B)∗

Y\¦ Z X\¦Y ⇒ X\¦ Z (< B)

∗These are rule schemata showing the compatibility of modalised slashes, ratherthan rules doing the actual work. The (> B) rules can be fleshed out asX/¦Y Y/· Z ⇒ X/· Z, X/¦Y Y/¦ Z ⇒ X/¦ Z, X/·Y Y/¦ Z ⇒ X/¦ Z etc.

54

Unlike Lambek (1958) calculus,∗ the following are possible rules of univer-sal syntactic projection; they satisfy all principles

The crossing functional composition rules

X/×Y Y\×Z ⇒ X\×Z (>B×)

Y/×Z X\×Y ⇒ X/×Z (<B×)

∗See Moortgat (1997) for a general introduction.

55

They have a re-ordering effect, but conserve directionality under syntacticprojection because of the Principle of Inheritance:

C A/×B B\×C> B×

A\×C<

A

A/×B C B\×C<

B>

A

A/×B B\×C C> B×

A\×C*** <

56

Adjuncts and second arguments can invert order, as e.g. in Heavy NPshift:

I introduced to Marcel some very heavy friends

S/(S\NP): ((S\NP)/PPTO)/NP: S\(S/PPTO): S\(S/NP):λp.p me′ λxλyλz.introduce′yxz λq.q marcel′ λr.r f riends′

> B2

(S/PPTO)/NP:λxλy.introduce′yx me′

< B×

S/NP: λx.introduce′marcel′ x me′<

S:introduce′marcel′ f riends′me′

Obliqueness of arguments is preserved (i.e. there is one LF), thus wewould expect a single category assignment for such lexical items.

Such derivations preserve the binding condition C at the level of logicalform as required by I introduced to each other some very heavy friends.

57

Having eliminated rules, we would expect constructions to follow from thelexical categories (of heads and specifiers of syntactic constructions) alone.

Combinatory syntax simply projects lexical properties, including direction-ality and LF.

The Principle of Lexical Head Government (PLHG):Both bounded and unbounded syntactic dependencies are specified by thelexical syntactic type of their head.

Syntactic derivation is purely syntactic type driven; LF cannot undo a deriva-tion (like GB/MP, and unlike HPSG and LFG).

This is not to say that LF plays no part in shaping the lexical syntactic type;cf. PCTT.

58

Unlike GPSG, HPSG and TAG, CCG also attempts to adhere to lexicaleconomy:

The Maxim of Head Categorial Uniqueness (HCU):A single nondisjunctive lexical category for the head of a given constructionspecifies both the bounded dependencies that arise when its complementsare in canonical position and the unbounded dependencies that arise whenthose complements are displaced under relativisation, coordination, andthe like.

59

• Johnson admires Monboddo.

• the man that I believe that Johnson admires

• I believe that Johnson admires and you believe that he despises, thecelebrated judge Lord Monboddo.

In both TAG and GPSG these dependencies are mediated by different ini-tial trees or categories, and in HPSG they are mediated by a disjunctivecategory.

60

Constructions:

Domain of locality and (un)boundedness

61

The Principle of Lexical Head Government (PLHG):Both bounded and unbounded syntactic dependencies are specified by thelexical syntactic type of their head.

Syntactic derivation is purely syntactic type driven; LF cannot undo a deriva-tion (like GB/MP, and unlike HPSG and LFG).

62

In a strictly lexicalised grammar, the domain of locality can only be definedby the lexical type of the head (there is no other locus to define it).

Domain of locality of pronouns is another point of departure for CCG.

CCG without GZ combinators leaves condition B (binding of pronouns) todiscourse (i.e. pronouns are not pro-terms in LF).

Something like Centering Theory (Grosz et al., 1995) or a discourse gram-mar (Webber, 2004) is needed for its capture. See e.g. Yüksel and Bozsahin(2002) for the use of centering theory for coreference possibilities that areleft open by condition B of the binding theory.

63

Pronominal reference possibilities are well-known to nest or intercalate(Jacobson, 1999), unlike syntactic dependencies, which either nest or cross(but not do both):

Every mani thinks that every boy j said that his j mother loves hisi dog.

Every mani thinks that every boy j said that hisi mother loves his j dog.

It is debatable whether binding restrictions on pronouns are combinatorialin nature.

But the issue is deeper, involving interaction of binding and extraction, lead-ing ultimately to having a linguistic LF or model-theoretic syntax, althoughboth approaches are type-driven; cf. (Jacobson, 1999; Szabolcsi, 1992;Steedman, 2005b) for further discussion.

64

Since a lexical item is only represented by an LF and a syntactic type, andits syntactic behaviour is regulated only by its syntactic type, the (un)boundednessof a construction must be a conspiracy of its syntactic type and LF.

that := (N\N)/(S/¦NP): λP︸︷︷︸

e7→tλQ.and′(Px)(Qx)

want := (S\NP)/(Sinf\NP): λP︸︷︷︸

e7→tλx.want′ (P(ana′ x))x

Although the syntactic type of their arguments is susceptible to long-distancecomposition, we know that relativisation is unbounded and control is not:

65

Çocuk i [[ i/∗ j adam-ak [ k/∗i/∗ j kitab-ı oku]]-t-mak ] ist-iyor

child man-DAT book-ACC read-CAUS-INF want-PROG

‘The child wants to have the man read the book.’ Bozsahin 2004* for ‘The childi wants the man (or someone) to have himi/heri read the book.’

Although subjects can be controlled in Turkish, the subject of read is fartoo embedded for the control verb to ‘see’ it.

66

Unboundedness of relativisation and boundedness of control follow fromdifferent syntactic manifestation of the e 7→ t type of P in their LF:

that := (N\N)/(S/¦NP): λP︸︷︷︸




P is of the form λz.pred′ · · ·z · · · for the relative pronoun:

a. The manx that Anna gave the book P = give′ xbook′anna′

b. The manx that Anna saw P = see′ xanna′

c. The bookx that Anna showed the man P = show′man′ xanna′

67

If P in turn takes a sentential or a VP complement for z, the x argument canbe passed down indefinitely (argument category S/NP can be obtained byrepeated composition):

The manx that Manny says [you claim that]S/S [Anna gave the book]S/¦NP

P = λx.say′(claim′(give′ xbook′anna′)you′)manny′

P is of the form λz.pred′ · · ·z for the controlled complement of want ; only thesubject can be missing, not a complement (hence no possibility of passingdown x indefinitely).

This is determined by the PCTT: infinitival VPs have all their complements(non-1s); a type such as λz.pred′ · · ·z · · · for English violates PCTT (not forDyirbal, more on this later).

68

Domains of locality for relativisation and control are embodied in their lexi-cal syntactic category:

The relative pronoun (as head) can only take non-subject residues (S/¦NPtype).

The controller-controllee relation can only be between the controlling argu-ment (NP) and the subject of the controlled clause (because P below is ofthe form λz.pred′ · · ·z).

that := (N\N)/(S/¦NP): λP︸︷︷︸




69

Subject and Object Control

Control is a lexical property. Control verb’s syntactic type and LF mustencode domain of locality and the controller-controllee relation.

If xi is the term for the controller, then (ana′ xi) is the term for the controllee,with ana′ providing the bound argument interpretation for ana′ x. The con-troller LF-commands the controllee.

Since control is a lexical property, it is conceivable that what is controlledcan be syntactically or semantically restricted (because, apart from a phono-logical form, that’s all we have in the category of a lexical item).

Bozsahin and Steedman (2005) claim that a fundamental asymmetry de-termines the categories of heads of constructions if they single out oneargument against all others (more on that later):

70

(1) The Principle of Lexicalised Asymmetry :Syntactic asymmetries are mediated by S , and semantic asymmetriesare mediated by 1, as determined by the syntactic or semantic type ofthe lexical category of the head of the construction.

PLA requires that the syntactic type of S and the semantic type of 1 bediscernible in the category of the lexical item.∗

This is possible in a purely lexicalised grammar, and it significantly con-straints the notion of “possible categories” in the lexicon.

∗PLA might be derivable from simpler considerations—as one reviewer noted, such asineffability; all verbs have an S , no matter what their LF is, and all verbs have a 1, nomatter what their syntactic type is. Other restrictions would exclude a certain lexical sub-class of verbs. Eg. for intransitive subjects, control might be imaginable but syntacticallyimpossible if eg. 2-argument can regulate control.

71

(2)a. promise := (S\NP)/(Sinf\NP)/NP : λx2λPλx1.promise′ (P(ana′ x1))x2x1

b. persuade := (S\NP)/(Sinf\NP)/NP: λx2λPλx1.persuade′ (P(ana′ x2))x2x1

The categories above capture the fact that syntactic subjects are the onlycontrollees in English.

(3)a. John wants to clean the window.

b. The dog wants to be petted.

c. Susie wants to grow up.

The controller is either the subject or the object.

72

Example (4) shows control of the unaccusative subject. NB. LF of grow up.

(4) Susie wants to grow up

NP (S\NP)/(Sinf\NP) (Sinf\NP)/(S\NP) S\NP: susie′ : λPλx.want′(P(ana′ x))x : λQ.Q : λx.growup′ xone′

>Sinf\NP

>S\NP

<S : want′(growup′ (ana′ susie′)one′)susie′

73

Agents of unergatives and transitives can be controlled, because they arealso syntactic subjects (in accusative languages):

(5)a. John promised him [ to exercise]

NP (S\NP)/(Sinf\NP)/NP NP (Sinf\NP)/(S\NP) S\NP> >

(S\NP)/(Sinf\NP) Sinf\NP>

S\NP<

S : promise′(exercise′ (ana′ john′))him′ john′

74

“Exceptional” case marking is not exceptional at all; him is an argument ofpersuade, and there is no “surface linking” of it to the controlled clause:

(6)a. John persuaded him [ to read the novel ]

NP (S\NP)/(Sinf\NP)/NP NP (Sinf\NP)/(S\NP) (S\NP)/NP NP> >

(S\NP)/(Sinf\NP) S\NP>

Sinf\NP>

S\NP<

S : persuade′(read′novel′ (ana′him′))him′ john′

75

Given the category of the infinitival VP for English, there can be no controlof the complements or adjuncts of the controlled clause:

(7) *John persuaded him [Sue see]

NP (S\NP)/(Sinf\NP)/NP NP NP (S\NP)/NP> > T

(S\NP)/(Sinf\NP) S/(S\NP)> B

S/NP*** *

This example works well in Dyirbal, and this is predicted by CCG becausethe category VPinf = (Sinf\NP) is lexicalised.

