Coordination and How to Distinguish Categories Author(s): Ivan A. Sag, Gerald Gazdar, Thomas Wasow, Steven Weisler Source: Natural Language & Linguistic Theory, Vol. 3, No. 2 (May, 1985), pp. 117-171 Published by: Springer Stable URL: http://www.jstor.org/stable/4047644 . Accessed: 10/05/2011 15:35 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=springer. . 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 support@jstor.org. Springer is collaborating with JSTOR to digitize, preserve and extend access to Natural Language & Linguistic Theory. http://www.jstor.org IVAN A. SAG, GERALD GAZDAR, AND STEVEN COORDINATION THOMAS WASOW, WEISLER AND HOW TO DISTINGUISH C ATEGORIES* 1. INTRODUCTION Most generative studies of coordinationhave assumedsomethinglike the following generalization(from Chomsky 1957, p. 36): (1) If SI and S2 are grammaticalsentences, and SI differs from S2 only in that X appears in Si where Y appears in S2 (i.e., and S2= .. Y -), and X and Y are conSI= .. *X. stituentsof the same type in SI and S2, respectively, then S3 iS a sentence, where S3 is the resultof replacing X by X + and + Y in Si (i.e., S3 = .*. X+ and+ Y. .. Yet many have observed there are primafacie examples of coordination not covered by this generalization,namely,exampleslike those in (2) and (3) where the coordinate elements appear to be of distinct syntactic categories. (2)a. b. c. d. e. Pat is either stupid or a liar. Pat is a Republicanand proud of it. Pat is healthy and of sound mind. Pat is either asleep or at the office. That was a rude remarkand in very bad taste. [AP or NP] [NP and AP] [AP and PP] [AP or PP] [NP and PP] * The authors wish to acknowledge their debt to a great many individualsfor helpful conversations,suggestions,and/or encouragementthroughoutthe usuallylong time taken for this paper to see the light of day. Special thanksare due to Carl Pollardand Geoffrey Pullum,who providedso muchassistanceand advice that they probablyought to have been listed as co-authors.Our thinkingabout what the Head FeatureConvention needed to do was considerablyclarified by conversationswith Fernando Pereira and Stuart Shieber. Among the other people from whose comments we have benefitted are: Emmon Bach, Robin Cooper, Elisabet Engdahl, Aryeh Faltz, Donka Farkas, Dan Flickinger,J. Mark Gawron,GeorgiaGreen, FrankHeny, MartinKay, Ed Keenan,Ewan Klein, Bill Ladusaw, Joan Maling, Dick Oehrle, AlmerindoOjeda, E. Anne Paulson,Jessie Pinkham,Graham Russell,Paul Schachter,Peter Sells, Hans Uszkoreit,EdwinWilliams,and three anonymous referees. In addition,we thank Michael Wescoat and Dan Flickingerfor valuable help in manuscriptpreparation.Supportfor workon this paperwas providedby grantsto Stanford Universityfrom the National Science Foundation(BNS-8102406) and the Sloan Foundation, by the Center for the Studyof Languageand Information,and by grantsfrom the Sloan Foundationand SystemDevelopmentFoundationto the Center for Advanced Study in the BehavioralSciences (Gazdar). Natural Language and Linguistic Theory 3 (1985) 117-171. 0167-806X/85.10 1985 by D. Reidel Publishing Company ? 118 IVAN f. g. h. i. j. k. (3)a. b. A. SAG ET AL. Sandy is either a lunatic or under the influence of drugs. I am hoping to get an invitationand optimistic about my chances. I am neither an authorityon this subject nor trying to portraymyself as one. I am both expecting to get the job and of the opinion that it is a desirableone. Pat was awardedthe Golden Fleece Award and very upset about it. Pat was neither recommendedfor promotionnor under any illusions about what that meant. Pat has become a banker and very conservative. I consider that a rude remarkand in very bad taste. [NP or PP] [VP and AP] [NP nor VP] [VP and PP] [VP and AP] [VP nor PP] [NP and AP] [NP and PP] In light of these examples, it might be suggested that (1) should be weakened along the lines sketched in (4): (4) If we have two sentences Z + X + W and Z + Y+ W, and if X and Y are actually constituentsof these sentences, we can generally form a new sentence Z - X + and + Y - W (Chom- sky 1957, p. 35). Such a revision, however, is too permissive, as it fails to predict the deviance of (5c), (6c) and many similarcases. (5)a. b. c. The scene of the movie was in Chicago. The scene that I wrote was in Chicago. *The scene of the movie and that I wrote was in Chicago. (Chomsky 1957, p. 36) (6)a. b. c. John sang beautifully. John sang a carol. *John sang beautifullyand a carol. (Peterson 1981, p. 449) What then is the appropriatecondition on conjunct identity? In this paper, we provide a simple answerto this fundamentalquestion. Taking grammatical categories to be sets of feature-value pairs, our central claim is the following: COORDINATION (7) AND CATEGORIES 119 If a phrase structurerule introduced a category a, then any conjunct of a is a supersetof a. Our thesis then is that the sentences of (2) are grammaticalprecisely because the verb be is introducedby a rule like (8). (8) VP- V XP. That is, once we adopt the view that syntactic categories are feature bundles, then the 'archicategory'XP can be