jackendoff, r. 1975. regularities in the lexicon

7/22/2019 Jackendoff, R. 1975. Regularities in the Lexicon

1/34

Linguistic Society of America

Morphological and Semantic Regularities in the LexiconAuthor(s): Ray JackendoffSource: Language, Vol. 51, No. 3 (Sep., 1975), pp. 639-671Published by: Linguistic Society of AmericaStable URL: http://www.jstor.org/stable/412891.

Accessed: 26/05/2011 02:40

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at.http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless

you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you

may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at.http://www.jstor.org/action/showPublisher?publisherCode=lsa..

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed

page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of

content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms

of scholarship. For more information about JSTOR, please contact [email protected].

Linguistic Society of Americais collaborating with JSTOR to digitize, preserve and extend access toLanguage.

http://www.jstor.org
http://www.jstor.org/action/showPublisher?publisherCode=lsahttp://www.jstor.org/stable/412891?origin=JSTOR-pdfhttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/action/showPublisher?publisherCode=lsahttp://www.jstor.org/action/showPublisher?publisherCode=lsahttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/stable/412891?origin=JSTOR-pdfhttp://www.jstor.org/action/showPublisher?publisherCode=lsa


2/34

MORPHOLOGICAL AND SEMANTIC REGULARITIES IN THE LEXICONRAY JACKENDOFF

BrandeisUniversityThis paper proposes a theory of the lexicon consistent with the LexicalistHypoth-esis of Chomsky's 'Remarks on nominalization' (1970). The crucial problem is todevelopa notion of lexicalredundancyruleswhichpermitsan adequatedescriptionofthe partial relations and idiosyncrasy characteristic of the lexicon. Two lexicalisttheories of redundancyrules, each equipped with an evaluation measure,are com-pared on the basis of their accounts of nominalizations;the superiorone, the FULL-

ENTRY THEORY, is then applied to a range of further well-known examples such ascausative verbs, nominal compounds, and idioms.The startingpoint of the LexicalistHypothesis, proposed in Chomsky's 'Remarkson nominalization' (1970), is the rejection of the position that a nominal such asBill's decision to go is derived transformationally from a sentence such as Billdecided to go. Rather, Chomsky proposes that the nominal is generatedby the baserules as an NP, no S node appearing in its derivation. His paper is concerned withthe consequences of this position for the syntactic component of the grammar.Thepresent paper will develop a more highly articulatedtheory of the lexical treatmentof nominals, show that it is independentlynecessary,and extend it to a wide rangeof cases other than nominalizations.1The goal of this paper is very similar to that of Halle 1973: the presentation of aframework in which discussion of lexical relations can be made more meaningful.I will not present any new and unusual facts about the lexicon; rather, I will tryto formulate a theory which accommodates a ratherdisparaterange of well-knownexamples of lexical relations. The theory presented here, which was developed

independently of Halle's, has many points of correspondence with it; I have,however, attempted a more elaborate working out of numerous details. I willmention important differencesbetween the theories as they arise.1. LEVELSOF ADEQUACYN DESCRIPTION.n a theory of the lexicon, we candistinguish three levels of adequacy in description, parallel to those discussed byChomsky 1965 for grammatical theory. The first level consists in providing eachlexical item with sufficient information to describe its behavior in the language.This corresponds to Chomsky's level of observational adequacy, in which thegrammar is required to enumerate correctly the set of sentences in the language.A theory of the lexicon meeting the second level of adequacyexpressesthe relation-ships, sub-regularities, and generalizations among lexical items of the language,e.g. the fact that decideand decisionare related in a systematic fashion. This levelcorresponds to Chomsky's level of descriptive adequacy, which requires the1My thanksgo to JohnBowers,FrancoisDell, Noam Chomsky,MorrisHalle, and to classesat the 1969LinguisticInstitute and BrandeisUniversity or valuablediscussion.Earlierversionsof this paperwerepresentedto the 1969summerLSA meetingat the Universityof Illinois andto the 1970 LaJolla SyntaxConference.Thanks also to Dana Schaulfor many usefulexamplesscatteredthroughout the paper. 639


3/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)grammar to express correctly relationships between sentences, such as the active-passive relation.A theory of the lexicon meeting the third level of'adequacy describes how theparticular relationships and sub-regularities in the lexicon are chosen-why theobserved relationships,and not other imaginable ones, form part of the descriptionof the lexicon in question. One of the questions that must be answered at this levelis, e.g., why decide rather than decision is chosen as the more 'basic' of the tworelated items. This element of the theory takes the form of an 'evaluation measure'which assigns relative values to competing lexical descriptions available within thetheory. This is the level of explanatory adequacy.As Chomsky emphasizes, the evaluation measure does not decide betweencompeting THEORIESf the lexicon, but between competing descriptions withinthe same theory. Each theory must provide its own evaluation measure, and acomparison of competing theories must be based on their success in meeting allthree levels of adequacy.Evaluation measures have typically been built into linguistic theories implicitlyas measures of length of the grammar,i.e. its numberof symbols. One place wheresuch a measure is made explicit is in Chomsky & Halle 1968, Chapter 8. Theabbreviatoryconventions of the theory-parentheses, braces etc.-are designed soas to represent linguistically significant generalizations in terms of reduced lengthof grammatical description. Similarly, Chomsky & Halle develop the concept ofmarking conventions in order to be able to distinguish more 'natural' (i.e. explana-tory) rules from less 'natural' ones, in terms of the number of symbols needed towrite the rules.In ?2 below, I will present two theories of the lexicon compatible with theLexicalist Hypothesis. One has a traditional evaluation measure which is appliedto the number of symbols in the lexicon; the other has a more unusual measureof complexity, referring to 'independent information content'. In ?3 I will showthat the latter theory is preferable. It is hoped that such an example of a n-on-traditional evaluation measure will lead to greater understanding of the issue ofexplanatory adequacy, which has been a source of great confusion in the field.

2. FORMULATION F TWO PRELIMINARYHEORIES. he fundamental linguisticgeneralization that must be captured by any analysis of English is that words likedecision are related to words like decide in their morphology, semantics, andsyntactic patterning. For Lees 1960, it seemed very logical to express this relation-ship by assuming that only the verb decide appears in the lexicon, and by creatingthe noun decision as part of a transformational process which derives the NPJohn's decision to go from the S John decided to go. However, for reasons detailed inChomsky 1970, this approach cannot be carried out consistently without expandingthe descriptive power of transformations to the point where their explanatorypower is virtually nil.Without transformations to relate decide and decision, we need to develop someother formalism. Chomsky takes the position that decide and decisionconstitute asingle lexical entry, unmarked for the syntactic feature that distinguishes verbsfrom nouns. The phonological form decision is inserted into base trees under the

640


4/34

MORPHOLOGICALAND SEMANTIC REGULARITIES IN THE LEXICON 641node N; decide is inserted under V. Since Chomsky gives no arguments for thisparticular formulation, I feel free to adopt here the alternative theory that decideand decision have distinct but related lexical entries. In regard to Chomsky'sfurther discussion, the theories are equivalent; the one to be used here extendsmore naturally to the treatment of other kinds of lexical relations (cf. ?5). Ourproblem then is to develop a formalism which can express the relations betweenlexical entries in accord with a native speaker's intuition.2It is important to ask what it means to capture a native speaker's intuition oflexical relatedness.It makes sense to say that two lexical items are relatedif knowingone of them makes it easier to learn the other-i.e. if the two items contain lessindependent information than two unrelated lexical items do. A grammar thatexpresses this fact should be more highly valued than one that does not. Theadvocate of a transformational relationship between decide and decision claimsthat this intuitive sense of relatedness is expressed by his transformation, in thatit is unnecessary to state the shared properties of the words twice. In fact, itis unnecessary to state the properties of decision at all, since they are predictablefrom the lexical entry of decide and the nominalization transformation.3 Hencea grammarcontaining the nominalization transformationcontains less independentinformation than one without it-since instead of listing a large number ofnominalizations, we can state a single transformation. Within such a grammar,the pair decide-decisioncontains fewer symbols than a random pair such as decide-jelly: given decide, there need be no lexical entry at all for decision, but jellyneeds a lexical entry whether or not decide is listed. Furthermore, the regularityof decide-decision means that many pairs will be related by the transformation,so a net reduction in symbols in the grammaris accomplished, and the evaluationmeasure will choose a grammar including the transformation over one without it.