76

In summary, control’s domain of locality encompasses a controller (NP),and a controlled clause (VPinf), both of which are lexically specified. Thereis no other domain that the control verb can control.

It is bounded, because the syntactic type of the controlled clause, VPinf,cannot pass the controller information down to its complements; only thenon-subcategorised argument can be missing.

This argument coincides with the maximally LF-commanding argument intransitives of accusative languages (i.e. 1). It is not necessarily so in erga-tive languages (e.g. Tagalog and Dyirbal).

77

Coordination : Ross’s (1967) Coordinate Structure Constraint (coord. struc-ture is an island)

This man loves animals and he dislikes children.

Animals this man loves and children dislike him.

Animals love this man and he dislikes children.

*a man who(m) loves animals and he dislikes children.

*a man who(m) animals love and children dislike him.

78

Across-the-board exception (ATB) to CSC: extraction is ok if the NP is miss-ing from all conjuncts:

A man who(m) animals love and children dislike.

A man who loves animals and dislikes children.

Exceptions to ATB exception: the extracted argument must bear the samegrammatical relation in all conjuncts:

*a man who dislikes children and animals love.

*a man who(m) animals love and dislikes children.

79

CCG’s account of coordination moved from syncategorematic rules to lex-icalisation.

and := (X$\?X$)/?X$: λpλq.p∧q

Lexical control of conjuncts assure like-category coordination (Baldridge,2002, p.97–99), without syncategorematic rules (e.g. X conj X → X)

*player that shoots and he misses

(N\N)/(S\NP) S\NP (S$\?S$)/?S$ NP S\NP<

S>

S\?S*** < B

S\NP

80

In OpenCCG:

cem-english> tccgLoading grammar from URL: file:/home/bozsahin/openccg/grammars/cem-english/grammar.xml

Enter strings to parse.Type ’:r’ to realize selected reading of previous parse.Type ’:h’ for help on display options and ’:q’ to quit.

tccg> shoots and he missesUnable to parsetccg> and he misses1 parse found.

Parse: s\*s------------------------------(lex) and :- s$1\*(s$1)/*(s$1)(lex) he :- np(lex) misses :- s\.np(<) he misses :- s(>) and he misses :- s\*s

81

Lexicalisation implies that languages might have different kinds of coordi-nation. Malagasy uses ary for sentential coordination, and sy for phrasalcoordination:

(8)a. Misotro taoka Rabe ary mihinam bary Rabe

drink alcohol R and eat rice R

‘Rabe is drinking alcohol and Rabe is eating rice.’

b. Misotro taoka sy mihinam bary Rabe

drink alcohol and eat rice R

‘Rabe is drinking alcohol and eating rice.’

82

As in GPSG (Gazdar, 1981), coordination is type-dependent, rather thanstructure-dependent (Steedman, 2000):

*Anna [met Manny]S/NP and [married](S\NP)/NP

Across-the-board rule application (Williams, 1978) and other structure-dependentaccounts need extra conditions on surface forms.

LFs of conjuncts get their copy of the variable, but there is no syntacticoperation for that; the lexically-specified LF ‘copies’ them.

This is possible because X’s in syntactic types and ∧ in semantic type areschematisations over valencies (Partee and Rooth, 1983):

and := (S$\?S$)/?S$: λpλq.p∧q

83

Anna met and married Manny

NP (S\NP)/NP (S$\?S$)/?S$ (S\NP)/NP NP: a′ : λx1λx2.meet ′x1x2 : λy1λy2.marry′y1y2 : m′

>((S\NP)/NP))\?((S\NP)/NP))

: λqλxλy.(λy1λy2.marry′y1y2[y][x])∧q[y][x])<

(S\NP)/NP: λxλy.(λy1λy2.marry′y1y2[y][x])∧λx1λx2.meet ′[y][x])

= λxλy.marry′xy∧meet ′xy>

S\NP: λy.marry′manny′ y∧meet ′manny′ y

We can expect the schematised syntactic types (Xs) to eliminate otherwiselegitimate LFs (NB. syntactic derivation is blind to LF):

* the man who married Mary and John disliked

(S\NP)/NP NP (S$\?S$)/?S$ NP (S\NP)/NP> > T

S\NP S/(S\NP)> B

S/NP*** &

and′(marry′mary′man′)(dislike′man′ john′)

The Xs (S$ above) are not token-identical but type-identical.

Since these are lexical types, we would expect different constructions toimpose different requirements (of agreement etc.) via their heads:

I like/*likes vodka, and Mary, beer.

84

As in GPSG, the type-dependent account of extraction and coordination,as opposed to the standard account using structure-dependent rules, makesthe across-the-board condition (ATB) on extractions from coordinate struc-tures (including the “same case” condition) a prediction of CCG.

a. A saxophonist [that(N\¦N)/¦(S/NP) [[Johnson admires]S/NP and [Mon-boddo detests]S/NP]S/NP]N\¦N

b. A saxophonist [that(N\¦N)/¦(S/NP) *[[Johnson admires]S/NP and [de-tests Monboddo]S\NP]S/NP]N\¦N

c. A saxophonist [that(N\¦N)/¦(S/NP) *[[Johnson admires]S/NP and [Mon-boddo detests him]S]]

d. A saxophonist that(N\¦N)/¦(S/NP) *[[Johnson admires him]S and [Mon-boddo detests]S/NP]

85

We would also expect that, in a verb-peripheral language (say SOV) withflexible word-order, the “same case restriction” is mediated not by theslash, but by morphological case; all arguments are on the same side ofthe verb:

(9) *Kız-ı [adam gördü ]S\NPacc[çocug-a da baktı]S\NPnom

girl-ACC man.NOM saw child-DAT and looked at

* for ‘The man saw the girl and the girl looked at the child.’

Still, the syntactic type is doing all the work.

86

Scrambling(or lack of it)

Although some languages have very flexible word order (like Warlbiri),most flexible-WO languages converge on basic word order(s):

Steele (1978) reports that for VSO languages, VOS is also observed to bevery frequent; for SOV languages, OSV seems to be an alternative.

Some syntactic processes, such as gapping, put the basicness of wordorder to the test.

87

Gapping : identical verb deletion under coordination (direction of deletiondepends on word order and—it seems—on nothing else; cf. Ross (1970)and ensuing discussion)

In a purely lexicalised grammar, the only word order is the lexical categoryof the verbs; there is not other locus for word order to arise.

Unlike Kayne (1994), there can be no universally assumed word order inCCG. Lexicalised word orders ought to reflect the limited cross-linguisticdiversity in syntax (e.g. forward gapping, backward gapping, both).

88

Backward gapping (Japanese):

• Ken-ga Naomi-o, Erika-ga Sara-o tazuneta

K-NOM N-ACC E-NOM S-ACC visit-PAST.CONCL

‘Ken visited Naomi, and Erika, Sara.’ Japanese

• *Ken-ga Naomi-o tazunete, Erika-ga, Sara-o.

89

Forward gapping (Irish and English):

• Chonaic Eoghan Siobhán agus Eoghnaí Ciarán.

saw Eoghan Siobhán and Eoghnaí Ciarán

‘Eoghan saw Siobhán, and Eoghnaí, Ciarán.’ Irish

*Eoghan Siobhán agus chonaic Eoghnaí Ciarán.

• John likes whisky, and Stuart wine.

*John whisky, and Stuart likes wine.

90

Both (Turkish):

• Adam dergi-yi oku-du, kız da kitab-ı

man.NOM magazine-ACC read-PAST girl.NOM and book-ACC

‘The man read the magazine, and the girl, the book.’

Dergi-yi adam oku-du, kitab-ı da kız

• Adam dergi-yi, kız da kitab-ı oku-du

man.NOM magazine-ACC girl.NOM and book-ACC read-PAST

‘The man read the magazine, and the girl, the book.’

Dergi-yi adam, kitab-ı da kız oku-du

91

Steele’s observation on common cooccurence of SOV and OSV, and VSOand VOS, can be captured with a single verbal category in CCG:

SOV and OSV S\{NPnom, NPacc} = S{\NPnom, \NPacc} Turkish

VSO and VOS S/{NPnom, NPacc} = S{/NPnom, /NPacc} Tagalog

SOV only (S\NP)\NP I.jo.∗

VSO only (S/NPacc)/NPnom Irish

SVO only (S\NP)/NP English

any order S{ |NPnom, |NPacc} Latin

∗I.jo. seems to have no case, cf. Greenberg (1963:10).

92

Set-CCG (Hoffman, 1995) allows schematisation over lexical categoriesto be formulated in the lexicon. Set-modal CCG (Baldridge, 2002) addsmodalities for lexical control over the combinatory aspects of the domainof locality of a construction.

Convention for syntactic type-LF pairing in set-based categories: The setsof syntactic and semantic arguments are paired in left-to-right order

vur (‘hit’) := S{\NPacc, \NPnom} : λ{x,y}.hit′ xy

This schematisation corresponds to a finite set of fully ordered categories:

{S\NPacc\NPnom: λx1λx2.hit′ x2x1, S\NPnom\NPacc: λx2λx1.hit′ x2x1 }

93

SVO and OVS cannot be captured with a single category, as eg.

(S\NP)/NP

If binding conditions do not reverse in OVS languages, these orders oughtto lead to different LFs:

SVO : (S\NP)/NP: λxλy.pred′ xy

OVS : (S\NP)/NP: λxλy.pred′ yx Hixkaryana (Derbyshire, 1977)

Another alternative for OVS is (S/NP)\NP: λxλy.pred′ xy

The latter would be the right category if O [SV] and [SV] coordinationis possible in the language (to be checked!)