viewed as an underspecified syntacticcategory, but a syntacticcategory nonetheless.Such a category, if coordinated,is requiredonly to be a subset of each of its conjuncts. XP is a subset of NP, PP, VP and AP, as will be made clear in our discussion. Example (6c) on the other hand, is ungrammaticalbecause the verb sing is introduced by a rule like (9). (9) VP-- V(NP). Rule (9) containsno archicategorywhich could be extended to NP on the one hand, and ADVP on the other.' In the process of providinga formalstatementof this proposal,we will in fact present a comprehensivegrammarof coordinationin English. Our analysis, couched in the framework of Generalized Phrase Structure Grammar(GPSG),will touch on several key theoreticalissues which are of independentinterest, e.g., the theory of syntactic features, principles of feature instantiation,and the analysis of Ross's (1967) Coordinate StructureConstraintand-'Across-the-Board'convention, and the treatment of variouskindsof coordinatestructureellipses. Our discussionwill thus focus on integrating the analysis of coordination into a broader syntactic framework.The fundamentalgeneralizationswhich, we argue, govern the syntax of coordinate constructions are naturallyexpressed within the frameworkof GPSG, and provide an intriguingillustrationof its explanatorypower. In particular,we will illustratehow the generalization in (7) can be deduced from a widely assumed and independently motivated grammaticalprinciple,namelythe principlewhich governs the percolationof features in heads. I We are makingthe assumptionthat adverbialslike beautifullyare introducedby a rule like (i): (i) VP- VP ADVP. Our argumentwould remain intact under the alternativeassumptionthat sing could be introducedby a rule like (ii): (ii) VP-*V (NP) (ADVP). 120 IVAN 2. A. SAG ET AL. BACKGROUND GPSGs as characterized in Gazdar, Klein, Pullum & Sag (1985 'GKPS' henceforth), differ from simple CF-PSGs of the type standardly characterizedin mathematicaltexts in a number of ways, two of which are relevant in the present context. (i) (ii) Syntactic categories are not taken to be unanalyzableobjects representedby monadic node labels, but rather are viewed as sets of feature-valuepairs. The syntactic structures of a language are admitted by immediate dominance and 1inear precedence rules taken together with markingconventions and universalprinciplesof feature instantiation.Thus, there are no phrasestructurerules, as such, in a grammaritself. Both of these differencesare designed to permitschematization,in the interests of 'capturinggeneralizations'.Thus (i) permits non-disjunctive statements of cross-categorial parallelismsin a manner familiar since Chomsky (1970). Similarly, the function of (ii) is to allow for the formulationof very general rules, neutral with respect both to various feature values and to the order of constituentson the right hand side of the rules. Unlike many similarproposalsin previous work, however, our general rules are employed directly in the admissionof phrase structure trees, ratherthan being expanded into a full CF-PSG first.2 2.1. Syntactic Categories and Syntactic Features FollowingChomsky(1970), we shall treat the traditionalcategories Noun (N) Verb (V), Adjective (A), and Preposition/Postposition (P) as decomposableby means of a feature system that includes a feature [+N] which only N and A have, and a feature [+V] which only V and A have. This enables us, for example, to refer to the class of all nouns and all prepositions/postpositionssimply by writing [-V]. In phonology, complex symbols for phonological units are commonly taken to be sets of (featurename, featurevalue) pairs, and it is essentiallythis conception of syntactic features that we will maintain.Thus (N, +) is equivalent to the 2 GPSGs,as characterizedin GKPS, are equivalentto CF-PSGsin the sense thatone could alwaysinterpretone as defininga large CF-PSGassociatedwith syntacticstructuresin the standardway (in virtue of a trivialrule-to-treemappingof the sort that is seldomdiscussed, but nonethelessimplicitin standardformulationsof CF-PSG).The constructabilityof such a CF-PSGis relevantonly with respect to assessingthe abstractmathematicalpropertiesof the framework. COORDINATION AND CATEGORIES 121 more usual [+N]; similarly,notationslike (BAR, 2) and [BAR 2] may be used interchangeably.Where no ambiguity will result, we use [feature value] to abbreviate (feature name, feature value). Further, we use [feature name] to indicate that a category is specified for the feature named. We will assume without argument that the maximal bar level for categories is two, and we follow Borsley (1983) in treating sentences as V2s that are distinguishedfrom verb phrases solely in virtue of being positively specified for the feature SUBJ (although we will continue to use the symbol S to designate this category when no confusionwill arise). The only other non-standardelaborationof the theoryof featuresthat we shall adopt is the idea (takenfrom Bear 1981, Pollard 1982) that features may take (certain)categories as their values. We refer to such features