Since the Lexicalist Hypothesis denies a transformational relationship betweendecide and decision, their relationship must be expressed by a rule within thelexical component. Transformational grammar has for many years had a namefor the kind of rule that expresses generalizationswithin the lexicon-it is called a2 Advocates of the theory of generative semantics might at this point be tempted to claimthat a formalismfor separatebut related lexical items is yet another frill requiredby lexicalistsyntax, and that generativesemanticshas no need for this type of rule. I hastento observethatthis claim would be false. In the generativesemantic theory of lexical insertion developed inMcCawley 1968 and adopted by Lakoff 1971a,lexical items such as kill and die have separatelexical entries, and are inserted to distinct derived syntactic/semantic structures. For aconsistent treatmentof lexical insertion,then,break n Thewindowbrokemustbe insertedontoa tree of the form [vBREAK],while break in John broke the windowmust be inserted onto

[vCAUSEBREAK],which has undergonePredicateRaising; in otherwords,breakhas two distinctlexical entries. Semantically, he two breaksare relatedin exactly the same way as die and kill;but clearly break and breakmust be relatedin the lexicon in a way that die and kill are not.A similarargumentholds for rule and rulervs. rule and king.Thusgenerativesemanticsrequiresrules expressinglexical relations for exactly the same reasons that the Lexicalist Hypothesisneeds them. Only in the earlier'abstract syntax' of Lees 1960and Lakoff 1971bare such rulessuperfluous.3 Of course, it also is difficult to expressthe numerousidiosyncrasiesof nominalizations,asChomsky 1970 points out at some length.


5/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)lexical redundancyrule; but little work has been done until now toward a formal-ization of such rules.The first question we must ask is: By what means does the existence of a lexicalredundancyrule reduce the independentinformation content of the lexicon ? Thereare two possibilities. The first, which is more obvious and also more akin to thetransformational approach, gives decide a fully specified entry; but the entry fordecision is either non-existent or, more likely, not fully specified. The redundancyrule fills in the missing information from the entry of decide at some point in thederivation of a sentence containing decision,perhapsat the stageof lexical insertion.As in the transformational approach, the independent information content ofdecide-decisionis reduced, because the entry for decision does not have to be filledin. The evaluation measure again can simply count symbols in the grammar. Wemay call this theory the IMPOVERISHED-ENTRYHEORY.Within such a theory, a typical lexical entry will be of the form given below. Allaspects of this form are traditional except for the 'entry number', which is simplyan index permitting reference to a lexical entry independent of its content:

(1) rentry number/phonological representation/syntactic eaturesSEMANTIC REPRESENTATION J

For example, decidewill have the form 2. The entry number is arbitrary,and thesemantic representationis a fudge standingfor some complex of semantic markers.The NP indices correlate the syntactic arguments of the verb to the semanticarguments (cf. Jackendoff 1972, Chapter 2, for discussion of this):

(2) -784/decid/+V+[[NPi on NP2]_NP1DECIDE NNP2_

We now introduce a redundancy rule, 3, in which the two-way arrow may be readas the symmetric relation 'is lexically related to'. The rule thus can be read: 'Alexical entry x having such-and-such properties is related to a lexical entry whaving such-and-such properties.'

x - ~w 1/y + ion/ ly(3) +N ^ +( ) + [NP1's (P) NP2] + [NP1 (P) NP2]

ABSTRACT RESULT OF ACT NP1 Z NP2OF NP1'S Z-ING NP2 - L

Given the existence of 3, decisionneeds only the following lexical entry:(4) -375derived from 784by rule 3

642


6/34

MORPHOLOGICAL AND SEMANTIC REGULARITIES IN THE LEXICON 643This theory thus reduces the lexical entry for decision to a cross-reference to therelated verb plus a reference to the redundancy rule. The entries of many othernouns will be simplified similarly by the use of a referenceto 3. The independentinformation content of the lexicon can be determined straightforwardlyby addingup the information in lexical entries plus that in redundancy rules; hence theevaluation measure can be stated so as to favor grammars with fewer symbols.A second possible approach to lexical redundancy rules, the FULL-ENTRYTHEORY,assumes that both decide and decision have fully specified lexical entries, and thatthe redundancyrule plays no part in the derivation of sentences, as it does in boththe transformational theory and the impoverished-entry theory. Rather, theredundancyruleplays a role in the information measure for the lexicon. It designatesas redundantthat information in a lexical entrywhich is predictableby the existenceof a related lexical item; redundantinformation will not be counted as independent.In the full-entry theory, lexical entries again have the form of 1, except that anentry number is unnecessary. Decide has the form of 2, minus the entry number.Decision, however, will have the following entry:

(5) [/decid + ion/+N+ [NP's __ on NP2]ABSTRACT RESULT OF ACT OF

NP1'S DECIDING NP2We evaluate the lexicon as follows: first, we must determine the amount of in-dependent information added to the lexicon by introducing a single new lexicalentry; then, by adding up all the entries, we can determine the information contentof the whole lexicon.

For a first approximation, the information added by a new lexical item, given alexicon, can be measured by the following convention:(6) (Information measure)Given a fully specified lexical entry W to be introduced into the lexicon,the independent information it adds to the lexicon is(a) the information that W exists in the lexicon, i.e. that W is a wordof the language; plus(b) all the information in W which cannot be predictedby the existenceof some redundancy rule R which permits W to be partiallydescribed in terms of information already in the lexicon; plus(c) the cost of referringto the redundancy rule R.

Here 6a is meant to reflect one's knowledge that a word exists. I have no clearnotion of how important a provision it is (it may well have the value zero), but Iinclude it for the sake of completeness. The heartof the rule is 6b; this reflectsone'sknowledge of lexical relations. Finally, 6c represents one's knowledge of whichregularities hold in a particular lexical item; I will discuss this provision in moredetail in ?6.To determine the independent information content of the pair decide-decision,let us assume that the lexicon contains neither, and that we are adding them one byone into the lexicon. The cost of adding 2, since it is related to nothing yet in the


7/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)lexicon, is the information that a word exists, plus the complete information contentof the entry 2. Given 2 in the lexicon, now let us add 5. Since its lexical entry iscompletely predictable from 2 and redundancy rule 3, its cost is the informationthat a word exists plus the cost of referringto 3, which is presumably less than thecost of all the information in 5. Thus the cost of adding the pair decide-decisionis the information that two words exist, plus the total information of the entry 2,plus the cost of referringto redundancy rule 3.Now note the asymmetry here: if we add decisionfirst, then decide,we arrive ata different sum: the information that two words exist, plus the informationcontained in 5, plus the cost of referringto redundancy rule 3 (operating in theopposite direction). This is more than the previous sum, since 5 contains moreinformation than 2: the four extra phonological segments +ion and the extrasemantic information represented by ABSTRACT RESULT OF ACT OF. To establishthe independent information content for the entire lexicon, we must choose anorder of introducing the lexical items which minimizes the sum given by successiveapplications of 6. In general, the more complex derived items must be introducedafter the items from which they are derived. The information content of thelexicon is thus measured as follows:

(7) (Information content of the lexicon)Given a lexicon L containing n entries, W1, ..., Wn, each permutation Pof the integers 1, ..., n determines an order Ap in which W1, ..., W,, canbe introduced into L. For each ordering Ap, introduce the words oneby one and add up the information specified piecemeal by procedure 6,to get a sum Sp. The independent information content of the lexiconL is the least of the n sums Sp, plus the information content of theredundancy rules.

Now consider how an evaluation measure can be defined for the full-entrytheory. Minimizing the number of symbols in the lexicon will no longer work,because a grammar containing decide and decision, but not redundancy rule 3,contains fewer symbols than a grammar incorporating the redundancy rule, byexactly the number of symbols in the redundancy rule. Since we would like theevaluation measure to favor the grammar incorporating the redundancy rule, wewill state the evaluation measure as follows:

(8) (Full-entry theory evaluation measure)Of two lexicons describing the same data, that with a lower informationcontent is more highly valued.The details of the full-entry theory as just presentedare somewhat more complexthan those of either the transformational theory or the impoverished-entry theory.However, its basic principle is in fact the same: the evaluation measure is set upso as to minimize the amount of unpredictable information the speaker knows

(or must have learned). However, the measure of unpredictable information is nolonger the number of symbols in the lexicon, but the output of informationmeasure 7: this expresses the fact that, when one knows two lexical items relatedby redundancy rules, one knows less than when one knows two unrelated itemsof commensurate complexity.I will argue that the full-entry theory, in spite of its apparent complexity, is

644


8/34

MORPHOLOGICALAND SEMANTIC REGULARITIES IN THE LEXICON 645preferable to the impoverished-entrytheory. As a prelude to this argument, I willmention two other discussions of redundancyrules.The formulation of morpheme-structurerules-those redundancy rules whichpredict possible phonological combinations within the words of a language-isalso open to an impoverished-entrytheory and a full-entry theory. The former isused in Halle 1959 and in the main presentation of Chomsky & Halle, where theredundancyrules are treated as part of the re-adjustmentrules. However, Chapter8of SPE describes some difficulties in this theory pointed out by Stanley 1967. Thealternative theory presented is (I believe) a notational variant of the full-entrytheory: the redundancyrules do not play an active role in a derivation, but ratherfunction as part of the evaluation measure for the lexicon.4 If the full-entry theoryturns out to be correct for the closely related area of morpheme-structurerules, weshould be inclined to prefer it for the rules relating lexical items.Halle 1973 proposes a variant of the full-entry theory for the processes of wordformation we are concerned with here. In his theory, the redundancyrules generatea set of'potential lexical items' of the language. He then uses the feature [+ LexicalInsertion] to distinguish actual words from non-existent but possible words.A 'special filter' supplies unpredictable information, including the value of[Lexical Insertion]. The filter thus contains all the information of 6a and 6b, buthas nothing that I can relate to 6c.Consider the contents of Halle's filter,an unorderedlist of idiosyncratic informa-tion. This list must include reference to everylexical item, including all potential butnon-existent ones. It is not rule-governed-rather, it is intended to state preciselywhat is not rule-governed.It is clear why Halle sets up the lexicon in this way: he istrying to retain a portion of the lexicon where the independent information can bemeasured simply by counting features, and the filter is just such a place. Ourformulation of the information measure in the full-entry theory has freed us of thenecessity of listing the independent information separately, or of distinguishing itextrinsicallyfrom the redundantinformation. Instead we have a lexicon containingmerely a set of fully specified lexical entries (giving exactly those words that exist),plus the set of redundancyrules. (I will mention Halle's theory again briefly at theend of ?5.1.)