94

The following word orders of Turkish are “unmarked” if Turkish is consid-ered verb-final, not just SOV:

adam kitab-ı oku-duman.NOM book-ACC read-PAST

NPnom NPacc S{\NPacc, \NPnom}<

S\NPnom<

S

kitab-ı adam oku-dubook-ACC man.NOM read-PAST

NPacc NPnom S{\NPacc, \NPnom}<

S\NPacc<

S

95

In OpenCCG:

tccg> adam kitab i okudu1 parse found.

Parse: s------------------------------(lex) adam :- n(>T) adam :- s$1/@i(s$1\@in)(lex) kitab :- n(lex) i :- n\*n(<) kitab i :- n(>T) kitab i :- s$1/@i(s$1\@in)(lex) okudu :- s{\.n\.n}(>) kitab i okudu :- s\.n(>) adam kitab i okudu :- s

tccg> kitab i adam okudu1 parse found.

Parse: s------------------------------(lex) kitab :- n(lex) i :- n\*n(<) kitab i :- n(>T) kitab i :- s$1/@i(s$1\@in)(lex) adam :- n(>T) adam :- s$1/@i(s$1\@in)(lex) okudu :- s{\.n\.n}(>) adam okudu :- s\.n(>) kitab i adam okudu :- s

96

Non-verb-final orders require backgrounding (detopicalisation) of the ar-gument: They need contraposition : reversal of directionality along withchange in information structure (Bozsahin, 2002).∗

(10) Leftward Contraposition (< T×): NP:a ⇒ S/(S/×.NP+top):λ f . f a

Rightward Contraposition (>T×): NP:a ⇒ S\(S\/×NP-top):λ f . f a

(< T×) is topicalisation, and (> T×) detopicalisation/backgrounding.

This is a lexical rule, referring to substantive categories only.

∗This schema simplifies and corrects the one in the paper.

97

The rule refers to modalities that are in the full repertoire of set-modal CCG:

·

/ .

¦/× ×.

?

NB. With this set, the × modality becomes a schema over the slashes‘/’ and ‘\’

\× is same as \/×, and /× is same as /×.. The modality in eg. \×. cannotbe suppressed.

98

kitab-ı oku-du adambook-ACC read-PAST man.NOM

NPacc S{\NPacc, \NPnom} NPnom<

S\NPnom> T×

S\(S\/×NP-top,nom)<

S

adam oku-du kitab-ıman.NOM read-PAST book-ACC

NPnom S{\NPacc, \NPnom} NPacc<

S\NPacc> T×

S\(S\/×NP-top,acc)<

S

99

oku-du kitab-ı adamread-PAST book-ACC man.NOM

S{\NPacc, \NPnom} NPacc NPnom> T×

S\(S\/×NP-top,acc)< B

S\NPnom> T×

S\(S\/×NP-top,nom)<

S

Long-distance scrambling can be lexically controlled: Turkish allows it,Japanese does not (example from Baldridge 2002:148; slightly modified):

Bu kitab-ı Fatma oku-mak iiste-dig-im-i bil-iyor ben-imithis book-ACC F.NOM read-INF want-COMP-AGR.1s-ACC know-PROG I-GEN

NPacc NPnomSinf {\NPgen, \NPacc} Svacc\Sinf S{\NPnom, \Svacc} NPgen< B

Svacc{\NPgen, \NPacc}< B

S{\NPnom, \NPgen, \NPacc}<

S{\NPgen, \NPacc}<

S\NPgen> T×

S\(S\/×NP-top,gen)<

S

‘Fatma knows that I want to read this book.’

Such derivations fail in Japanese at the point of (> T×). Verb’s argumentswould have the \¦ modality, rather than the most permissive \.

100

English topicalisation is an instance of leftward contraposition:∗

This book, I think you recommended that Mary read< T× > B > B > B

S/(S/×.NP+top) S/¦Sfin S/¦Sfin S/· NP> B

S/¦Sfin> B

S/· NP>

S

∗NB. The composition S/¦S S/· NP yields S/·NP, which projects NP’s slash modality,rather than S/¦NP, which does not.

101

Rightward contraposition should be disallowed (somehow) in the Englishlexicon to avoid examples like the following:

*I think irecommended that Mary read this book your brotheri> B > B > B > T×

S/¦Sfin (S\·NP)/¦Sfin S S\(S\/×NP-top)>

S\·NP<

S

One possibility, proposed by Baldridge (2002:114) but not adopted by him,∗

is to type the English subject NP as \¤NP, rather than \·NP, which wouldmaintain English subject/object asymmetry in extraction without overgen-erating for rightward displacement of the subject above.∗The concern was that this type would necessitate an instance of crossing compositionthat seems inactive in English. The example above might be the evidence that \¤and\/×modalities are active.

102

English is still a rigid word-order language without collapsing into Turkish-style local or long-distance scrambling (example from Baldridge, p.74):

* John Brazili knew that would idefeat China

NP NP (S\¤NP)/Sfin S/¦S S\¤NP> B

(S\¤NP)/¦S*** > B×

103

The lexical category of the verbs is the source of cross-linguistic variationfor the possibility of scrambling, and its range.

Universal combinatory syntax of CCG can only project these propertiesonto surface derivation.

Thus the syntactic consequences of having scrambling or no scramblingare directly accounted for, e.g. in coordination and topicalisation; cf. Baldridge(2002) and Bozsahin (2005) for more on this issue.

104

Summary of universally available combinatory rules:

X/?Y Y ⇒ X (>)

Y X\?Y ⇒ X (<)

X/¦Y Y/¦ Z ⇒ X/¦ Z (> B)

Y\¦ Z X\¦Y ⇒ X\¦ Z (< B)

X/×Y Y\×Z ⇒ X\×Z (>B×)

Y/×Z X\×Y ⇒ X/×Z (<B×)

105

Summary of universally available lexical rules

Type Raising

A : a ⇒ T/(T\A) : λ f . f a (> T)

A : a ⇒ T\(T/A) : λ f . f a (> T)

Contraposition

NP:a ⇒ S/(S/×.NP+top):λ f . f a (> T×)

NP:a ⇒ S\(S\/×NP-top):λ f . f a (< T×)

106

Ergativity and accusativity

as lexical systems

Moravcsik (1978) argues that there are ergative patterns in accusative lan-guages, and accusative patterns in ergative languages.

But there is skewness in the distribution of patterns and number of lan-guages: Most ergative languages are only morphologically ergative, splitergative or partially ergative, whereas most accusative languages are noteg. split accusative or partially accusative.

107

Grammatical roles v grammatical relations:

S : Single argument of a simple intransitive.∗

A : The agent-like (primary) argument of a transitive action verb like “hit.”

P : The more patient-like (secondary) argument of such a transitive verb.

Pivotal category is a category restricted to S -agreement. (Dixon defined“subject” as an S =A pivot, which might be called the “deep subject”. Ourusage is similar to LFG’s).

Pivot is an item with a pivotal category.∗Most of this terminology is due to Dixon (1979, 1994), eg. “pivot”, S and A . We use Prather than O that Dixon suggested, following Palmer (1994).

108

Morphological Ergativity : Same case for S and P :∗

Arnaq yurar-tuq

woman-ABS dance-IND.3SG

S

‘The woman dances.’ Yup’ik

Angutem tangrr-aa arnaq

man-ERG see-IND.3SG.3SG woman-ABS

A P

‘The man sees the woman.’

∗The examples of this section are from the sources cited in Bozsahin and Steedman(2005), unless stated otherwise.

109

Morphological accusativity : Same case for S and A :

Kadın dans ediyor

woman.NOM dance do-AOR.3SG

S

‘The woman dances.’ Turkish

Adam kadın-ı gör-dü

man.NOM woman-ACC see-PAST.3SG

A P

‘The man saw the woman.’

110

Syntactic ergativity is operation of S and P alignment in syntax (S =P ):

[bayi burrbula ba

�

gul gubi-

�

gu bara-n] [ baji-gu ]

I.ABS.TH B.ABS I.ERG.TH doctor-ERG punch-NFUT fall.down-PURP

P A S

‘The doctor punched Burrbulai and i fell down.’ Dyirbal

111

Mixed ergativity is presence of both an ergative system and a non-ergativeone:

m.33 ka55 m.

33 ko44 ndu21

Muga Mugo beat

‘Muga beats Mugo’ or ‘Mugo beats Muga’ Liangshan Nuosu(superscripts represent tones)

L.Nuosu has no case (ergative split along resultative verbs v others).

Disambiguating syntactic contexts provide alternative interpretations:

Muga beat Mugo and ran away : accusative interpretation (Muga runsaway)

Muga was beaten by Mugo and cried : ergative interpretation (Muga cries)

112

Partial ergativity : some syntactic constructions follow the ergative pattern,but not all:

nanuq Piita-p tugu -ta-a

polar.bear.ABS P-ERG kill -TR.PART-3SG

‘A polar bear killed by Piita.’ Inuit‘A polar bear that Piita killed’

miiraq kamat -tu-q

child.ABS angry -REL.INTR-SG

‘The child that is angry.’

*angut aallaat tigu-sima -sa-a

man.ABS gun.ABS take-PRF -REL.TR-3SG.SG

*‘the man who took the gun’

Manning (1996:99) claims that, In Inuit, any argument can be omitted undercoordination (ie., no pivot).

113

Split ergativity : Morphology of lexical classes might differ (pronouns vnouns in Dyirbal):

nana banaga-nyu

we.PL.NOM return-NFUT

‘We returned.’

nyurra nana-na bura-n

you.PL.NOM we.PL-ACC see-NFUT

‘You all saw us.’

nana nyurra-na bura-n

we.PL.NOM you.PL-ACC see-NFUT

‘We saw you all.’

nyurra nana-na bura-n banaga-nyu

[you.PL.NOM we.PL-ACC see-NFUT] [return-NFUT]

‘You all saw us and (we) returned.’