as category-valued.3 It appears to be widely assumed that grammarsemploying features must adopt the following condition: (10) Only a fully specifiedcategory may label a node. However, we shall not make this stipulation.In the present framework, specificationslike (11) are just as much syntactic categories as those in (12). (1 1)a. {(SING,+)}. b. {(SLASH,{(BAR, 2)})}. (12)a. {(SING,+), (N, +), (V, -), (BAR, 2)}. b. {(N, -), (V, +), (SUBJ,-), (BAR, 2), (SLASH,{(N, +), (V, -), (BAR, 2)})}. Here, (12a) representsthe category of singular noun phrase, and (12b) the category VP/NP, familiarfrom earlierwork in GPSG. It may seem at first blush that such a relaxationof the theory of categories will create wild overgeneration. Thus is not true, however. As we shall see, principles of feature instantiationwill ensure that the featurespecificationsof lexical items, which we take to be fully specified in an appropriate manner, are inherited by the phrasal categories of which they are the heads. The rejection of (10) is the cornerstone of our analysis of coordination. I See GKPS chapter 2, section 6, for a detailed mathematical specification of the theory of features assumed in the present paper. 122 IVAN A. SAG ET AL. The notion of extension plays an importantrole in characterizingthe relation between the underspecifiedcategories mentioned in rules and their more fully specified counterparts in phrase structure trees. We define this notion informallyas follows: (13) A category C2 is an extension of a category Cl if an only if: (1) Every atom-valuedfeature-valuepair in Cl is in C2, and (2) For every category-valued feature-value pair in Cl, the value of the feature in C2 is an extension of the value of that feature in Cl. This recursive definition says first of all that any specification for an atom-valued(i.e., non-categoryvalued) feature in a category is also in all extensions of that category. It also guaranteesthat if a category specifies a value v for some category-valued feature, then any extension of that category specifies a value for that same feature that is an extension of v. Note that an extension of a category C may contain a specificationfor a category-valued feature which is unspecified in C. The notion 'is an extension of' is thus a generalizationof the notion 'is a supersetof' which takes proper account of category-valuedfeatures. An importantoperationon categories is that of UNIFICATION, a notion whose linguisticrelevance was first suggested by Kay (1979). It is closely analogous to the operationof union on sets except that, as in the case of extension, the resulting set must be a function. Unification is undefined for sets containingfeature specificationsthat contradicteach other. (14) The UNIFICATION of a set of categories is the smallest category which is an extension of every memberof the set, if such a category exists, otherwise the unificationis undefined. In (15) we list features which will be employed in subsequent discussion, together with the set of possible values for each. Boolean features (value set = {+, -}): N, V, PAST, AUX, PAS, INV, PRD, SUBJ. b. Category-valuedfeatures:AGR, RE, WH, SLASH. c. Unary features (value set = {+}): SING, XSP, THP, NULL. d. Other features: BAR {0, 1, 2} CASE {NOM, ACC} VFORM {FIN, INF, PRP, PSP, GER, BSE}. (15)a. The features BAR, N, and V are used to differentiateamong the major categories and their projections, as is standard in various versions of COORDINATION AND CATEGORIES 123 X-bar theory. Our use of CASE differs from that in much recent literaturein that we use it only for morphologicallyrealized case marking, which in English is limited to parts of the pronominal system. Categories containing a value for SLASH are to be thought of as dominating a 'gap' whose category is the value for SLASH; thus, for example, as already noted, (12b) is the category VP/NP familiar from earlier work in GPSG. The various values of VFORM serve to distinguish finite, infinitival, present participle, past (perfect) participle, gerund, and base forms of verbs and verb phrases (as discussed by Gazdaret al., 1982). PAST specifies a particularinflectionclass for finite verb forms. AUX and INV are featuresthat applyto auxiliaryverbs, also as discussed by Gazdar et al., 1982). PAS distinguishespassive verbal forms from others. PRD distinguishes predicative categories, about which we will have considerablymore to say later. SUBJ, as we mentioned above, is used to distinguish S from VP, since both are being treated as V2s. RE and WH mark constituents that either are, or else properlycontain, reflexive/reciprocalor wh elements, respectively. Note that both these features are category-valued,as is AGR, the agreement feature present in numerous syntactic categories; verbs and VPs will contain (AGR, N2); DETs contain (AGR, N'). AGR plays a central role in the theory of semantically-basedagreementoutlined in Sag and Klein (1982). SING is used to mark singularnouns: the only permissiblevalue for SING is +. We assume that pluralnouns are distinguishedby lacking a specification for SING. XSP and THP