3. WHICHTHEORY?he argument for fully specified entries comes from con-sideration of words whose affixation is predictable by a redundancy rule, but whoseputative derivational ancestors are not lexical items of English. Examples areaggression, retribution,and fission, which have the morphological and semanticproperties of the nouns described in redundancyrule 3, but for which there are nocorresponding verbs *aggress, *retribute,or *fiss. Our intuition about these itemsis that they contain less independent information than comparable items whichcannot be partially described by a redundancy rule (e.g. demise and soliloquy),but that they contain more than comparable items which are related to genuinelexical items (e.g. decision, attribution).

4 Chomsky & Halle retainimpoverished exical entries,but only for the purposeof countingup featuresnot predictedby redundancyrules and listing what potential words actually exist.Pairedwith each impoverishedentry, however,is a fully specifiedentry,which is what actuallytakes part in the derivation.


9/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)How can the three theories we have discussed describe these verbs? The trans-

formational theory must propose a hypothetical lexical item marked obligatorilyto undergo the nominalization transformation (cf. Lakoff 1971b).Thus the lexiconmust be populated with lexical items such as *fiss which are positive absoluteexceptions to various word-formation transformations. The positive absoluteexception is of course a very powerful device to include in grammatical theory(see discussion in Jackendoff 1972). Furthermore,the use of an EXCEPTION eatureto prevent a lexical item from appearing in its 'basic' form is counter-intuitive: itclaims that English would be simpler if *fiss were a word, since one would not haveto learnthat it is exceptional. Lakoff n fact claims that there must be a hypotheticalverb *king, corresponding to the noun king as the verb rule corresponds to thenoun ruler. Under his theory, the introduction of a real verb king would makeEnglish simpler, in that it would eliminate an absolute exception feature from thelexicon. In other words, the evaluation measure for the transformational theoryseems to favor a lexicon in which every noun with functional semantic informationhas a relatedverb. Since there is little evidence for such a preference,and since it isstrongly counter-intuitivein the case of king, the transformationalaccount-besidesrequiring a very powerful mechanism, the absolute exception-is incorrect at thelevel of explanatory adequacy.Next consider the impoverished-entrytheory. There are two possible solutionsto the problem of non-existent derivational ancestors. In the first, the entry ofretribution s as unspecified as that of decision (4); and it is related by redundancyrule 3 to an entry retribute,which however is marked [-Lexical Insertion]. Thecost of adding retribution o the lexicon is the sum of the information in the entry*retribute,plus the cost of retribution's eferencesto the redundancyrule and to the(hypothetical) lexical item, plus the information that one word exists (or, morelikely, two-and the information that one of those is non-lexical). Under thereasonable assumption that the cost of the cross-referencesis less than the cost ofthe phonological and semantic affixes, this arrangement accurately reflects ourinitial intuition about the information content of retribution. Furthermore, iteliminates the use of positive absolute exceptions to transformations, replacingthem with the more restricteddevice [-Lexical Insertion]. Still, it would be nice todispense with this device as well, since it is rather suspicious to have entries whichhave all the properties of words except that of being words. The objections tohypothetical lexical items in the transformationaltheory at the level of explanatoryadequacy in fact apply here to [-Lexical Insertion] as well: the language is alwayssimpler if this feature is removed.We might propose eliminating the hypothetical lexical entries by building theminto the entries of the derived items:

(9) 511 1derived by rule 3 from/retribut/

+V+ [NP1 for NP2]

_ NP2 RETRIBUTENP2 _ -

646


10/34

MORPHOLOGICALAND SEMANTIC REGULARITIES IN THE LEXICON 647The cost of 9 is thus the information that there is a word retribution,plus theinformation within the inner brackets, plus the cost of referringto the redundancyrule. Again, the assumption that the cross-referencecosts less than the additionalinformation /ion/ and ABSTRACTRESULT OF ACT OF gives the correct descriptionof our intuitions. This time we have avoided hypothetical lexical items, at theexpense of using rather artificial entries like 9.This artificiality betrays itself when we try to describe the relation between setslike aggression-aggressive-aggressor, aviation-aviator, and retribution-retributive.If there are hypothetical roots *aggress,*aviate,and *retribute,each of the membersof these sets can be related to its root by the appropriate redundancy rule 3,lOa, or 10b, where lOa and 10b respectively describe pairs like predict-predictiveand protect-protector (I omit the semantic portion of the rules at this point forconvenience-in any case, ?4 will justify separating the morphological andsemantic rules):

[x ] [w 1(10) a. l/y+ ive/ IUy[+A J L+VJ

b. ly + or/ - [/y/ ]+N _ +VSuppose we eliminate hypothetical lexical items in favor of entries like 9 forretribution.What will the entry for retributive ook like? One possibility is:

(11) [65 1derived by rule lOafrom

"/retribut/+V+NP1 for NP2_NP1 RETRIBUTE NP2 _

But this solution requires us to list the information in the inner brackets twice, inretributionand retributive:such an entry incorrectlydenies the relationshipbetweenthe two words.Alternatively, the entry for retributivemight be 12 (I use 3' here to denote theinverseof 3, i.e. a rule that derivesverbs from -ion nouns; presumably the presenceof 3 in the lexical component allows us to use its inverse as well):

(12) -65 -derived by 3' and lOafrom511Thus retributive s related to retributionby a sequence of redundancyrules, and theindependent information content of the pair retribution-retributives the informa-tion that there are two words, plus the information within the inner brackets of 9,plus the cost of referringto 3' once and lOa twice. This is closer to the intuitivelycorrect solution, in that it relates the two words. However, it is still suspicious,


11/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)because it claims retribution s more basic than retributive.Clearly the entries couldjust as easily have been set up with no difference in cost by making retributivebasic. The same situation will arise with a triplet like aggression-aggressor-aggressive,where the choice of one of the three as basic must be purely arbitrary.Intuitively, none of the three should be chosen as basic, and the formalization of thelexicon should reflect this. The impoverished-entry theory thus faces a choice:either it incorporates hypothetical lexical items, or it describes in an unnaturalfashion those related lexical items which are related through a non-lexical root.Consider now how the full-entry theory accounts for these sets of words,beginning with the case of a singleton like perdition (or conflagration),which hasno relatives like *perdite, *perditiveetc., but which obviously contains the -ionending of rule 3. We would like the independent information content of this itemto be less than that of a completely idiosyncratic word like orchestra-but morethan that of, say, damnation,which is based on the lexical verb damn. The impover-ished-entry theory resorts either to a hypothetical lexical item *perdite or to anentry containing another entry, like 9, which we have seen to be problematic.The full-entry theory, on the other hand, captures the generalization withoutextra devices. Note that 6b, the measure of non-redundant information in thelexical entry, is cleverly worded so as to depend on the existence of redundantinformation somewhere in the lexicon, but not necessarily on the existence ofrelated lexical entries. In the case of perdition, the only part of the entry whichrepresents a regularity in the lexicon is in fact the -ion ending, which appears aspart of the redundancyrule 3. What remains irregular is the residue described inthe right-hand side of 3, i.e. that part of perdition which corresponds to the non-lexical root *perdite.Hence the independent information content of perdition s theinformation that there is a word, plus the cost of the root, plus the cost of referringto rule 3. Perditionadds more information than damnation,then, because it has aroot which is not contained in the lexicon; it contains less information thanorchestra because the ending -ion and the corresponding part of the semanticcontent are predictable by 3 (presumably the cost of referringto 3 is less than theinformation contained in the ending itself; see ?6).

We see then that the full-entry theory captures our intuitions about perditionwithout using a hypothetical lexical item. The root *perditeplays only an indirectrole, in that its COST ppears in the evaluation of perdition as the differencebetweenthe full cost of perditionand that of the suffix; nowhere in the lexicon does the rootappear as an independent lexical entry.Now turn to the rootless pair retribution-retributive.Both words will have fullyspecified lexical entries. To determine the independent information content of thepair, suppose that retribution s added to the lexicon first. Its independent informa-tion, calculated as for perdition above, is the information that there is a word, plusthe cost of the root *retribute,plus the cost of referring to 3. Note again that*retributedoes not appear anywhere in the lexicon. Now we add to the lexiconthe entry for retributive,which is entirely predictable from retributionplus redun-dancy rules 3 and lOa. According to information measure 6, retributiveadds theinformation that it is a word, plus the cost of referring to the two redundancyrules.The cost of the pair for this order of introduction is therefore the information that

648


12/34

MORPHOLOGICALAND SEMANTIC REGULARITIES IN THE LEXICON 649there are two words, plus the information in the root *retribute,plus the cost ofreferringto redundancy rules three times. Alternatively, if retributive s added tothe lexicon first, followed by retribution, he independentinformation content of thepair comes out the same, though this time the cost of the root appears in theevaluation of retributive.Since the costs of these two orders are commensurate,there is no optimal order of introduction, and thus no reason to consider eitheritem basic.