114

Lexicalising the mapping

between strings and LF

In CCG, the category of a free-order transitive action verb is as in (11a),while the category of an intransitive verb is as in (11b):

(11)a. S{|NPA , |NPP }

b. S{|NPS }

The grammatical roles S , A and P are descriptive labels, to be replacedby realisation of these roles by grammatical case.

115

The syntactic type |NPS stands for the NP bearing the grammatical caseof the single argument of the intransitive: ‘\NP’ for English, ‘\NPnom’ forTurkish, ’\NPc’ for Icelandic, where c can be any morphological case.

NB. There is no notion of ‘abstract case’ in the theory, but the fact thatan NP cannot both be an argument and not have a slash seems to befunctionally equivalent to the condition that all NPs must bear case.

Ergativity and accusativity can be modeled as different ways of organisingthe alignment of S with an argument of the transitive verb (if any):

116

12

ACCERG

predicate

s

APS { |NP , |NP }simple transitive:

simple intransitive: S {|NP }

117

(12) Lexical categories for verbs in an ergative system:

a. S|NPP :λx.pred′ x

b. S{|NPA , |NPP }:λ{x1,x2}.pred′ x2x1

Lexical categories for verbs in an accusative system:

c. S|NPA :λx.pred′ x

d. S{|NPA , |NPP }:λ{x1,x2}.pred′ x2x1

118

These are actually lexical schemas, to be instantiated by grammatical caseand directionality.

For example, the Japanese/Turkish accusative lexicon comes out as in (13a–b) whereas the Basque/Inuit ergative lexicon comes out as in (13c–d):∗

(13)a. S\NPnom:λx.pred′ x b. S{\NPnom, \NPacc}:λ{x1,x2}.pred′ x2x1

c. S\NPabs:λx.pred′ x d. S{\NPerg, \NPabs}:λ{x1,x2}.pred′ x2x1

∗Thus, empirically verifiable morphological ergativity sets the verbal categories as erga-tive, not whether the syntax of the language is analysed as ergative.

119

We can similarly define SVO, OVS, VSO and VOS variety of accusativeand ergative languages.

VP constituency that is associated with VSO languages such as Welsh isevident in the LF: V and O form an LF constituent in

©©

©©

HH

HH

©©

©©

HH

HH

predicate 2

1

NB. CCG does not correlate configurationality of languages with presenceor absence of VP nodes; all languages have a VP type , ie. S|NP, wherethe residual NP is the subject. (Non)configurationality is thus an outcomeof lexically-specifiable but rather limited directional liberties the languagescan exploit or ignore.

120

Using OpenCCG’s type system

to model ergativity

Bozsahin, Kruijff, and White (2005) describe OpenCCG in detail. This sec-tion covers some material specific to parameterising the grammar specifi-cations.

121

OpenCCG home page: http://openccg.sourceforge.net

You need Java SDK 1.5

Use CVS version of OpenCCG:

cvs -d:pserver:[email protected]:/cvsroot/openccg login

cvs -d:pserver:[email protected]:/cvsroot/openccg checkout openccg

mini-basque, mini-inuit etc. under grammars directory are small grammarswith worked out examples.

122

Building CCG grammars in OpenCCG

lexicon.xml

morph.xml

rules.xml

tccg

ccg−build grammarlexicon−base.xml

dict.xml

ccg−build skeleton preset−families.xmlparameters.xmltypes.xml

123

The following parameter file specifies the basic skeleton of verbal cate-gories in an ergative language:

124

<language name="Dyirbal" type="erg">

<parameters>

<infinitive subject-type="syntactic"/>

<iv><predicate syn-type="s"/><arg>

<s-argument dir="\" syn-type="np" case="abs"/></arg>

</iv>

<tv><predicate syn-type="s"/><setarg>

<a-argument dir="\" syn-type="np"/><p-argument dir="\" syn-type="np"/>

</setarg></tv>

</parameters></language>

125

We obtain verbal categories such as the following for ‘see’ and ‘return’(more detailed output shows syntactic features and LF)

tccg> buran1 parse found.

Parse: s{\.np\.np}------------------------------(lex) buran :- s{\.np\.np}

tccg> banaganyu2 parses found.

Parse 1: s\.np------------------------------(lex) banaganyu :- s\.np

Parse 2: s------------------------------(lex) banaganyu :- s\.np(gram) pd: s\.np => s(pd) banaganyu :- s

126

OpenCCG’s basic types (aka. sorts) can model accusative and ergativecase marking patterns, independent of language-particular case realisa-tion:

<types><type name="s-case" parents="subject-case"/><type name="subject-case"/><type parents="a-case" name="erg"/><type name="a-case"/><type parents="s-case p-case" name="abs"/><type name="p-case" parents="subject-case"/>

</types>

s−case

subject−case

p−case a−case

abs erg

127

An ‘accusative’ types file:

<types><type name="s-case" parents="subject-case"/><type name="subject-case"/><type parents="a-case s-case" name="nom"/><type name="a-case" parents="subject-case"/><type parents="p-case" name="acc"/><type name="p-case"/>

</types>

s−case

subject−case

a−case p−case

nom acc

All grammatical cases can be localised in types.xml, so that the grammaronly refers to non-leaf labels above cross-linguistically.

128

Which would be generated from the following parameters file:

<language name="English" type="acc">

<parameters>

<infinitive subject-type="syntactic"/>

<iv><predicate syn-type="s"/><arg>

<s-argument dir="\" syn-type="np" case="nom"/></arg>

</iv>

<tv><predicate syn-type="s"/><arg>

<a-argument dir="\" syn-type="np"/><p-argument dir="/" syn-type="np"/>

</arg></tv>

</parameters></language>

129

<family closed="true" pos="V" name="unergative"><entry name="primary">

<complexcat><atomcat type="s">

<fs id="0"><feat attr="index"><lf>

<nomvar name="E"/></lf>

</feat></fs>

</atomcat><slash dir="\"/><atomcat type="np">

<fs id="1"><feat attr="case" val="s-case"/><feat attr="index"><lf>

<nomvar name="X1"/></lf>

</feat></fs>

</atomcat>

130

Degrees of ergativity

and strictly lexicalised grammars

Although syntactically ergative languages such as Dyirbal seem to applythe same restrictions to all constructions (more on this later), this is a rarity.

Most ergative languages are like Inuit, in the sense that at least someconstructions escape the ergative pattern.

In other words, lexical heads of syntactic constructions have access toall and only the surface-syntactic categories to regulate the syntactic be-haviour of their arguments.

Thus, syntactic/partial/mixed ergativity is a question of what heads (actingas pivots or non-pivots) impose syntactically.

Split/morphological ergativity is a matter of the lexicon-morphology.

131

Bozsahin and Steedman (2005) claim that the limited diversity aroundergativity is due to systematic capture of the verbal categories and eg.infinitivals in the languages’ lexicons.

Currently, the systematicity is captured with the aid of parameters, whichare part of the theory (ie. they are not switches in the mind/brain etc.)

Which means, the value space of the parameters ought to be explained bya theory of the lexicon, to eliminate them from the theory.

The bias towards accusative languages involves interaction of global fac-tors shaping the lexical categories of argument-taking elements, includingat least topic-prominence, subject-prominence and agent-prominence.

132

A strictly lexicalised grammar attributes all degrees of ergativity to the lexi-con.

The syntactic behaviour of bounded and unbounded constructions is reg-ulated by their heads (cf. the Principle of Lexical Head Government).

The heads are lexical items; they can adopt an asymmetric behaviour orignore it.

If adopted, the lexical head of a syntactic dependency has one option:choose S as pivot (syntactic processes depend only on the syntactic type).

It has two options for lexical properties such as control: choose S or 1.

These limited degrees of freedom—we claim—can capture the apparentdiversity of languages, and different kinds of ergativity.

133

(14) The Principle of Lexicalised Asymmetry :Syntactic asymmetries are mediated by S , and semantic asymmetriesare mediated by 1, as determined by the syntactic or semantic type ofthe lexical category of the head of the construction.

PLA requires that the syntactic type of S and the semantic type of 1 bediscernible in the category of the lexical item.∗

This is possible in a purely lexicalised grammar, and it significantly con-straints the notion of “possible categories” in the lexicon.

∗PLA might be derivable from simpler considerations—as one reviewer noted, such asineffability; all verbs have an S , no matter what their LF is, and all verbs have a 1, nomatter what their syntactic type is. Other restrictions would exclude a certain lexicalsubclass of verbs. Eg. for unergatives, control might be imaginable but syntacticallyimpossible if eg. 2-argument can regulate control.

134

Full ergativity : Dyirbal syntax.

Both bounded and unbounded constructions follow the ergative pattern, ie.the S =P relation is singled out as the pivot.

The lexical category for -

�

u acts as a pivot for relativization.

bayi yara [miyanda] -

�

u yanuI.ABS.TH man.ABS laugh -REL go.NFUT

NPa/NPa NPa S\NPa (NP↑\NPa)\(S\NPa) S\NPa> <

NPa NP↑\NPa<

NP↑=S/(S\NP)>

S

‘The man who was laughing went.’

135

balan yibi ba�

gul yara-

�

gu [miyanda] -

�

u -ru bura-nwoman.ABS man-ERG laugh -REL -ERG see-NFUT

NPa NPe S\NPa (NP↑\NPa)\(S\NPa) (NP↑\NPe)\(NP↑\NPa) S{\NPe, \NPa}<

NP↑\NPa<

NP↑\NPe<

NP↑=(S\NPτ)/(S\NP\NPτ)>

S\NPa>

S :and′(laugh′man′)(see′woman′man′)

‘The man who was laughing saw the woman.’