are person features whose semantic correlates are 'excluding the speaker' and 'third person', respectively. We will explain the functioning of these unary features in section 4.2 below. It should be noted that we do not impose a hierarchicalstructureon categories, of the sort presented in Gazdarand Pullum(1982). However, we presume that categories are classifiedin ways that allow grammatical principlesto make reference to certain subsets. In particular,we want to make reference to the subset of a given category C that involves all the HEAD features present in the domain of C. Likewise, we want to distinguishsubsetsof categories that involve all and only FOOT features. These two classes will play an importantrole in the feature distribution principles we formulate in section 2.4, their members are listed in (16) and (17). (16) HEAD = {N, V, BAR, SING, XSP, THP, CASE, VFORM, PAS, PAST, AUX, INV, PRD, SLASH, AGR, SUBJ, PRD}. (17) FOOT = {SLASH,RE, WH}. 124 IVAN A. SAG ET AL. 2.2. MarkingConventions Our theory of syntactic categories will employ both FEATURE CORESTRICTIONS (FCRs) and FEATURE SPECIFICATION DEFAULTS (FSDs). However the two kindsof markingconvention have a rather different formal status: FCRs are substantive constraintson the sets of feature-valuepairswhich constitutewell-formedcategories. Some FCRs will be universaland thus be part of a characterizationof 'possible natural language syntactic category', and some will be language-particular, and thus be part of a characterizationof, for example, 'possible syntactic category in English.' A generalization such as 'only nominals inflect for case' would be stated as an FCR, notated as in (18). OCCURRENCE (18) FCR: [CASE] = [+N] & [-V]. The interpretationof (18) is that any category containing a value for CASE will also contain the values + and - for N and V, respectively. This ensures that something of the form of, for example, {[+V], [-N], [NOM]}is simply not a well-formedsyntactic category. Another example of an FCR is (19): (19) FCR: [+INV] v [+AUX] & [FIN]. This says that a category which carriesthe feature [+INV] will also carry the features [+AUX] and [VFORM FIN]. [+INV] is a feature that appearson sentences which include a subject but begin with a verb (see Gazdaret al., 1982), and also appearson that sentence's initialverb. This FCR thus has as a consequence that such a verb will alwaysbe a tensed auxiliaryverb. Let us turn our attention now to the second type of marking convention, namelyFSDs. These form an importantpart of the link between the highly schematicrules listed in the grammar,and the trees they admit (as outlined in the next two sections). As the firstillustration,consider the phrasestructurerules in (20) and (21) in relation to the feature INV. (20) VP-* V S. (21) S[+INV]-). V[+AUX] NP VP[BSE]. We will use the expression treefragmentto refer to fragmentsof a tree which consist simplyof a mother node and all of its daughternodes. Two of the tree fragmentsthat we might expect to get from (20) are displayed in (22). COORDINATION (22)a. AND CATEGORIES 125 VP V b. S [-INV] VP V S [+INV] That is, there is no particularreason for having [-INV] rather than [+INV] on the S daughter. But if we allow the grammarto admit tree fragmentslike (22b), then we will end up generatingexamples like *Lee believes will the children be late. So INV needs to have a default specification,namely [-INV]. We can state this default as follows: (23) FSD: [-INV]. Since there is no reason for INV not to have that specificationon S, it must have it, accordingto the approachto defaultsthat we assumein this paper (see GKPS, chapter 5, section 5 for a precise formal theory of FSDs). There is, of course, one class of structuresin which [+INV] is obliged to be present, namely those arisingfrom rules such as (21). But, since this rule stipulatesthe presence of [+INV], the default will not be invoked. In general a feature is exempt from assumingits defaultspecificationif it has been assigned a differentvalue in virtue of some ID rule or some principle of feature instantiation. Oversimplifyingsomewhat, suppose that in every tree fragment legitimated by rule r, the instantiatedcategories C' that correspondto some category C in r, always agree on the feature f. Then we can conclude that the value of f is indeed the way it is as a result of some phrase structure rule or principle of feature instantiation:such a feature specificationis privileged.By contrast, if the categories in the various tree fragmentsresultingfrom r assign different values to f, values which vary independently of anything else in the fragments, then we want to exclude those in which f fails to take its default value. It will be helpful to illustratethis fundamentalpoint with an example. (24) FSD: - [CONJ]. Consider (24). Sometimesthe default situationfor a given feature is for it not to be present at all, i.e., for it to be absent from the domain of a given category. In such cases, we will speak of a feature as being 126 IVAN A. SAG ET AL. 