Similarly, the triplet aggression-aggressor-aggressivewill have, on any order ofintroduction, an independent information content consisting of the informationthat there are three words, plus the information content of the root *aggress, plusthe cost of referring to redundancy rules five times (once for the first entry in-troduced, and twice for each of the others). Since no single order yields a signi-ficantly lower information content, none of the three can be considered basic to theothers.Thus the full-entry theory provides a description of rootless pairs and tripletswhich avoids either a root in the lexicon or a claim that one memberof the group isbasic, the two alternativesencountered by the impoverished entry theory. The full-entry theory looks still more appealing when contrasted with the transformationaltheory's account of these items. The theory of Lakoff 1971b introduces a positiveabsolute exception on *perdite, requiring it to nominalize; but *aggress mayundergo either -ion nominalization, -or nominalization, or -ive adjectivaliza-tion, and it must undergo one of the three. Lakoff is forced to introduce Booleancombinations of exception features, together marked as an absolute exception, inorder to describe this distribution-patently a brute force analysis.In the full-entry theory, then, the lexicon is simply a repository of all informationabout all the existing words; the information measure expresses all the relation-ships. Since the full-entry theory escapes the pitfalls of the impoverished-entrytheory, without giving up adequacy of description, we have strong reason toprefer the former, with its non-standard evaluation measure. From here on, theterm 'lexicalist theory' will be used to refer only to the full-entry theory.Before concluding this section, let us consider a question which frequentlyarises in connection with rootless pairs and triplets: What is the effecton the lexiconif a back-formation takes place, so that a formerly non-existent root (say *retribute)enters the language? In the transformational theory, the rule feature on the hypo-thetical root is simply erased, and the lexicon becomes simpler, i.e. more regular.In the lexicalist theory, the account is a bit more complex, but also more sophis-ticated. If retribute were simply added without disturbing the previous order formeasuring information content, it would add to the cost of the lexicon the informa-tion that there is a new word plus the cost of referringto one of the redundancyrules. Thus the total cost of retribution-retributive-retribute ould be the informa-tion that there are three words, plus the information in the root retribute, plus thecost of four uses of redundancy rules. But now that retribute is in the lexicon, arestructuringis possible, in which retributeis taken as basic. Under this order ofevaluation, the information content of the three is the information that there arethree words, plus the information in retribute, plus only two uses of redundancyrules. This restructuring, then, makes the language simpler than it was before


13/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)retribute was introduced, except that there is now one more word to learn thanbefore. What this account capturesis that a back-formation ceases to be recognizedas such by speakers precisely when they restructure the evaluation of the lexicon,taking the back-formation rather than the morphological derivatives as basic.I speculate that the verb *aggress, which seems to have only marginal status inEnglish, is still evaluated as a back-formation, i.e. as a derivative of aggression-aggressor-aggressive, and not as their underlying root. Thus the lexicalist theoryof nominalizations provides a description of the diachronic process of back-formation which does more than simply erase a rule feature on a hypotheticallexical item: it can describe the crucial step of restructuringas well.

4. SEPARATEMORPHOLOGICALND SEMANTIC ULES. At the outset of the dis-cussion, I stated redundancy rule 3 so as to relate lexical items both at the mor-phological and semantic levels. In fact, this formulation will not do. It claims thatthere is a particular meaning, ABSTRACTESULTOFACTOFV-ING,associated with theending -ion. However, several different semantic relations obtain between -ionnominals and their related verbs, and several nominalizing endings can express thesame range of meanings. Some of the morphological rules are stated in M1 (themorphological part of 3), M2, and M3:

(13) ..M: [/+ ion/] [ ]M2: [Y+ ment] ]M3: [Y+al] [+VI]

Some of the semantic rules areS1 (the semantic part of 3), S2, and S3:[+N+ [NPi's - ((P)NP2)] "+ V

(14) S1: ABSTRACT RESULT OF + [NP1 - ((P)NP2)]ACT OF NP1'S Z-ING NP1 Z NP2NP2

[+N 1 [+VS2: + [ (NP2)] + [NP1 (NP2)]GROUPTHATZ-S (NP2) NP1 Z (NP2)

+N+[(NP1's) ((P)NP2)] [+VS3:

(ACT^*- +[NP1--((P)NP)]I

.^P^ pROCSOF0 NP1ZNP2NP1'S) {PROCESS) JF LN ZN2l Z-ING NP2An example of the cross-classification of the morphological and semanticrelations is the following table of nouns, where each row contains nouns of the

650


14/34

MORPHOLOGICAL AND SEMANTIC REGULARITIES IN THE LEXICON 651

same semantic category, and each column contains nouns of the same morpho-logical category.

(15) Ml M2 M3S1: discussion argument rebuttalS2: congregation governmentS3: copulation establishment refusalThat is, John's discussion of the claim, John's argument against the claim, andJohn's rebuttal of the claim are semantically related to John discussed the claim,John argued against the claim, and John rebuttedthe claim respectively in a wayexpressed by the subcategorization conditions and semantic interpretationsof S1;the congregationand thegovernmentofFredonia are relatedby S2 to theycongregateand they govern Fredonia;and John'scopulationwith Mary,John's establishmentofa new order,and John's refusal of the ojfferare related by S3 to JohncopulatedwithMary, John establisheda new order,and John refusedthe offer.5There are furthernominalizing endings such as -ition (supposition),-ing (writing)and -0 (offer); andfurther semantic relations, such as ONE WHO Z's (writer, occupant) and THING INWHICH ONE Z's (residence, entrance).The picture that emerges is of a family ofnominalizing affixesand an associated family of noun-verb semantic relationships.To a certainextent, the particularmembersof each family that are actually utilizedin forming nominalizations from a verb are chosen randomly. Insofar as the choiceis random, the information measure must measure independently the cost ofreferring to morphological and semantic redundancy rules (cf. ?6 for furtherdiscussion).How do we formalize the information measure, in light of the separation ofmorphological rules (M-rules) and semantic rules (S-rules)? An obvious first pointis that a semantic relation between two words without a morphological relationshipcannot be counted as redundancy:thus the existence of the verb rule should renderthe semantic content of the noun ruler redundant, but the semantic content ofking must count as independent information. Hence we must requirea morpho-logical relationship before semantic redundancycan be considered.A more delicate question is whether a morphological relationship alone shouldbe counted as redundant.For example, professorand commission as in 'a salesman'scommission') are morphologically related to profess and commit;but the existenceof a semantic connection is far from obvious, and I doubt that many Englishspeakers other than philologists ever make the association. What should theinformation content of these items include? A permissive approach would allowthe phonology of the root to be counted as redundant information; the onlynon-redundant part of professor, then, would be some semantic information likeTEACH. A restrictive approach would require a semantic connection before mor-phology could be counted as redundant; professor then would be treated likeperdition, as a derived word with a non-lexical root, and both the phonology/profess/ and the semantics TEACHwould count as independent information.

5 I assume,with Chomsky1970, that the of innominalizations s transformationallynserted.Note also that the semantic relations are only approximate-the usual idiosyncrasiesappear inthese examples.


15/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)In ?5.1 I will present two cases in which only morphological rules play a rolebecause there are no semantic regularities. However, where semantic rules exist,it has not yet been established whether their use should be requiredin conjunctionwith morphological rules. Therefore the following restatement of informationmeasure 6 has alternative versions:

(16) (Information measure)Given a fully specified lexical entry W to be introduced into the lexicon,the independent information it adds to the lexicon is(a) the information that W exists in the lexicon; plus(b) (permissive form) all the information in W which cannot bepredicted by the existence of an M-rule which permits W to be

partially described in terms of information already in thelexicon, including other lexical items and S-rules; or(b') (restrictive form) all the information in W which cannot bepredictedby the existence of an M-rule and an associated S-rule(if thereis one) which together permitW to be partiallydescribedin terms of information already in the lexicon; plus(c) the cost of referringto the redundancyrules.