136

Sensitivity to S =P is evident in relativising non-absolutive arguments: thesame head is involved, which requires an antipassive before a relativemarker:

bayi yara [ jilwal -

�

a ] -

�

u guda-gu yanuman.ABS kick -ANTIP -REL dog-DAT go.NFUT

NPa S{\NPe, \NPa} (S\NPa)\TV (NP↑\NPa)\(S\NPa) (S\NPa)|(S\NPa) S\NPa<

S\NPa<

NP↑\NPa>

NP↑=S/(S\NP)> B

S/(S\NPa)>

S :and′(antip′(kickdog′man′))(go′man′)

‘The man who kicked the dog went.’

NB. The heads of relativisation and antipassive are lexical items, with fullinterpretation and a syntactic type.

137

Antipassive : Traditionally considered to be the ergative equivalent of pas-sive (non-subject complement becomes subject, subject becomes oblique;NB. loss of transitivity marking below).

(15)a. Na’e kai-i ’a e ika ’e he tamasi’i.

PAST eat-TRANS ABS DEF fish ERG the boy

‘The boy ate the fish.’ Tongan (Clark, 1973, p.600)

b. Na’e kai ’a e tamasi’i ’i he ika.

PAST eat ABS DEF boy OBL the fish

‘The boy ate some of the fish.’

Tongan is VSO and ergative (Otsuka, 2000). Hence common associationof ergativity with SOV languages is only a tendency; word order seems tobe orthogonal to ergativity/accusativity of verbal categories.

138

And, according to Postal (1977), antipassive also manifests itself in ac-cusative languages (e.g. French).

In a strictly lexicalised grammar, linking passives with accusativity and an-tipassives with ergativity is unnecessary, as both can be conceived as ap-plying to the lexical category of transitive verbs across the lexicon and do-ing the same thing, be it ergative or accusative.

139

Coordination : Dyirbal lacks overt conjunctions.

Topic chains in Dyirbal suggest a pivotal conjunction category; NPabscanbe the topic in a chain of clauses (Dixon, 1972, p.67):

bayi burrbula [ba�

gul gubi-

�

gu bara-n] [baji-gu ]I.ABS.TH B.ABS I.ERG.TH gubi-ERG punch-NFUT fall.down-PURP

S/(S\NPa) S$/(S$\NPe) S{\NPe, \NPa} S\NPa>

S\NPa&

S\NPa>

S :and′(punch′burrbula′gubi′)(falldown′burrbula′)

‘The gubi punched Burrbulai and (hei) fell down.’

NB. Same-case condition for ATB exceptions hold.

140

bayi yara [yanu ] [ba

�

gun yibi-

�

gu bura-n]I.ABS.TH man.ABS go.NFUT II.ERG woman-ERG see-NFUT

S/(S\NPa) S\NPa S$/(S$\NPe) S{\NPe, \NPa}>

S\NPa&

S\NPa>

S:and′(go′man′)(see′man′woman′)

‘The man went and the woman saw him.’

141

An NPerg cannot act as the topic in a topic chain, and this follows from thecategory of conjunctions and type-dependent coordination:∗

(16) ba

�

gul gubi-�

gu [ bayi burrbula bara-n] [ baji-gu ]I.ERG.TH gubi-ERG I.ABS.TH B.ABS punch-NFUT fall.down-PURP

S/(S\NPe) S$/(S$\NPa) S{\NPe, \NPa} S\NPa>

S\NPe*** &

:∗and′(punch′burrbula′gubi′)(falldown′gubi′)

* for ‘The gubi j punched Burrbulai and (he j) fell down.’

∗In principle, left-node raising of ergative NPs to share across all conjuncts should bepossible with an additional non-pivotal conjunction category, but we do not know whetherthis includes fronting of the ergative NP in Dyirbal to become a topic.

142

Dyirbal is syntactically (fully) ergative because heads of all constructionsappear to adopt the asymmetry, which, for syntactic constructions, canonly be an S -based asymmetry, by the Principle of Lexicalised Asymmetry(PLA).

143

Full ergativity : Dyirbal control.

Dyirbal seems unable to control the A role of non-finite complements (butnote the purpose clause glosses for the controlled clause); antipassive hasto apply for it to be controlled:

bayi yara wal

�

garra [bural -

�

a-ygu bagun yibi-gu ]I.ABS.TH man.ABS want-NFUT see -ANTIP-PURP II.DAT.TH woman-DAT

NPa S{\NPa, |(Sinf |NPa)} S{\NPe, \NPa} IV\TV IV|IV<

S\NPa<

S\NPa>

S\NPa<

S :want′(antip′(seewoman′ (ana′man′)))man′

‘The man wanted to see the woman.’

144

S and P can be controlled; the control verb takes as argument an infini-tival VP in which the residual NP is not necessarily the agent NP but thesyntactic subject.

(17)a. wal

�

garra := S{\NPabs, |(Sinf|NPabs)} :λ{x1,P}.want′ (P(ana′ x1))x1

b. giga := S{\NPerg, |(Sinf|NPabs), \NPabs} : λ{x1,P,x2}.tell′ (P(ana′ x2))x2x1

bayi yara walngarra [�

aba-ygu ]I.ABS.TH man.ABS want-NFUT bathe-PURP

NPa S{\NPa, |(Sinf |NPa)} S\NPa>

S\NPa<

S :want′(bathe′ (ana′man′))man′

‘The man wanted to bathe.’

145

bayi yara wal

�

garra [ba

�

gun yibi-

�

gu bura-li ]I.ABS.TH man.ABS want-NFUT II.ERG.TH woman-ERG see-PURP

NPa S{\NPa, |(Sinf|NPa)} NPe S{\NPe, \NPa}<

S\NPa>

S\NPa<

S :want′(see′ (ana′man′)woman′)man′

‘The man wanted the woman to see him.’

NB. Although the English gloss does not correspond to control, this is aninstance of control in Dyirbal, because ‘man’ is a semantic argument ofboth clauses and necessarily missing from the subordinate clause.

146

This is in sharp contrast to all accusative languages, which seem alwaysto be able to control 1, and to other ergative languages, which can alsocontrol 1s.

Eg., Previous example is ruled out in English; non-subject NPs (/NP) doesnot align with 1.

*John persuaded him [Sue see]

NP (S\NP)/(Sinf\NP)/NP NP NP (S\NP)/NP> > T

(S\NP)/(Sinf\NP) S/(S\NP)> B

S/NP*** *

147

Manning (1996) attributes Dyirbal’s apparent lack of A -control to such con-structions being chains of clauses, therefore a syntactic construction, ratherthan being a lexical property of the verb as in control, which would explaintheir strict S =P sensitivity.

But clause chaining in Dyirbal is restricted to NPabs as Dixon (1972:67)noted.

148

(18b) should be problematic under clause-chaining account:

(18)a. yabu�

uma-

�

gu giga-n [banaga-ygu ]

mother.ABS father-ERG tell-NFUT return-PURP

‘Father told mother [to return].’

b.

�

uma-

�

gu yabu giga-n [banaga-ygu ]

father-ERG mother.ABS tell-NFUT return-PURP

‘Father told mother [to return].’

149

Thus, Dyirbal’s infinitival VP category differs from that of other ergativelanguages, and from accusative languages, in a way that can be predictedby PLA:

Dyirbal inf. VP: Sinf\NPabs: λx. · · ·x · · ·

Inuit inf. VP: Sinf\NPabs: λx. · · ·x

English inf. VP: Sinf\NP: λx. · · ·x · · ·

Turkish inf. VP: Sinf\NPnom: λx. · · ·x · · ·

Since infinitivalisation is a lexical process, PLA applies to the lexical cate-gory of its head, hence it can have a 1-asymmetry or S -asymmetry.

In case of accusative languages, 1 aligns with S and A , thus reading con-trol off LF or syntactic type is the same (i.e. x above is always 1 for Englishand Turkish), but it must be enforced in Inuit, Basque, Tagalog and manyothers if Dyirbal is not the only exception to universality of 1-control.

150

Lexicalising the VPinf category suggests the possibility of having no controlin a language.

Mayali (Evans, 1991), Mohawk (Baker, 1996) and Nunggubuyu (Heath,1975) seem to have no control.

Baker (1996) shows convincingly that languages that lack infinitival clausesare likely to have no control. This is not a necessity but a strong tendency.

It is not a necessity because, assuming the conventional definition of aninfinitival clause for the moment in that they lack syntactic subjects, lackof infinitivals means no S -control in our terminology, but 1-control wouldstill be possible as in ergative languages that exhibit accusative patterns incontrol.

151

Partial ergativity : Inuit control.

A and S can be controlled, but not S ; P below can only be of type λx. · · ·x

(19)a. niriursui := S{\NPabs, \VPinf} : λ{x1,P}.promise′ (P(ana′ x1))x1

b. Miiqqat [Juuna ikiu-ssa-llu-gu ] niriursui-pp-u-tchildren.ABS J.ABS help-FUT-INF-3SG promise-IND-INTR-3PL

NPa NPa S{\NPe, \NPa} S{\NPa, \VPinf}<

S\NPe<

S\NPa

<S:promise′(help′ juuna′(ana′ children′))children′

‘The children promised to help Juuna.’

152

c. Miiqqat [qiti-ssa-llu-tik ] niriursui-pp-u-tchildren.ABS dance-FUT-INF-4PL promise-IND-INTR-3PL

NPa S\NPa S{\NPa, \VPinf}<

S\NPa<

S:promise′(dance′(ana′ children′))children′

‘The children promised to dance.’

Partial ergativity : Inuit syntax.

Some syntactic constructions follow the ergative pattern, ie. the S =P rela-tion is singled out as the pivot, but not all.