'default-absent'.Loosely speaking, (24) says that the default specification for CONJ is to have no specificationat all. We interpretthis to mean that if it is possible for a category in a candidate projection to lack a CONJ specification, then that category must lack a CONJ specification.Or, in other words, a CONJ specification is only permissibleif it is explicitly required by a rule. Defaults excluded, nothing prevents (9) from being instantiatedas the tree fragmentin (25). (25) VP V [CONJ or] NP [CONJ and] On the other hand, we can easily see that there are candidate tree fragments induced by (9) which lack these CONJ specifications. Since CONJ is not requiredby the rule to have anythingother than its default specification,our convention for defaults says that CONJ must have its default specification, i.e., must be absent altogether. Consequently,we will not generate strings like *Lee or sang and "Rule Brittania". So far, we have not said anythingabout defaults occurring on mother categories. Suppose that we allowed a CONJ specificationto freely occur on the root of a tree fragmentarisingfrom (9): (26) VP [CONJ and] V NP If this were permissible,it would allow us to generate stringssuch as *Lee both plays piano sings "Rule Brittania"where there is a coordinating conjunction missing from the position between piano and sings. Examples like these show that FSDs must apply to mother categories in a rule as well as to daughters.Once again, the principleof defaultsthat we have suggested will exclude the feature specificationillustratedon the mother of (26), since nothing obliges [CONJ and] to be there. In fact, the only way in which a CONJ specificationis ever present is by virtue of a rule (such as (49a), below) which explicitly requiresthe presence of a CONJ specification. Although this treatment of defaults gives the right results in most cases, there is an importantclass of cases for which it is still inadequate, namely those in which a given feature is found in two distinct categories in the tree fragment, and the values of the two occurrences covary. Feature instantiationprinciples such as the Head Feature Convention COORDINATION AND CATEGORIES 127 give rise to instances of such covariation. For example, the situation occurs whenever the grammarpermits recursion via a head category: rules introducing adverbial modifiers typically have this form, as do coordinationrules. This leads us to define a more liberalnotion of default satisfactionthan that presupposedin our discussionso far. Under this more liberal notion, a tree fragment meets the defaults if and only if for every category and default, (i) the default is true of the category, or (ii) no candidate tree fragmentsexist in which the default is true of that category, or (iii) the default is false of the category but making it true would necessitate changing some other category in the tree fragment.4 2.3. Rules A phrasestructurerule of the familiarsort specifies two distinctrelations: (i) immediate dominance relations, and (ii) linear precedence relations among sisters. Consider,for example, the rules shown in (27): (27) A-*BCD B--ACD C-*ABD D -ABC. Inspectionshows that a generalizationcan be made about the set of trees admitted by this grammar, namely that the sister constituents always appearin an order that happensto correspondto the order of the letters A, B, C, D in the Roman alphabet.This generalizationis not expressed by the grammarshown in (27). Following Gazdarand Pullum(1981), we adopt a mode of analysisthat factors out the two relations of immediatedominance and linear precedence which are conflated in phrase structurerules of the familiarsort.' For immediate dominance, we use the format shown in (28), which we shall call an immediatedominance(ID) rule. (28) A -B, C, D. This rule specifies part of the conditions that must hold of a structure rooted in A: namely, that it consist of exactly three daughters whose categories are B, C and D, respectively.However, it does not in itself say anything about the linear order in which B, C and D must occur under A. For linear precedence, we introduce the antisymmetric,transitive 4 See GKPS chapter 5, section 5, for a detailed mathematicaltreatment of default satisfaction. ' Falk (1993) presentsa similarproposal,developed independentlyof Gazdarand Pullum. 128 IVAN A. SAG ET AL. relation "<", where "A < B" is to be read as "A must precede B if they share the same mother". The structures admitted by a set of ID rules are constrained by a further set of rules, which we refer to as Linear Precedence (LP) rules. The structuresadmittedby an ID rule grammarare just those which are consistent with some ID rule and all LP rules. B, C, D A,C,D A,B,D A, B, C. (29) A B C D (30) A<B<C<D. Taken together (29) and (30) are extensionally equivalent to grammar (27), i.e., the two admit exactly the same set of tree fragments.But the ID/LP grammarin (29)-(30) does what (27) does not do: it expressesthe generalizationabout sister constituentorder. We adopt the essential insight, but not the exact practice of Gazdar and Pullum (1981) in allowing general feature instantiationprinciplesto ensure that major category features, as well as other syntactic features, are identical on mothers and heads in ID rules. We use the symbol H to refer to daughtersthat are designatedas heads in ID rules, and adopt the obvious usage of referring to an element of a syntactic structure as a head just in case its correspondingelement is a head in the ID rule used to admit that structure.Thus we will have ID rules like those in (3 1):6 (3 1)a. b. c. d. S ->NP, H[-SUBJ]. S[+INV] -* Ho, NP, VP[BSE]. VP-- H,NP. VP-) H0, NP, VP[INF]. And feature instantiationprinciples (such as the Head Feature Convention, discussed below) taken together with the LP rules in (32), guarantee that these admit tree fragments whose form is shown in (33) (irrelevantdetails not shown). 