Examples below will show that the permissive form of 16 is preferable.5. OTHERAPPLICATIONS.he redundancy rules developed so far describe therelationbetween verbs and nominalizations. It is clear that similar rules can describe

de-adjectival nouns (e.g. redness, entirety, width), deverbal adjectives (predictive,explanatory), denominal adjectives (boyish, national, fearless), de-adjectival verbs(thicken, neutralize, yellow), and denominal verbs (befriend, originate, smoke).Likewise, it is easy to formulate rules for nouns with noun roots like boyhood,6adjectiveswith adjective roots like unlikely, and verbs with verb roots like re-enterand outlast.7 Note, by the way, that phonological and syntactic conditions such aschoice of boundary and existence of internal constituent structure can be spelledout in the redundancyrules.Note also that a complex form such as transformationalistcan be accountedfor with no difficulty: each of the steps in its derivation is described by a regularredundancyrule. No question of ordering rules or of stating rule features need everarise (imagine, by contrast, the complexity of Boolean conditions on exceptionswhich would be needed in the entry for transform, f we were to generate this wordin a transformational theory of the lexicon). Transformationalist s fully specifiedin the lexicon, as are transform, transformation,and transformational.The totalinformation content of the four words is the information that there are four words,plus the information in the word transform,plus idiosyncratic information addedby successive steps in derivation (e.g. that transformation n this sense refers to acomponent of a theory of syntax and not just any change of state, and that atransformationalistn the sense used in this paper is one who believes in a particular

6 The abstractnessof nouns in -hood,mentionedby Halle 1973as an example, is guaranteedby the fact that the associated semantic rule yields the meaning STATE(or PERIOD) OF BEINGA Z.7 Cf. Fraser 1965 for an interesting discussion of this last class of verbs.

652


16/34

MORPHOLOGICALAND SEMANTIC REGULARITIES IN THE LEXICON 653form of transformational theory), plus the cost of referringto the three necessaryredundancy rules. Note that the information measure allows morphologicallyderived lexical items to contain more semantic information thanthe rulepredicts;weuse this fact crucially in describingthe information content of transformationalist.More striking examples will be given below.With these preliminary observations, I will now present some more diverseapplications of the redundancyrules.

5.1. PREFIX-STEMERBS.Many verbs in English can be analysed into one of theprefixes in-, de-, sub-, ad-, con-, per-, trans- etc. followed by one of the stems -sist,-mit, -fer, -cede, -cur etc. Chomsky & Halle argue, for phonological reasons, thatthe prefix and stem are joined by a special boundary =. Whether a particularprefix and a particular stem together form an actual word of English seems to bean idiosyncratic fact:

(17) *transist transmit transfer *transcede *transcurpersist permit prefer precede *precurconsist commit confer concede concurassist admit *affer accede *accursubsist submit suffer succeed *succurdesist *demit defer *decede *decurinsist *immit infer *incede incur

We would like the information measure of the lexicon to take into account theconstruction of these words in computing their information content. There are twopossible solutions. In the first, the lexicon will contain, in addition to the fullyspecified lexical entries for each actually occurring prefix-stem verb, a list of theprefixes and stems from which the verbs are formed. The redundancy rules willcontain the following morphological rule, which relates three terms:

E+ temJJThe information content of a particularprefix-stem verb will thus be the informa-tion that there is a word, plus the semantic content of the verb (since there is nosemantic rule to go with 18, at least in most cases), plus the cost of referring tomorphological rule 18. The cost of each individual prefix and stem will be countedonly once for the entire lexicon.Since we have up to this point been unyielding on the subject of hypotheticallexical items, we might feel somewhat uncomfortable about introducing prefixesand stems into the lexicon. However, this case is somewhat different from earlierones. In the case of perdition, the presumed root is the verb *perdite. If perditewere entered in the lexicon, we would have every reason to believe that lexicalinsertion transformations would insert perdite into deep structures, and that thesyntax would then produce well-formed sentences containing the verb perdite.In order to prevent this, we would have to put a feature in the entry for perdite toblock the lexical insertion transformations. It is this rule feature [-Lexical


17/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)Insertion] which we wish to exclude from the theory. Consider now the lexicalentry for the prefix trans-:

(19) [/trans/ ]+lyPrefix]Trans-has no (or little) semantic information, and as syntactic information has onlythe marker [+Prefix]. Since the syntactic category Prefix is not generated by thebase rules of English, there is no way for trans- alone to be inserted into a deepstructure.It can be insertedonly when combined with a stem to form a verb, sincethe category Verb does appear in the base rules. Hence there is no need to use theoffending rule feature [-Lexical Insertion] in the entry for trans-, and no need tocompromise our earlier position on *perdite.

However, there is another possible solution which eliminates even entries like 19,by introducingthe prefixesand stems in the redundancyrule itself. In this case, theredundancy rule consists of a single term, and may be thought of as the simplesttype of'word-formation' rule:

trans sistper mitcon fer

(20) / aD = cede /( suB tain /de

The information content of prefix-stemverbs is the same as before, but the cost ofthe individual prefixesand stems is counted as part of the redundancy rule, not ofthe list of lexical items. The two solutions appear at this level of investigation to beequivalent, and I know as yet of no empirical evidence to decide which should bepermitted by the theory or favored by the evaluation measure.This is the case of morphological redundancy without semantic redundancypromised in ?4. Since, for the most part, prefixesand stems do not carry semanticinformation, it is not possible to pair 18 or 20 with a semantic rule.8 Obviouslythe information measure must permit the morphological redundancy anyway.Besides complete redundancy,we now have three cases to consider: those in whicha semantic redundancy rule relates a word to a non-lexical root (e.g. perdition),those in which the semantic rule relates a word incorrectly to a lexical root (e.g.professor), and those in which there is no semantic rule at all. The three cases areindependent, and a decision on one of them need not affect the others. Thus thedecision to allow morphological redundancyfor prefix-stemverbs still leaves openthe question raised in ?4 of how to treat professor.It should be pointed out that word-formation rules like 20 are very similar to

8 If they did carry semantic information, it would be more difficult,but not necessarilyimpossible,to state the rulein the form of 20. This is a potentialdifferencebetweenthe solutions,concerningwhich I have no evidenceat present.

654


18/34

MORPHOLOGICAL AND SEMANTIC REGULARITIES IN THE LEXICON 655Halle's word-formation rules (1973). The major differencehere between his theoryand mine is that his lexicon includes, in addition to the dictionary, a list of allMORPHEMES n the language, productive and unproductive. The present theorylists only WORDSn the lexicon. Productive affixesare introduced as part of lexicalredundancy rules, and non-productive non-lexical morphemes (such as *perdite)do not appear independently anywhere in the lexical component. Other than theargumentsalready stated concerning the feature [Lexical Insertion], I know of littleevidence to distinguish the two solutions. However, since Halle has not formulatedthe filter, which plays a crucial role in the evaluation measure for his theory of thelexicon, it is hard to compare the theories on the level where the present theorymakes its most interesting claims.

5.2. NOUN COMPOUNDS.he compound nouns in 21 are all formed by conca-tenating two nouns:(21) a. garbage man, iceman, milkman, breadbasket, oil drumb. snowman, gingerbreadman, bread crumb, sand castlec. bulldog, kettledrum, sandstone, tissue paper

Although the meaning of each compound is formed from the meanings of the twoconstituent nouns, the way in which the meaning is formed differsfrom line to line.Part of a speaker's knowledge of the English lexicon is the way in which themeanings of compounds are related to the meanings of their constituents: thus wewould say that someone did not know English if he (seriously) used garbageman tomean 'a man made out of garbage', by analogy with snowman.If one brought Lees 1960 up to date, one would get an approach to compoundswhich uses transformations to combine nouns randomly, controlled by exceptionfeatures so as to produce only the existing compounds with the correct meanings.But how can such exception features be formulated? Either noun in a compoundcan be changed, with a corresponding change in acceptability: we have garbageman, garbage truck, but not *garbage gingerbread, *garbage tree; we also havegarbage man, gingerbreadman, but not *ant man, *tissue man. Thus the use ofexception features will require each noun in the lexicon to be cross-listed withevery other noun for the compounding transformations. Furthermore, sincegingerbread s itself a compound, ginger,bread,and manwill all somehow have to berelated by the exception features. In the end, the exception features appear to beequivalent to a listing of all the existing compounds along with their meanings.In the lexicalist theory, we can dispense with exception features in the descriptionof compounds. We simply give each actually occurringcompound a fully specifiedlexical entry, and in the list of redundancy rules we enter morphological rule22 and semantic rules 23a,b,c, describing the data of 21a,b,c respectively. Ofcourse, there are a great number of additional semantic rules (cf. Lees, chapter 4);I list only these three as a sample:

(22) [I[Nx] [NY]/ ] [|+N]


19/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)

m-i\, +N JLZJ(3) a. [Z THAT CARRIES W] (\+N] }

1,~~~~I W\\f+N1lbZ MADEOF WI' +NN}

C.LZ LIKE A W\ \[+N] [z+. A W]+

N{ [\}

The redundancyrules thus define the set of possible compounds of English, and thelexicon lists the actually occurring compounds.The information measure 16gives an intuitivelycorrect result for the independentinformation in compounds. For example, since the nouns garbage and man arein the lexicon, all their information will be counted as redundant in evaluating theentry for garbage man. Thus the independent information content of garbage manwill be the information that such a word exists, plus any idiosyncratic facts aboutthe meaning (e.g. that he picks up rather than delivers garbage), plus the cost ofreferringto 22 and 23a. The information of a complex compound like gingerbreadman is measured in exactly the same way; but the independent information in itsconstituent gingerbread is reduced because of its relation to ginger and bread.Gingerbreadman is thus parallel in its evaluation to the case of transformationalistcited earlier.