Relative markers are pivots:

(20)a. nanuq [Piita-p tugu ] -ta-apolar.bear.ABS P-ERG kill -TR.PART-3SG

NPa NPe S{\NPe, \NPa} (NP↑\NPa)\(S\NPa)<

S\NPa<

NP↑\NPa<

NP↑ :λQ.and′(kill′pbear′piita′)(Qpbear′)

‘A polar bear killed by Piita.’‘A polar bear that Piita killed’

153

b. miiraq [kamat ] -tu-qchild.ABS angry -REL.INTR-SG

NPa S\NPa (NP↑\NPa)\(S\NPa)<

NP↑\NPa<

NP↑ :λQ.and′(angry′ child′)(Qchild′)

‘The child that is angry.’

c. *angut [aallaat tigu-sima] -sa-aman.ABS gun.ABS take-PRF -REL.TR-3SG.SG

NPa NPa S{\NPe, \NPa} (NP↑\NPa)\(S\NPa)<

S\NPe*** <

*‘the man who took the gun’

Non-pivotal coordination: Fortescue (1984:131) reports that, in contem-porative mood (ie. true coordination), two absolutive NPs share acrossconjuncts if they are both P or S , but not S =P (NB. passive morphologyon the second conjunct, therefore a missing NPabs):

Hansi [ isir-puq ]S\NPabs[Kaala-mil =lu taku-niqar-luni ]S\NPabs

H come-in.3S.INDIC K-ABL and see-PASSIVE-4S.CONT

‘Hansi came in and was seen by Kaala.’

‘Hansi [came in]S\NPabsand [Kaala saw (him)]S\NPabs

’

would be grammatical only in non-contemporative moods (ie. non-coordinatingsubordination).

154

Inuit is partially ergative because not all constructions follow the ergativepattern, but some do. In Basque, no construction follows the ergative pat-tern (no head adopts the asymmetry although verbal categories are erga-tive).

155

Morphological ergativity : Basque syntax and control.

Basque’s control is similar to Inuit’s; only 1s can be controlled (as in ac-cusative languages, cf. Anderson 1976).

Unlike accusative languages, the S and A arguments corresponding to 1do not bear the same case in Basque.

And, unlike Inuit, unbounded dependencies engendered by heads of coor-dination and relativisation do not seem to arise from asymmetric syntactictypes.

This is a possibility predicted by the Principle of Lexical Head Governmentand the Principle of Lexicalised Asymmetry: heads are free to adopt orignore the asymmetry as lexical items, independent of other heads’ be-haviour.

156

The same relative marker, -n, is used to relativise all arguments of a verb;the marker has a non-pivotal category:

(21) a. emakume-a-ri liburu-a eman dio -n gizon -awoman-the-DAT book-the give has -REL man -the

NPd NPa S{\NPe, \NPa, \NPd} (S\NPτ)\(S\NPτ) (N/N)\(S\NP) N NP↑\N<

S{\NPe, \NPd}<

S\NPe<

S\NPe<

N/N>

N<

NP↑

‘The man who has given the book to the woman’

157

b. gizon-a-k liburu-a eman dio -n emakume -aman-the-ERG book-the give has -REL woman -the

NPe NPa S{\NPe, \NPa, \NPd} (S\NPτ)\(S\NPτ) (N/N)\(S\NP) N NP↑\N<

S{\NPe, \NPd}<

S\NPd<

S\NPd<

N/N>

N<

NP↑

‘The woman to whom the man has given the book’

Gizon-a-k emakume-a-ri liburu-a eman dio.

man-the-ERG woman-the-DAT book-the.ABS give has

‘The man has given the book to the woman.’ Comrie (1989:141–2)

emakume-a-ri liburu-a eman dio-n gizon-a

woman-the-DAT book-the give has-REL man-the

‘The man who has given the book to the woman’

gizona-k emakume-a-ri eman dio-n liburu-a

man-the-ERG woman-the-DAT give has-REL book-the

‘The book which the man has given to the woman’

gizon-a-k liburu-a eman dio-n emakume-a

man-the-ERG book-the give has-REL woman-the

‘The woman to whom the man has given the book’

158

Basque coordination without pivots: Example below is considered an in-stance of conjunction reduction by Manning (1996).

Ama-k seme-a eskola-n utzi (zuen) eta klase-ra joan (zen)

mother-ERG son-ABS school-at leave AUX.3SG.3SG and class-to go AUX.3SG

‘Mother left her son at school and went to class.’

159

It is not only the S and the A NP, but the P NPs (22a) and S and P NPscan be co-indexed under coordination as well (22b–c).

(22) a. Sagarrak emakumeak egosten ditu eta gizonak jaten dituApples-ABS woman-ERG cook AUX.3SG.3PL and man-ERG eat AUX.3SG.3PL‘The woman cooks and the man eats apples.’ Alan King (p.c.)

b. Sagarrak emakumeak egosten ditu eta usteltzen diraApples-ABS woman-ERG cook AUX.3SG.3PL and rot AUX.3PL‘The woman cooks apples and they rot.’

c. Sagarrak usteltzen dira eta emakumeak egosten dituApples-ABS rot AUX.3PL and woman-ERG cook AUX.3SG.3PL‘Apples rot and the woman cooks them.’

160

The choice of auxiliaries in (22b–c) show that any syntactic argument canbe deleted in Basque, and the auxiliaries provide a constrained way ofidentifying them.

That can be made explicit in the conjunction’s category: it will be

((S\NPα)\?(S\NPα))/?(S\NPβ)

to take care of case and agreement under coordination.

This is a non-pivotal category (α and β are not restricted to a combinationof S , A and P ).

There seems to be no head in the Basque lexicon with a pivotal category,which would explain Basque’s “apparent” ergativity.

161

Split ergativity : Dyirbal’s pronouns.

nana banaga-nyu

we.PL.NOM return-NFUT

‘We returned.’

nyurra nana-na bura-n

you.PL.NOM we.PL-ACC see-NFUT

‘You all saw us.’

nana nyurra-na bura-n

we.PL.NOM you.PL-ACC see-NFUT

‘We saw you all.’

nyurra nana-na bura-n banaga-nyu

[you.PL.NOM we.PL-ACC see-NFUT] [return-NFUT]

‘You all saw us and (we) returned.’

162

The operation of Dyirbal’s split morphology might suggest an ergative sys-tem for nominals, and an accusative system for pronominals.

The accusative system corresponds to 1 binding of the nominative pro-noun (23a–b) and 2 binding of the accusative pronoun (23c), as in ac-cusative languages.

(23)a. nyurra := S/VP: λP.Pyou′

b. nana := S/VP: λP.Pwe′

c. nana-na:= VP/TV= (S{\NPerg})/(S{\NPerg, \NPabs}): λ{y,P}.Pus′ y

163

(24) nyurra nana-na bura-n banaga-nyuyou.PL.NOM we.PL-ACC see-NFUT return-NFUT

S/VP VP/TV S{\NPabs, \NPerg} S\NPabs> pd

VP S>

S&

S :and′(see′us′ you′)(return′ pro)for ‘You all saw us and (we) returned.’

Dyirbal’s split behaviour in morphology can be accounted for by the lexicalcategories of the pronominals, without a need for another set of verbalcategories for case marking and binding of pronominal arguments (ie. itsverbal categories are all ergative).

Dyirbal can be characterised as a split ergative language with fully ergativesyntax.

164

Mixed ergativity : Liangshan Nuosu (see Gerner 2004).

The P A V order is ergative, and A P V order is accusative in Nuosu.

Gerner claims P /A identity in P A V & P V coordination (his ex.37)

This involves a pivot because P in the first conjunct is part of S =P ; as theword-order in the first conjunct suggests, it is an ergative clause

If Nuosu can only delete clause-initial positions as Gerner claims, then thesecond conjunct is an accusative clause due to the ordering [A ]P V (withsubject relation S =A ).

This is merely an ergative subject in an ergative clause co-referring with anaccusative subject in an accusative clause, that is, a pivot.

165

Skewness in the ergative-accusative language spectrum

There are clearly global factors that bias languages towards some ordersrather than others.

Earliness in the sentence seems to correlate with various kinds of promi-nence, including topic- or focus- prominence, agent or 1-prominence, andsubject or S -prominence.

Languages that are extremely Topic/Focus-prominent, such as the Slaviclanguages, are thereby committed to some degree of order-freedom at thesyntactic and semantic levels.

166

(The labels S and O below are closer to typological tradition, similar to 1and 2 of CCG):

For accusative languages, in which the nominative subject and the topicusually coincide, the SO orders (SOV, SVO, VSO) maximise 1-prominence,S -prominence, and topic-prominence,

While the OS orders (OSV, OVS, VOS) minimise all three.

Since the accusative languages appear to considerably outnumber theergative, this fact alone may explain the preponderance of SO orders amongconfigurational languages in general, as Manning (1996:22) has suggested.

167

For ergative languages however, SO orders maximise only 1-prominence,minimising S -prominence.

Since the absolutive S in ergative languages is often claimed to coincidewith topic in much the same sense as nominative does in accusative lan-guages (Mallinson and Blake 1981, 103-115, 155-158):

SO orders may also minimise Topic-prominence as well.

On the other hand, while OS orders minimise 1-prominence, they maximiseS -prominence, and by implication Topic-prominence.

One might therefore expect the ergative languages to split across SO andOS linearization, depending on whether they favor 1-prominence over S /Topic-prominence.

168

Case and Agreement: Icelandic case

As syntactic information, case and agreement can only be part of the syn-tactic type of a lexical item.

By the Principle of Lexical Head Government, the syntactic type of thehead should enforce agreement, regardless of the (un)boundedness of theconstruction it heads:

The man who likes/*like chocolate

The man who I think that you claimed likes/*like chocolate

The embedded verb projects, and the relative pronoun enforces, agree-ment.

Similarly, case must be regulated by the syntactic type of the head.

169

The Icelandic challenge (for the theory) is that the notion of subject seemsto be seriously dissociated from morphological case; subjects are leftward-looking functors independent of their case (Icelandic is considered to beSVO, and subjects can be nominative, accusative, dative or genitive).

And, agreement seems to be dissociated from nominative case and direc-tionality of the NP (ie. from its CCG equivalent of “structural” case):

dative subjects do not agree with the verb, but nominative subjects do.

accusative subjects agree, just like nominative subjects.

170

Stelpunum brá við.

the-girls (DAT-FEM-PL) startled (SG)

‘The girls were astonished.’ (Andrews, 1990, p.189–190)

Ég tel strákana (hafa verið) kitlaða.