6 We use the symbols X, Y, and Z to designate underspecified categories. X' stands for {(BAR, 1)}, XI[+INV] for {(BAR, 0), (INV, +)}, and so forth. Abbreviations such as H', H'1+INV], etc., are no different from these, apart from the further information conveyed that a given element is a head in its ID rule. We reserve the symbols C, C", etc., for use as variables ranging over arbitrary syntactic categories in the statement of metagrammatical principles. See GKPS, chapter 3, section 4, for a technical reconstruction of the theory of heads presupposed here. COORDINATION (32)a. AND CATEGORIES 129 X0 < X2. b. N2<V2. S [FIN] (33)a. NP VP [FIN] S [+INV, +AUX] b. V0 [+INV, +AUX] c. VP [BSE] VP [FIN] V [FIN] d. NP NP VP [GER] V [GER] NP VP [INF] 2.4. The Head Feature Convention The Head Feature Convention (HFC, hereafter)is part of the mapping from ID rules to structures.That is, it imposes certain conditionson how categories may be assigned to nodes in a tree fragment. Like other feature instantiation principles, the HFC narrows down the class of structureswhich are compatiblewith a given ID rule. In order to give an informalaccount of what it means for a structureto meet the HFC, we will firststate an oversimplifiedversion, and then progressivelyrefine it.7 To begin with, we will assume that every rule introduces exactly one head. Then the simplestconceivable version of the HFC will just require identity between the HEAD features on the mother and those on the head daughter. ' What follows is an informal discussion of the HFC as presupposed in the present paper. Full technical details and an extended formal development of the issues treated here can be found in GKPS, chapter 5, section 4. 130 (34) IVAN A. SAG ET AL. HFC: version 1 The HEAD features on the mother are identical to those on the head. The problem with this definitionis that it enforces an absolute identity, and makes no allowance for the fact that the daughter may be independently required to carry (or not to carry) a HEAD feature specificationwhich (or whose absence) is incompatiblewith those on the mother (i.e., no legal extension of the mother carries (fails to carry)such a specification)or the mothermay be requiredto carry(or not to carry)a HEAD feature specificationwhich (or whose absence) is incompatible with the daughter. These requirements can arise either because the 'problematic'feature specificationis stipulatedin the rule, or because its presence or absence is required by FCRs, or because its presence or absence is required by the other principles. If we leave the HFC as a simple identity statement, then rules that give rise to 'problematic' feature specificationswill simply not play any role in the admissionof structures. To avoid this consequence we need to modify the HFC so that it only seeks to equate those HEAD feature specificationswhich can be freely equated. In looking at the head we must restrict attention to those feature specificationswhich can appearon the mother, and in looking at the mother we must restrict attention to those feature specificationsthat can appearon the head. In order to restrictour attention in this way, we need a notion of 'free' feature specifications, i.e., the set of feature specificationsthat can be instantiatedon the category in the context of the rule in which it finds itself. Intuitively,the free feature specifications on a category are the featurespecificationswhich can legitimatelyappear on extensions of that category: feature specificationswhich conflict with what is alreadypart of the category, either directly, or in virtue of other feature principlesof FCRs, are not free on that category. Given this notion of free features, our revised definition needs to run as follows. (35) HFC: version 2 (i) The HEAD feature specificationson the head are an extension of the HEAD features of the category created by taking the intersection of the mother with the free feature specificationson the head. COORDINATION AND CATEGORIES 131 (ii) The HEAD feature specifications on the mother are an extension of the HEAD features of the category created by taking the intersection of the head with the free feature specificationson the mother. This achieves the result we want. Unfortunately,such a definitiononly caters for the situationin which a rule has a single head. Catering for the possibility of