Now consider the problem of evaluating the following nouns:(24) a. blueberry, blackberryb. cranberry, huckleberryc. gooseberry, strawberry

Blueberry and blackberry are obviously formed along the lines of morphologicalrule 25 and semantic rule 26. This combination of rules also forms flatiron, high-chair, madman, drydock, and many others.

(25) [I[Ax] [NY]/] JI+A]}

(26) [Z WHICHIS WJ {[+A ]}

Thus blueberryand blackberryare evaluated in exactly the same way as garbageman.Cranberryand huckleberrycontain one lexical morpheme and one non-lexicalmorpheme. The second part (-berry)and its associated semantics should be redun-

656


20/34


dant, but the phonological segments /craen/and /hukl/, and the semantic charac-teristics distinguishing cranberries and huckleberriesfrom other kinds of berries,must be non-redundant. Hence this case is just like perdition,where a non-lexicalroot is involved, and the information measure formulated for the case of per-dition will yield the intuitively correct result. One problem is that the lexicalcategories of cran- and huckle- are indeterminate, so it is unclear which mor-phological ruleapplies. Likewise, it is unclearwhich semantic rule applies. However,I see nothing against arbitrarily applying the rules which cost least; this conven-tion will minimize the information in the lexicon withoutjeopardizing the generalityof the evaluation procedure.We observe next that gooseberryand strawberrycontain two lexical morphemesand are both berries, but gooseberries have nothing to do with geese and straw-berries have nothing to do with straw. This case is thus like professor, which hasnothing to do semantically with the verb profess, and exactly the same questionarises in their evaluation: should they be intermediate in cost between the previoustwo cases, or should they be evaluated like cranberry,with straw- and goose-counted as non-redundant? The fact that there is pressure towards phonologicalsimilarity even without semantic basis (e.g. gooseberry was once groseberry) issome evidence in favor of the permissive form of 16, in which morphologicalsimilarity alone is sufficient for redundancy.Another semantic class of compound nouns (exocentric compounds) differsfrom those mentioned so far in that neither constituent describes what kind ofobject the compound is. For example, there is no way for a non-speaker of Englishto know that a redhead is a kind of person, but that a blackhead is a kind ofpimple.9 Other examples are redwing (a bird), yellow jacket (a bee), redcoat (asoldier), greenback (a bill), bigmouth (a person), and big top (a tent). The mor-phological rule involved is 25; the semantic rule must be

[?N 1 [[?N](27) THING WITH A Z j


21/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)these elaborate rule features stems from the nature of transformations. Anyphrase-markertaken as input to a particular transformation corresponds to a setof fully specified output phrase-markers.In the case of exocentric compounds, thecombination of the two constituent words by rule 27 does not fully specify theoutput, since the nature of THINGn 27 is inherently indeterminate.We thus see an important empirical difference between lexical redundancyrules and transformations: it is quite natural and typical for lexical redundancyrules to relate items only partially, whereas transformationscannot express partialrelations. Several illustrations of this point have appeared already, in the mor-phological treatment of perditionand cranberryand in the semantic treatment oftransformationalist.However, the case of exocentric compounds is perhaps themost striking example, since no combination of exception featuresand hypotheticallexical items can make the transformationaltreatmentappearnatural.The lexicalisttreatment, since it allows rules to relate exactly as much as necessary, handlesexocentric compounds without any remarkableextensions of the machinery.

5.3. CAUSATIVE ERBS.There is a large class of verbs which have both transitiveand intransitiveforms;10 e.g.,(28) a. The door opened.b. Bill opened the door.(29) a. The window broke.b. John broke the window.(30) a. The coach changed into a pumpkin.b. Mombi the witch changed the coach from a handsome young maninto a pumpkin.

It has long been a concern of transformational grammarians to express the factthat the semantic relations of door to open, of windowto break, and of coach tochange are the same in the transitive and intransitive cases.There have been two widely accepted approaches, both transformational incharacter. The first, that of Lakoff 1971b, claims that the underlying form of thetransitive sentence contains the intransitive sentence as a complement to a verb ofcausation-i.e., that the underlying form of 28b is revealed more accurately in thesentence Bill causedthe door to open. The other approach, case grammar, is that ofFillmore 1968. It claims that the semantic relation of door to open is expressedsyntactically in the deep structuresof 28a and 28b, and that the choice of subjectis a purelysurfacefact. The deep structuresaretaken to be 31laand 31b respectively:

(31) a. past open [objective the door]b. past open [objective the door] [Agentive by Bill]These proposals and their consequences have been criticized on diverse syntacticand semantic grounds (cf., e.g., Chomsky 1972, Fodor 1970, and Jackendoff

1972, Chap. 2); I do not intend to repeat those criticisms here. It is of interest tonote, however, that Lakoff's analysis of causatives is the opening wedge into thegenerative semanticists' theory of lexicalization: if the causative verb break is theresult of a transformation, we would miss a generalization about the nature of10 This class may also include the two forms of begin proposed by Perlmutter 1970.

658


22/34


agentiveverbsby failingto derivethe causative verb kill by the same transformation.But since kill has as intransitiveparallel not kill but die, and since there are manysuch causative verbs without morphologically related intransitives, the only wayto avoid an embarrassingnumber of exceptions in the lexicon is to perform lexicalinsertion AFTERhe causative transformation, as proposed by McCawley 1968.Again, the difficulty in this solution lies in the nature of transformations. Thereare two cross-classifying generalizations which a satisfactory theory must express:all causative verbs must share a semantic element in their representation; and theclass of verbs which have both a transitive causative form and an intransitivenon-causative form must be described in a general fashion. Expressing the secondgeneralization with a transformation implies a complete regularity, which in turnloses the first generalization; McCawley's solution is to make a radical move torecapturethe first generalization.There remains the alternative of expressing the second generalization in a waythat does not disturb the first. Fillmore's solution is along these lines; but he stillrequiresa radical change in the syntactic component, viz. the introduction of casemarkers.The lexicalist theory can leave the syntactic component unchanged by using thepower of the lexicon to express the partial regularityof the second generalization.The two forms of break are assigned separate lexical entries:

7/brak/(32) a. NP ]

,NPi BREAK7/brtk/

+V+[NP2 ' NP1]

_NP2 CAUSE (NP, BREAK)_The two forms are related by the following morphological and semantic rules:"

(33) a[KV] [+Vb. +[NP1 ]| - +[NP2 NP1]

[NP, W NP2 CAUSE NP, W)JThus the independent information contained in the two entries for breakis the factthat there are two words,12plus the independent information in the intransitiveform 32a, plus the cost of referring to the redundancy rules. Hence the relationbetween the (a) and (b) sentences in 28-30 is expressed in the lexicon and not inthe transformational component.

11Since 32a is an identity rule,it is possibly dispensable.I have included t herefor thesakeofexplicitness, and also in orderto leave the form of the information measureunchanged.12 Perhaps the use of the identityrule32acould makethe two wordscount as one, if this weredesirable.I have no intuitions on the matter,so I will not botherwith the modification.


23/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)This solution permits us still to capture the semantic similarity of all causativeverbs in their lexical entries; thus die and kill will have entries 34a and 34b

respectively:-/d/ -

(34) a. ++a [NP,__]NP1 DIE

-/kil/ 1+V+ [NP2 NP1]NP2 CAUSE NP1 DIE)

Die and kill are related semantically in exactly the way as the two entries of break:one is a causative in which the event caused is the event described by the other.However, since there is no morphological rule relating 34a-b, the informationmeasure does not relate them; the independent information contained in the twoentries is the fact that there are two words, plus all the information in both entries.Thus the lexicalist theory successfully expresses the relation between the twobreaksand their relation to kill and die, without in any sense requiringkill and dieto be exceptional, and without making any radical changes in the nature of thesyntactic component.A further possibility suggested by this account of causative verbs is that thepartial regularitiesof the following examples from Fillmore are also expressedin thelexicon:(35) a. Bees swarmed in the garden.We sprayed paint on the wall.

b. The garden swarmed with bees.We sprayed the wall with paint.Fillmore seeks to express these relationships transformationally,but he encountersthe uncomfortable fact that the (a) and (b) sentences are not synonymous: the (b)sentences imply that the garden was full of bees and that the wall was covered withpaint, but the (a) sentences do not carry this implication. Anderson 1971 showsthat this semantic differenceargues against Fillmore's analysis, and in favor of onewith a deep-structuredifference between the (a) and (b) sentences. A lexical treat-ment of the relationship between the two forms of swarm and spraycould expressthe difference in meaning, and would be undisturbed by the fact that some verbs,such as put, have only the (a) form and meaning, while others, such as fill, haveonly the (b) form and meaning. This is precisely parallel to the break-break vs.die-kill case just discussed.