I believe the-boys (ACC-MASC-PL) (to have been) tickled (ACC-MASC-PL)

‘I believe the boys to have been tickled.’

This is a possibility envisaged (though not explained) by CCG: directionalityof an argument is orthogonal to its syntactic features, and both are lexicallydiscernible.

171

Thus, if a verb takes an infinitival VP, ie. Sinf\NP, it can ignore morpholog-ical case and agreement of the residual NP, in the residue Sinf\NPCase,Agr,although that NP is LF-coindexed with the subject of the matrix verb. (thereis no threading of syntactic and semantic features via co-indexation as inHPSG).

Stelpan/*Stelpuna [vonast til ](S\NPnom/(Sinf\NP) [að vanta ekki efni í ritgerðina]S\NPaccthe-girl-NOM/*ACC hopes toward to lack not material in the-paper

‘the girl hopes to not lack material for the paper.’

S: hope′(not′lack′material′ (ana′girl′))girl′

172

Icelandic control .∗

All cases can be controlled in Icelandic, but only if they are also syntacticsubjects in the above sense: the leftward NP (cf. Zaenen et al. 1990, p.106;Andrews 1990, p.198 ).

a. Stelpuna vantar efni í ritgerðina

the-girl-ACC lacks material in the-paper‘The girl lacks material for the paper.’

b. Stelpan/*Stelpuna vonast til að vanta ekki efni í ritgerðina

the-girl-NOM/*ACC hopes toward to lack not material in the-paper‘the girl hopes to not lack material for the paper.’

∗cf. Maling and Zaenen 1990 for extensive discussion of its syntax, and Bozsahin andSteedman 2005 for a CCG formulation of some aspects of coordination in Icelandic.

173

Thus the category of the verb vantar ‘lacks’ is as in (25a), rather than(25b):

(25) a. vantar:= (S\NPacc,3sing)/NPnom

b. lacks := (S\NPnom,3sing)/NPacc

And the category of vonast (til) ‘hopes (for)’ is

vonast := (S\NPnom)/(Sinf\NP): λPλx.hope′(P(ana′ x))x

Stelpan/*Stelpuna vonast til [að vanta ekki efni í ritgerðina]the-girl-NOM/*ACC hopes toward to lack not material in the-paper

NPnom (S\NPnom)/(Sinf\NP) Sinf\NPacc>

S\NPnom<

S: hope′(not′lack′material′ (ana′girl′))girl′

174

The matrix verb has the option of 1) ignoring case and agreement comingfrom the residue and imposing its own requirement as in the control verbabove, or 2) adopting them.

The second option might be the Inuit strategy, rather than the verb incor-poration analysis of Bok-Bennema (1991:169):

Hansi [sinik]S\NPnom -kuma-vuq

H-NOM sleep -want-IND.3SG(INTR)

‘Hansi wants to sleep.’ (Bok-Bennema, ibid:222)

Hansi-p [qajaq atur]S\NPgen -uma-vaa

H-GEN kayak-NOM borrow -want-IND.3SG.3SG(TR)

‘Hansi wants to borrow the kayak.’

Bok-Bennema’s -GEN and -NOM can be glossed as -ERG and -ABS.

175

NB. embedded verb is incorporated in both examples.

Unlike Basque, the case of control verb’s subject appears to depend on thetransitivity of the controlled verb.

Nik [ joan ]S\NPabsnahi dut

I-ERG go.INF want have.1SG.3SG

‘I want to go.’

Nik [kafea egin ]S\NPerg nahi dut

I-ERG coffee-ABS do.INF want have.1SG.3SG

’I want to make coffee.’

The matrix verb cannot impose its own requirements on the embeddedverb, this category violates Projection Principles (Principles of Inheritanceand Lexical Head Government), and it is not formulable in a strictly lexi-calised grammar with no record of derivations.

176

Thus, morphological case, structural case and agreement can diverge orconverge, because, as syntactic information embodied in a lexical syntactictype, the choice is up to the head of the construction.

The only locus for their origin is the lexical syntactic type of argument-taking entities, eg. verbal categories, to be projected by a completelytransparent and monostratal universal combinatory syntax (the ‘syntacticprojector’).

In a lexicalised domain of locality, the possibilities for the heads are pre-dictable, and not unlimited;

and they seem to have been attested cross-linguistically.

177

Summary

CCG attempts to provide

1) a framework (or toolbox) to explicitly formulate hypotheses in the form ofsyntactic-semantic type pairings,

2) a theory (of grammar-lexicon) to narrow down the notion of “possiblelexical category”, therefore possible grammars,

3) ... and a somewhat user-friendly computational environment to unleashthe ‘syntactic projector’ and hypotheses on real data.

178

Appendix:

Tagalog: Ergative? Accusative? Both? Neither?

In a lexicalised grammar where verbal categories project all lexical restric-tions on combinatory modalities and syntactic information such as caseand agreement, all of these options are possible.

179

Tagalog’s voice system, and cases.∗

NOM GEN DAT

Common noun markers : ang ng sa

Personal name markers: si ni kay

∗All data in this section are from Kroeger 1993.

180

B-um-ili ang=lalake ng=isda sa=tindahan

PERF.AV-buy NOM=man GEN=fish DAT=store

‘The man bought (*the) fish at the store.’

B-in-ili ng=lalake ang=isda sa=tindahan

PERF-buy-OV GEN=man NOM=fish DAT=store

‘The man bought the fish at the store.’

B-in-ilh-an ng=lalake ng=isda ang=tindahan

PERF.AV-buy GEN=man GEN=fish NOM=store

‘The man bought fish at the store.’

181

Binding possibilities are such that less oblique arguments can bind moreobliques, independent of case:

Nag-iisip sila sa=kanilang sarili

AV-think.about NOM.they DAT=their self

‘They think about themselves.’

* Iniisip sila ng kanilang sarili

DV.think.about NOM.they GEN their self

‘They think about themselves.’

Sinaktan ng=babae ang=kaniyang sarili

DV.hurt GEN=woman NOM=her self

‘A/the woman hurt herself.’

182

Voice system of Tagalog seems to set up the following correspondence tosemantic arguments:

Voice

1 2 3

AV NOM GEN/DAT

NOM GEN DAT

DV GEN NOM

GEN GEN NOM

OV GEN NOM DAT

183

If we replace NOM with ABS, and GEN with ERG, we get

Voice

1 2 3

AV ABS ERG/DAT

ABS ERG DAT

DV ERG ABS

ERG ERG ABS

OV ERG ABS DAT

Objective Voice is considered more basic than Active Voice in Tagalogstudies.

184

Every voiced verbal clause must have an ang-NP, ie. absolutive.

But glossing Tagalog as an ergative language needs further support.

Tagalog’s relativisation is pivotal, headed by the linker ng: Only ang-markedNPs can be relativised, independent of their LF role:

*Iyon ang=baro=ng [b-um-ili ang=babae ]S/NPergthat ABS=dress=LNK PERF.AV-buy ABS=woman

* for ‘That’s the dress that a/the woman bought.’

*Iyon ang=babae=ng [b-in-ili ang=baro ]S/NPergthat ABS=woman=LNK PERF-buy.OV ABS=dress

* for ‘That’s a/the woman who bought a dress.’

185

In OV, the patient can extract, although the agent cannot:

Iyon ang=baro=ng [b-in-ili ng=babae ]S/NPabs

that ABS=dress=LNK PERF-buy.OV ERG=woman

‘That’s the dress that a/the woman bought.’

If linker (relative marker) is a pivot in the lexicon, these facts follow. And itwill be the kind of pivot that we see in ergative languages:

186

Iyon ang baro ng b-in-ili ng=babaeThat ABS dress LNK PERF-buy.OV ERG=woman

S/NP NPa↑/¦(S/NP) S/NP (S/NP){\?(S/NP), /?(S/NPa)} S{/NPe, /NPa} NPe>

S/NPa>

(S/NP)\?(S/NP)<

S/NP>

NPa↑=S\(S/NPa)<

S

‘That’s the dress that a/the woman bought.’

187

Coordination does not seem to be asymmetric: any NP can be sharedacross conjuncts:

[Huhugasan ko ] at [pupunasan mo ] ang=mga=pinggan

FUT-wash-DV 1.SG.ERG and FUT-dry-DV 2.SG.ERG ABS=PL=dish

‘I will wash and you dry the dishes.’

?*[Niluto ang=pagkain ] at [hinugasan ang=mga=pinggan ] ni=Josie

PERF-cook-OV ABS=food and PERF-wash–DV ABS=PL=dish ERG=J

for ‘The food was cooked and the dishes washed by Josie.’

?*[Nanghuhuli ang=ama ko ] at [nagtitinda ang=ina ko ] ng=isda

AV(antierg).IMPERF-catch ABS=father my and AV.IMPERF(antierg)-sell ABS=mother my ERG=fish

for ‘My father catches and my mother sells fish.’

188

[Nagbigay ng=regalo si=Maria] at

PERF.AV(antierg)-give ERG=present ABS=M and

[nagpadala ng=Liham ang=mga=bata] kay=Juan

PERF.AV(antierg)-send ERG=letter ABS=PL=child DAT=J

‘Maria gave a present and the children sent a letter to Juan.’

Tagalog seems to be a partially ergative language because there are pivotsin the lexicon (of ergative kind; ie. missing absolutive NPs).

189

Foley (1998) considers Tagalog to be a third type distinct from ergativeand accusative patterns. Machlachlan and Nakamura agree on Tagalog’sergativity, and so do Payne (1982) and de Guzman (1978), but Guilfoyle etal.(1992) consider it accusative.

Bozsahin and Steedman (2005) claim that there is no need to posit anaccusative sub-system for Tagalog due to AV voice; so it is not mixed erga-tive.

190

A strictly lexicalised grammar equipped with PLA predicts that, if subjectis not discernible in the lexical syntactic type of a verb, there can be noasymmetry.