multiple heads complicates matters considerably. If we ignore the issue that 'problematic'feature specificationsgive rise to, then a multi-headedversion of the HFC has a straightforwarddefinition. (36) HFC: version 3 The HEAD features on the mother are identical to the HEAD features on the category that results from taking the intersectionof all the head daughters. Notice that in the single-headed case, this definition simply reduces to the one we started out with (since the intersection of a singleton whose only memberis a category is that category itself). But, as we have said, it makes no allowance for 'problematic' feature specifications. A definitionwhich does make the appropriateallowance is given below. (37) HFC: version 4 (final version) (i) The HEAD feature specificationson each head are an extension of the HEAD features of the category created by taking the intersection of the mother with the free feature specificationson that head. (ii) The HEAD feature specifications on the mother are an extension of the HEAD features of the category created by taking the intersection of the heads with the free feature specificationson the mother. This definition reduces to version 3 in the case of multiheaded constructionwhich does not involve 'problematic'features, and it reduces to version 2 in the case of single-headed constructions.In the special case of a single-headed construction that does not involve 'problematic' features, the definition reduces to version 1. The possibility of these reductions shows that the definition maintains the advantages of the simplerformulations.However, it also overcomes their limitations. In order to illustrate the HFC, we will confine our attention for the moment to single-headedconstructions.Considerthe pair of rules in (38) and (39): 132 A. SAG ET AL. IVAN (38) VP-- Ho,S[BSE]. (39) S -- NP, H[-SUBJ]. Rule (38) is the ID rule which introduces verbs like insist that take a complement clause whose head is in base form (cf. I insist that they be here by noon), while (39) is the familiar S expansion rule. Two of the structurescompatiblewith (38) are the following: VP [FIN] (40) S [BSE] X? VP [FIN, -PLUR] (41) V0 [FIN, +PLUR] S [BSE] Both of these trees fail to meet the HFC with respect to rule (38). The set of free HEAD featureson the mother category in (40) is {(N, -), (V, +), (VFORM, FIN)}. Note that (BAR, 2), though a HEAD feature and present on the mother, is not a free HEAD feature in this example, since the lexical head already contains (BAR, 0). Example (40) fails to meet the HFC on rule (38) because none of the free HEAD featuresfrom the mother have been instantiatedon the head daughter.On the other hand, (41) fails because the set of free HEAD featureson the mother category is {(N, -), (V, +), (VFORM,FIN), (-PLUR)} whereas the instantiated free HEAD featureson the head daughterare {(N, -), (V, +), (VFORM, FIN), (+PLUR)}. These two sets intersect, but neither is an extension of the other, and (41) fails to meet the HFC with respect to the rule in (38). Next, examine the tree fragmentsin (42) and (43). VP [FIN] (42) S [BSE] V [FIN] S [BSE] (43) NP VP [BSE] Example (42).does meet the HFC on rule (38), and (43) meets the HFC on rule (39). Consequently, the rules together will admit the tree (44), COORDINATION AND CATEGORIES 133 because each tree fragmentin (44) is compatiblewith the rules. (44) VPfFIN] S [BSE] V [FIN] NP VP [BSE] Imaginewe were randomlygeneratingall the tree fragmentsadmittedby the two ID rules just discussed. There is nothing to guarantee that the mother in any given tree fragmentcompatiblewith the S expansionrule will match the complementdaughterin a given tree fragmentcompatible with the VP expansionrule. And indeed none of the feature instantiation principlesin our grammaris intended to ensure such a matching. But in order for the grammarto admit the tree (44), it is quite sufficientif each of its component tree fragments turns out to be one of the many such structurescompatiblewith the relevant ID rules, and this we can see to be true. What the HFC ensures in these cases is that we get a [VFORM FIN] specification on the VP category if and only if we get the same specification on the V category, and that we get a [VFORM BSE] specification on the S category if and only if we get it on the VP category. What happens in this example is that the subcategorization requirementfor a bare infinitive demanded by verbs of this class gets transmitteddown to the VP head of the S complement in the way just described. Analogous reasoning shows that the requirementwould also be transmitteddown to the lexical head of any VP that sproutedfrom the appropriatenode in (44). 