Consider also some of the examples mentioned in Chomsky 1970. The relation ofHe was amused at the stories and The stories amused him can be expressed in thelexicon, and no causative transformation of the form Chomsky proposes need beinvoked. The nominalization his amusementat the stories contrasts with *thestories'amusement of him because amusementhappens to be most directly related to theadjectival amused at rather than to the verb amuse. Other causatives do havenominalizations, e.g. the excitation of theprotons by gamma rays. I take it then that

660


24/34

MORPHOLOGICALAND SEMANTIC REGULARITIES IN THE LEXICON 661the existence of only one of the possible forms of amusementis an ad-hoc fact,expressed in the lexicon.

Chomsky also cites the fact that the transitive use of grow, as in John growstomatoes, does not form the nominalization *the growth of tomatoes by John.Rather the growth of tomatoes is related to the intransitive tomatoes grow. Againwe can express this fact by means of lexical relations. This time, the relation isperhapsmore systematic than with amusement,since nouns in -th, such as width andlength, are generally related to intransitivepredicates. Thus the meaning of growthcan be predictedby the syntacticpropertiesof the redundancyrule which introducesthe affix -th. The transitive grow does in fact have its own nominalization: thegrowing of tomatoesby John. Thus Chomsky's use of causatives as evidence for theLexicalist Hypothesis seems incorrect-in that causatives do have nominalizations,contraryto his claim. But we can account for the unsystematicityof the nominaliza-tions, as well as for what regularitiesdo exist, within the present framework.Note also that our account of causatives extends easily to Lakoff's class ofinchoative verbs (1971b). For example, the relation of the adjective open to theintransitive verb open ('become open') is easily expressed in a redundancy rulesimilar to that proposed for causatives.As further evidence for the lexicalist theory, consider two forms of the verbsmoke:

(36) a {The cigar } smoked.(The chimneys, , , , (the cigar.b. Johnsmoked1..*thehi. f(*the chimney.JThe intransitive verb smoke means 'give off smoke'; it is related to the noun smokeby a redundancy rule that applies also to the appropriate senses of steam, smell,piss, flower, and signal. The transitive form of smoke in the sense of 36b is partiallyrelated to the intransitive form by 33 in that it means 'cause to give off smoke',but it contains additional information-something like 'by holding in the mouthand puffing'. This information is not predictable from the redundancy rule, but itprovides the clue to the anomaly of *John smoked the chimney (so, if John werea giant, he might well use a chimney like a pipe, and then the sentence might beacceptable).A transformationaltheory has no way to capturethis partial generaliza-tion without artificiality. The lexicalist theory simply counts the unpredictableinformation as non-redundant, and the predictable information as redundant.While we are on the subject of smoke, it may be interesting to point out someother senses of smoke as illustration. Call the noun smoke1, and the intransitiveand transitive senses just discussed smoke2 and smoke3 respectively. There isanother transitive verb smoke4, which means 'permeate or cover with smoke' asin John smoked the ham. The redundancy rule relating smoke4 to smoke1 is alsoseen in verbs like paint, another sense of steam, water (in water the garden), andpowder (as in powder your nose), flour, and cover. There is another intransitivesmoke5, meaning 'smoke3 something'. The ambiguity in Johnis smokingis betweensmoke2 and smoke5. Smoke5 is related to smoke3 by a redundancy rule that alsohandles two forms of eat, drink, draw, read, cook, and sing. From smoke3 we alsoget the nominalization smoke6, 'something that is smoked3' (e.g. A cigar is a good


25/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)smoke) by the redundancyrule that also gives the nouns drink, desire, wish, dream,find, and experience. The verb milk (as in milk a cow) is related to the noun assmoke1 and smoke3 are related, but without an intermediate *The cow milked('The cow gave off milk'); the relation between the two milks requirestwo sets ofredundancyrules used together. We thus see the rich variety of partial regularitiesin lexical relations: their expression in a transformational theory becomes hard toconceive, but they can be expressed quite straightforwardly n the lexicalist frame-work.

5.4. IDIOMS.dioms are fixed syntactic constructions which are made up of wordsalready in the lexicon, but which carry meanings independent of the meanings oftheir constituents. Since the meanings are unpredictable, the grammar must rep-resent a speaker's knowledge of what constructions are idioms and what theymean. The logical place to list idioms is of course in the lexicon, though it is notobvious that the usual lexical machinery will suffice.Fraser 1970 discusses three points of interest in the formalization of idioms.First, they are constructed from known lexical items; the information measure,which measures how much the speaker must learn, should reflect this. Second,they are for the most part constructed in accordance with known syntactic rules(with a few exceptions such as by and large), and in accordance with the syntacticrestrictions of their constituents. Third, they are often resistant to normally applic-able transformations; e.g., The bucket was kicked by John has only the non-idio-matic reading. I have nothing to say about this third consideration, but the firsttwo can be expressed in the present framework without serious difficulty.Let us deal first with the question of the internal structure of idioms. Since wehave given internal structure to items like compensation and permit, there seemsto be nothing against listing idioms too, complete with their structure. The onlydifferencein the lexical entries is that the structure of idioms goes beyond the wordlevel. We can thus assign the lexical entries in 37 to kick the bucket, give hell to, andtake to task.13

(37) a. [NP1 [vp [vkik] [NP [Art63] [Nbukat]]]]b [NP1 [vp [vgiv] [NP [Nhel]] [pp [pto] NP2]LNP1YELL AT NP2 J

[NPi [v [vtak] NP2 [pp [pto] [NP [Ntaesk]]]]1cLNP.RITICIZE P2 JThe lexical insertion rule will operate in the usual way, inserting the lexical entriesonto deep phrase markers that conform to the syntactic structure of the lexicalentries. Since the structure of the entries goes beyond the word level, the idiommust be inserted onto a complex of deep-structure nodes, in contrast to ordinarywords which are inserted onto a single node.

13 The normal notation for strict subcategorizationrestrictions is difficult to apply in thiscase, so I have for convenienceadopted a notation in which the strict subcategorizationcon-ditions are combined with the phonological and syntactic representations, in an obviousfashion. No particular theoretical significance is intended by the change in notation. Thisproposal, which appearsto be much like that of Katz 1973, was arrived at independently.

662


26/34


As with ordinary lexical entries, the strictly subcategorized NP's must have aspecific grammatical relation with respect to the entry, and this is indicated in theentries of 37. In the case of take NP to task, the strictly subcategorized directobject is in fact surroundedby parts of the idiom; i.e., the idiom is discontinuous.But in the presenttheory, this appearsnot to be causefor despair,as our formalismsseem adequate to accommodate a discontinuous lexical item.This last observation enables us to solve a puzzle in syntax: which is the under-lying form in verb-particle constructions, look up the answer or look the answerup? The standard assumption (cf. Fraser 1965) is that the particle has to form adeep-structureconstituent with the verb in order to formulate a lexical entry; hencelook up the answer is underlying, and the particle movement transformation is arightward movement. But Emonds 1972 gives strong syntactic evidence that theparticlemovement rulemust be a leftwardmovement. He feels uncomfortableaboutthis result because it requires that look ... up be discontinuous in deep structure;he consoles himself by saying that the same problem exists for take ... to task,but does not provide any interesting solution. Having given a viable entry fortake ... to task, we can now equally well assign discontinuous entries to idiomaticverb-particle constructions, vindicating Emonds' syntactic solution.By claiming that the normal lexical-insertion process deals with the insertion ofidioms, we accomplish two ends. First, we need not complicate the grammar inorder to accommodate idioms. Second, we can explain why idioms have thesyntactic structureof ordinary sentences: if they did not, the lexical insertion rulescould not insert them onto deep phrase markers. Our account of idioms thus hasthe important virtue of explaining a restriction in terms of already existing con-ventions in the theory of grammar-good evidence for its correctness.Now that we have provided a way of listing idioms, how can we capture thespeaker's knowledge that idioms are made up of already existing words? To relatethe words in the lexicon to the constituents of idioms, we need morphologicalredundancy rules. The appropriate rules for kick the bucket must say that a verbfollowed by a noun phrase forms a verb phrase, and that an article followed by anoun forms a noun phrase. But these rules already exist as phrase-structurerulesfor VP and NP. Thus, in the evaluation of idioms, we must use the phrase-structurerules as morphological redundancy rules. If this is possible, the independent in-formation in kick the bucket will be the information that it is a lexical entry, plusthe semantic information DIE, plus the cost of referringto the phrase-structurerulesfor VP and NP.

Though mechanically this appears to be a reasonable solution, it raises thedisturbing question of why the base rules should play a role in the informationmeasure for the lexical component. Some discussion of this question will appearin ?7. At this point I will simply note that this solution does not have very drasticconsequences for grammatical theory. Since the base rules can be used as redun-dancy rules only if lexical entries go beyond the word level, no descriptive poweris added to the grammaroutside the description of idioms. Therefore the proposalis very limited in scope, despite its initially outrageous appearance.If the base rules are used as morphological redundancy rules for idioms, wemight correspondinglyexpect the semantic projection rules to be used as semantic


27/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)redundancyrules. But of course this cannot be the case, since then an idiom wouldhave exactly its literal meaning, and cease to be an idiom. So we must assume thatthe permissiveversion of the information measure is beingused: both morphologicaland semantic redundancy rules exist, but only the morphological rules apply inreducingthe independentinformation in the idiom. This is furtherevidence that thepermissive version of the information measure must be correct.Note, by the way, that a transformational theory of nominalization containsabsolutely no generalization of the approach that accounts for idioms. Thus thelexicalist hypothesis proves itself superior to the transformational hypothesis in away totally unrelated to the original arguments deciding between them.