This seems to be borne out in Tagalog: verbs in recent past do not carryvoice, and none of their arguments are ang-marked (Baldridge, 2004).

191

And, any argument can extract:

Kabibili lang ni Juan ng tela

buy-RECPAST just ERG J ERG cloth

‘Juan has just bought some cloth.’

Sino ang kabibili lang ng tela?

who ABS buy-RECPAST just ERG cloth

‘Who has just bought some cloth?’

Ano ang kabibili lang ni Juan?

what ABS buy-RECPAST just ERG J

‘What has Juan just bought?’

192

*References

Ades, A. E. and Steedman, M. (1982). On the order of words. Linguistics and Philosophy ,4, 517–558.

Anderson, S. (1976). On the notion of subject in ergative languages. In C. Li, editor,Subject and Topic. Academic Press.

Andrews, A. (1990). Case structures and control in modern Icelandic. In J. M. Maling andA. Zaenen, editors, Modern Icelandic Syntax , volume 24 of Syntax and Semantics,pages 187–234. Academic Press, San Diego.

Bach, E. (1979). Control in Montague grammar. Linguistic Inquiry , 10, 513–531.

Baker, M. C. (1996). The Polysynthesis Parameter . Oxford University Press.

Baldridge, J. (2002). Lexically specified derivational control in Combinatory CategorialGrammar . Ph.D. thesis, University of Edinburgh.

Baldridge, J. (2004). Extraction asymmetries in English, Tagalog, and Toba Batak. ms.,University of Edinburgh.

Blake, B. J. (1990). Relational Grammar . Routledge, London.

Bok-Bennema, R. (1991). Case and Agreement in Inuit . Foris, Berlin.

Bozsahin, C. (2002). The combinatory morphemic lexicon. Computational Linguistics,28(2), 145–176.

Bozsahin, C. (2004). On the Turkish controllee. In Proc. of 12th Int. Conf. on TurkishLinguistics (ICTL), Izmir, Turkey.

Bozsahin, C. (2005). Word order, word order flexibility and the lexicon. In S. Ozsoy, editor,Theoretical Issues in Word Order . Kluwer. forthcoming.

Bozsahin, C. and Steedman, M. (2005). Lexicalized asymmetry and syntactic projection.ms., University of Edinburgh.

Bozsahin, C., Kruijff, G.-J. M., and White, M. (2005). Specifying Grammars for OpenCCG:A Rough Guide.

Chomsky, N. (1965). Aspects of the Theory of Syntax . MIT Press, Cambridge, MA.

Chomsky, N. (1995). The Minimalist Program. MIT Press, Cambridge, Mass.

Clark, R. (1973). Transitivity and case in Eastern Oceanic languages. Oceanic Linguistics,12, 559–606.

Comrie, B. (1989). Language Universals and Linguistic Typology . The University ofChicago Press. 2nd ed.

Curry, H. B. and Feys, R. (1958). Combinatory Logic I. North-Holland, Amsterdam.

Derbyshire, D. (1977). Word order universals and the existence of OVS languages. Lin-guistic Inquiry , 8, 590–598.

Dixon, R. (1972). The Dyirbal language of North Queensland . Cambridge UniversityPress.

Dixon, R. (1979). Ergativity. Language, 55, 59–138.

Dixon, R. (1994). Ergativity . Cambridge University Press.

Dowty, D. (1996). Nonconstituent coordination, wrapping, and multimodal categorial gram-mars: Syntactic form as logical form. In K. Doets, editor, Proceedings of the Tenth In-ternational Congress of Logic, Methodology, and Philosophy of Science, 1995. NorthHolland, Amsterdam.

Epstein, S. D., Groat, E., Kawashima, R., and Kitahara, H. (1998). A Derivational Ap-proach to Syntactic Relations. Oxford University Press, Oxford.

Evans, N. (1991). A draft grammar of Mayali. ms., University of Melbourne.

Foley, W. A. (1998). Symmetrical voice systems and precategoriality in Philippine lan-guages. In LFG98 Conference, Workshop on Voice and Grammatical Functions inAustronesian Languages, The University of Queensland, Australia.

Fortescue, M. (1984). West Greenlandic. Croom Helm, Sydney.

Gazdar, G. (1981). Unbounded dependencies and coordinate structure. Linguistic In-quiry , 12, 155–184.

Gerner, M. (2004). On a partial, strictly word-order based definition of grammatical rela-tions in Liangshan Nuosu. Linguistics, 42(1), 109–154.

Greenberg, J. H. (1963). The languages of Africa. Int. Journal of American Linguistics,29(1), part II. publ. by The Hague: Mouton, for Indiana University.

Grosz, B., Joshi, A., and Weinstein, S. (1995). Centering: A framework for modeling thelocal coherence of discourse. Computational Linguistics, 2, 203–225.

Guilfoyle, E., Hung, H., and Travis, L. (1992). Spec of IP and spec of VP: Two subjects inAustronesian languages. Natural Language and Linguistic Theory , 10, 375–414.

Guzman, V. (1978). Syntactic Derivation of Tagalog Verbs. University of Hawaii Press,Honolulu. Oceanic Linguistics Special Publication no.16.

Heath, J. (1975). Functional relationships in grammar. Language, 51, 89–104.

Hoffman, B. (1995). The Computational Analysis of the Syntax and Interpretation of “Free”Word Order in Turkish. Ph.D. thesis, University of Pennsylvania.

Jacobson, P. (1996). The syntax/semantics interface in Categorial Grammar. In S. Lappin,editor, The Handbook of Contemporary Semantic Theory . Blackwell.

Jacobson, P. (1999). Towards a variable-free semantics. Linguistics and Philosophy , 22,117–184.

Karttunen, L. (1989). Radical lexicalism. In M. Baltin and A. Kroch, editors, AlternativeConceptions of Phrase Structure. University of Chicago Press, Chicago.

Kayne, R. (1994). The Antisymmetry of Syntax . MIT Press.

Kroeger, P. (1993). Phrase Structure and Grammatical Relations in Tagalog. CSLI, Stan-ford, CA.

Lambek, J. (1958). The mathematics of sentence structure. American MathematicalMonthly , 65, 154–170.

Maling, J. M. and Zaenen, A., editors (1990). Modern Icelandic Syntax , volume 24 ofSyntax and Semantics. Academic Press, San Diego.

Mallinson, G. and Blake, B. (1981). Language Typology . North Holland, Amsterdam.

Manning, C. D. (1996). Ergativity: Argument structure and grammatical relations. CSLI,Stanford, CA.

Moortgat, M. (1997). Categorial type logics. In J. van Benthem and A. ter Meulen, editors,Handbook of Logic and Language. MIT Press.

Moravcsik, E. A. (1978). On the distibution of ergative and accusative patterns. Lingua,45, 233–279.

Morrill, G. (1994). Type-Logical Grammar . Kluwer, Dordrecht.

Oehrle, R. T. (2000). Multi-modal type logical grammar. In R. Borsley and K. Borjars, ed-itors, Non-Transformational Syntax: A Guide To Current Models, page in preparation.Blackwell, Oxford.

Otsuka, Y. (2000). Ergativity in Tongan. Ph.D. thesis, Oxford University.

Palmer, F. (1994). Grammatical Roles and Relations. Cambridge University Press.

Partee, B. H. and Rooth, M. (1983). Generalized conjunction and type ambiguity. InR. Bauerle, C. Schwarze, and A. v. Stechow, editors, Meaning, Use, and Interpretationof Language. de Gruyter, Berlin.

Payne, T. E. (1982). Role and reference related subject properties and ergativity in Yup’ikEskimo and Tagalog. Studies in Language, 6, 75–106.

Postal, P. (1977). Antipassive in French. Linguisticæ Investigationes, 1, 333–374.

Reinhart, T. and Siloni, T. (2004). Against an unaccusative analysis of reflexives. InA. Alexiadou, E. Anagnostopoulou, and M. Everaert, editors, The Unaccusativity Puz-zle. Oxford Univ. Press.

Ross, J. R. (1967). Constraints on Variables in Syntax . Ph.D. thesis, MIT. Published asInfinite Syntax!, Ablex, Norton, NJ, 1986.

Ross, J. R. (1970). Gapping and the order of constituents. In M. Bierwisch and K. E.Heidolph, editors, Progress in Linguistics, pages 249–259. Mouton, The Hague.

Steedman, M. (2000). The Syntactic Process. MIT Press, Cambridge, MA.

Steedman, M. (2002). Plans, affordances, and combinatory grammar. Linguistics andPhilosophy , 25, 723–753.

Steedman, M. (2005a). Interfaces and the grammar. In Proceedings of the 24th WestCoast Conference on Formal Linguistics, Vancouver, Canada.

Steedman, M. (2005b). Semantics in categorial grammar. ms., University of Edinburgh.

Steele, S. (1978). Word order variation: A typological study. In J. Greenberg, editor,Universals of Human Language. Stanford University Press.

Szabolcsi, A. (1989). Bound variables in syntax: Are there any? In R. Bartsch, J. van Ben-them, and P. van Emde Boas, editors, Semantics and Contextual Expression, pages295–318. Foris, Dordrecht.

Szabolcsi, A. (1992). Combinatory grammar and projection from the lexicon. In I. A. Sagand A. Szabolcsi, editors, Lexical Matters. CSLI, Stanford.

Webber, B. (2004). D-LTAG: Extending lexicalized TAG to discourse. Cognitive Science,28, 751–779.

Williams, E. (1978). Across-the-board rule application. Linguistic Inquiry , 9, 31–43.

Yüksel, O. and Bozsahin, C. (2002). Contextually appropriate reference generation. Nat-ural Language Engineering, 8(1), 69–89.

Zaenen, A., Maling, J. M., and Thráinsson, H. (1990). Case and grammatical functions:the Icelandic passive. In J. M. Maling and A. Zaenen, editors, Modern Icelandic Syn-tax , volume 24 of Syntax and Semantics, pages 95–136. Academic Press, San Diego.

combinatory categorial grammar and linguistic...

Documents