6. THE COSTOF REFERRINGTO REDUNDANCYRULES.In evaluating the independentinformation of lexical entries, we have continually included the cost of referringtoredundancyrules. We have not so far specifiedhow to calculate this cost, or how torelate it quantitatively to other costs in the lexicon. In this section I will proposesome preliminaryanswers to those questions.In the discussion of the full-entry theory in ?2, 1said that the cost of referringto aredundancyrule in evaluating a lexical entry representsone's knowledge of whichregularitieshold in that particularlexical entry. In order to be more specific, let usreconsider the meaning of the information measure in the full-entry theory. Inmeasuring the independent information contained in a lexical entry, we are ineffect measuring how much new information one needs in order to learn thatlexical item. If the lexical item is totally unrelated to anything else in the lexicon,one must learn it from scratch.But if there is other lexical information which helpsone know in advance some of the propertiesof the new word, there is less to learn;this is captured in clause (b) of the information measure.In learning that a new lexical item can be formed on the basis of an old lexicalitem and a redundancy rule, however, something must be learned besides theidentity of the old lexical item: namely, which redundancy rule. to apply. Forexample, part of one's knowledge of the lexicon of English is the fact that thenominalizations of refuse and confuse are refusal and confusion, not *refusion and*confusal, although in principle the latter forms could exist. That is, in learning thewords refusaland confusion,one must learn the arbitraryfact that, of the choice ofpossible nominal affixes, refuse uses -al and confuse uses -ion. Clause (c) of theinformation measure, the cost of referring to the redundancy rule, is meant torepresent this knowledge. I am claiming therefore that the evaluation of refusalmust take into account the fact that it, and not *refusion, s the proper nominaliza-tion of refuse.For a clear case of the use of clause (c), let us turn to another example. Botha1968 discusses the process of nominal compounding in Afrikaans, which containsmany compounds which are morphologically simple concatenations of two nouns,as in English. But there are also many compounds in which the two nouns arejoined by a 'link phoneme' s or a. Botha demonstrates at great length that thereis no phonological, morphological, syntactic, or semantic regularity in the useof link phonemes; i.e., the link phoneme must be learned as an idiosyncrasy ofeach individual compound.

664


28/34

MORPHOLOGICAL AND SEMANTIC REGULARITIES IN THE LEXICON 665In the present theory, the Afrikaans lexicon contains three morphological rulesfor noun compounds:

(38) a. [[NX] [NY]/] {[I I}[+,, LSN ]lu { i

[I[NX][NY]/] {[L+ N]}

1+N]b-L+N l Wyl

[I[NX] [NY] ] {[IxI]Since all the morphological information of a particularcompound is predicted byone of the three rules in 38, clause (b) of the information measure contributesnothing to the information content of the compound. But since the speakermust learn which of the three is appropriate,clause (c) must contribute the cost ofthe information involved in making this choice.A third example involves inflectional morphology. Halle 1973 argues thatparadigmatic information should be representedin the dictionary, and in fact thatonly and all fully inflected forms should be entered. As a consequence, the lexicalinsertion rules must enter partial or complete paradigms into deep structures,andthe rules of concord must have the function of filtering out all but the correctforms, rather than that of inserting inflectional affixes.14Under Halle's proposal,part of the task of the lexical component of English is to list the correspondencesbetween the present and past tense forms of verbs. Accordingly, we can state afew morphological redundancy rulesrelatingpresent to past tense forms in English:

(39)a ]^ r/xd ](a) [+ [V+pres]] [+[V+past]Jf/CoVCo/ ir/CoVCo+t/b L+[V+pres]J [+[V+past]]

[/CO back -CO -aback CO/c.round -caround+[VY+pres] _ +[V+past]

df/CoVCo/ 1 f/COoOx+d/1d. [V+pres]J [[V+past]]14 This of course requiresrules of concord to be of a differentformal nature than ordinary

transformations.But perhapsthis is not such a bad result,consideringthat the most convincingcases for Lakoff's global rules seem to be in this area. An independentargumentthat concordrules differ formally from transformationscould serve as evidence that transformationsneednot be global: only the very limited classof concordrules,which are no longertransformationsat all, need information from various levels of derivation.This more highlystructuredtheoryreduces the class of possible grammars.


29/34

LANGUAGE, VOLUME 51, NUMBER 3 (1975)Here 39a is the regular rule for forming past tenses, and the other three representvarious irregularforms: 39b relates keep-kept, dream-dreamt,lose-lost, feel-feltetc.; 39c relatestell-told, cling-clung,hold-held,break-brokeetc.; the very marginaland strange 39d relatesjust the six pairs buy-bought, bring-brought,catch-caught,fight-fought, seek-sought, and think-thought. Note that 39b-c take over thefunction of the 'precyclic re-adjustment rules' described by Chomsky & Halle(209-10).15A final preliminary point in this example: in the evaluation of a paradigm by theinformation measure, I assume that the information that a word exists is countedonly once for the entire paradigm. Although one does have to learn whethera verbhas a nominalization, one knows for certain that it has a past tense, participles, anda conjugation. Therefore the information measure should not count knowledgethat inflections exist as anything to be learned.Now let us return to the problem of measuringthe cost of referringto a redun-dancy rule. Intuitively, the overwhelminglyproductiverule 39a should cost virtuallynothing to refer to; the overwhelmingly marginal rules 39b-d should cost a greatdeal to referto, but less than the information they renderpredictable.The disparityin cost reflects the fact that, in choosing a past tense form, 39a is ordinary andunremarkable,so one must learn very little to use it; but the others are unusual or'marked' choices, and must be learned. We might furtherguess that 39b-c, whicheach account for a fair number of verbs, cost less to refer to than 39d, which appliesto only six forms (but which is nevertheless perceived as a minor regularity). Still,the pair buy-bought contains less independent information than the totally irregularpair go-went, which must be counted as two independent entries.These considerations lead to a formulation of the cost of reference somethinglike

(40) The cost of referring to redundancy rule R in evaluating a lexical entryW is IR,W x PR,W, where IR,W is the amount of information in Wpredicted by R, and PR.W s a number between 0 and 1 measuring theregularity of R in applying to the derivation of W.

For an altogether regular rule application, such as the use of 39a with polysyllabicverbs, PR,W will be zero. With monosyllabic verbs and 39a, PR,W will be almost butnot quite zero; the existence of alternatives means that something must be learned.For 39b-d, PR.Wwill be close to 1; their being irregularmeans that their use doesnot reduce the independent information content of entries nearly as much as 39a.In particular, 39d will reduce the independent information content hardly at all.In fact, it is quite possible that the total information saved by 39d in the evaluation

15 I have not considered the question of how to extend the phonological generalizationof39c to other alternationssuch as mouse-mice, ong-length.Perhapsthe only way to do this is toretainthe rulein the phonology, and simplylet the lexicalredundancyrulesupplya rule feature.But a more sophisticatedaccount of the interaction of the morphologicalrules might capturethis generalizationwithout a rulefeature; e.g., one could considerfactoringmorphologicalrulesinto phonological and syntactic parts,as we factored out separate morphologicaland semanticrules in ?4. In any event, I am includingall the phonology in 39 becausemany people have beendissatisfiedwith the notion of re-adjustment ules: I hope that bringing up an alternativemaystimulate someone to clarifythe notion.

666f


30/34

MORPHOLOGICAL AND SEMANTIC REGULARITIES IN THE LEXICON 667of the six relevant pairs of lexical entries is less than the cost of stating the rule.Our evaluation measure thus reflects the extremely marginal status of this rule. Inother cases, perhapsthe nominalizing affixes and Afrikaans compounds, the variouspossible derived forms are in more equal competition, and PR.Wwill have a valueof, say, 0.3.I will not suggest a precise method of calculating PR,W, s I believe it would bepremature. However, the general concept of how it should be formulated is fairlyclear. Count a lexical pair related by R as an ACTUALuse of R. Count a lexicalentrywhich meets one term of the structuraldescription of R, but in whose evalua-tion R plays no role, as a NON-USE f R. For example, confusecounts as a non-useof the rule introducingthe -al nominal affix, since it meets the structuraldescriptionof the verbal term of the rule, but there is no noun confusal.The sum of the actualuses and the non-uses is the number of POTENTIALses of R. PRW should be nearzero when the number of actual uses of R is close to the number of potential uses;PRw should be near 1 when the number of actual uses is much smaller than thenumber of potential uses; and it should rise monotonically from the formerextreme to the latter.If phonological conditions can be placed on the applicability of a redundancyrule, PR,Wdecreases

jackendoff, r. 1975. regularities in the lexicon

Documents