3. THE SYNTAX OF COORDINATION Traditionalgrammariansrefer to a constructioncalled "coordinateconjunction" or simply "coordination".All languages, as far as we know, make use of constructionsof this type.8 This apparentlyinnocent claim ' We include here not only languageslike English,Japanese,and Latinin which there are specific morphemesreservedfor logical conjoining and disjoiningof statements,but also languages like Dyirbal (Dixon 1972; see e.g., p. 154) in which coordinate constructions appear without overt conjunction morphemes like English and. Note also that some languages use an overt coordinatingmorphemefor NPs but not for clauses; Hausa is an example: see Abraham (1941, p. 92), Kraft and Kraft (1973, p. 330). Realization of coordinatingmorphemes is a highly parochial matter. This paper concerns itself with mattersthat we take to be much less parochial,and in manycases probablyuniversal. 134 IVAN A. SAG ET AL. implicitly embodies an important metatheoretical assumption to the effect that there is a unitary notion of coordination,one that abstracts away from the evident differencesbetween the coordinate constructions of, say, English, Japanese, and Latin. In this section, we give an outline of our theory of the coordinateconstruction. It is worth observing that transformationalgrammarhas never been able to capture a unitary notion of coordination,for reasons that were fundamentalto the nature of the theory. Consider the following examples. (45) Kim sang and Sandydanced. (46) Kim and Sandymet. (47) Kim sang and was accompaniedby Sandy. Examples (45) and (46) would have been, and could only have been, generateddirectly by base rules. But example (47) had to be derived in a completely different way, via a transformationof Conjunction Reduction, in any grammarthat handledpassive constructionstransformationally. Analogous triads of examples can be constructed for almost every transformationever proposed, so the problemis not specific to passives. Conjunction Reduction, though more often assumed than formulatedin the transformationalliterature,had to be formulatedin such a way as to produce structuresthat were isomorphicto those that would have been produced if everything had been base-generatedin the first place. Thus obvious generalizations,such as that VPs participatein coordinate constructionsof just the same sort as NPs and Ss (and precedes the last VP in the coordinatestructure,for instance, ratherthan following it or being infixed into it) fail to be expressed. The structures needed can be generated by brute force adjustments,but similarityof coordinatestructures across categories is not therebyexplained. Recent transformationalgrammarhas largely abandonedcoordination as a topic of study, and what work there is bears a marginalrelationship to the mainstreamof work in that paradigm.Thus George (1980) offers a highly problematic approach in which everything but sentential coordination is derived by deletion from larger paraphrases.9More recently, Goodall (1983) has proposed a treatment of coordination involving sentence (actually phrase marker) unions that he claims are not two' For example, George is forced to invoke a transderivational constraint, though not eo nomine,in order to get the subject-verb agreement right in a sentence like Kim and Lee like Koreanfood. COORDINATION AND CATEGORIES 135 dimensionallyexpressible.On his account, conjuncts get realized linearly in the phonological component (see p. 146: 'the question of how the phonology interprets union of sentences'). Since at least the syntax of constituent order and the syntax of agreement interact with the linearizationof conjuncts, it is clear that much of contemporarysyntax is in fact a branch of phonology from Goodall's perspective."' Williams (1978, 1981) defends another transformationaltheory of coordination which abandons tree-representablephrase markers in favor of objects whose precise character is left open. We have discussed Williams'proposals at length elsewhere (Gazdar et al., 1982) and will not repeat ourselves here. What strikes us most about these three proposals,and that of Pesetsky (1982) which we consider below, is (i) the way in which their authors embrace formal devices that have little or no precedent within the framework within which they work, and (ii) the extent to which the formaldetails of their proposalsand their consequencesfor other aspects of the grammarsimply have not been worked out in a serious way. 3.1. Coordination Schemata The theory of coordinationwe present is able to locate all the parochial aspects of coordinate constructionsin two components of the grammar that constitute natural repositories for parochial facts, namely (i) rules responsible for realizationof specific morphosyntacticfeatures, and (ii) principlesdetermininglinearprecedence among constituents.Everything else, including the rule schemata that define the internal structure of coordinate constituents, is handled by mechanisms that we claim are universal. We begin with a terminologicalclarification.Informally,we shall use the word conjunct to refer to a constituentparticipatingin a coordinate construction,and we shall refer to items like and, or, but, either, neither, nor as conjunctions.Despite the fact that, for example, or correspondsto logical disjunctionrather than conjunction, there seems little chance of confusion arising throughthis usage. Our syntax of coordination makes use of a feature CONJ, whose values (in English) are membersof the set in (48). (48) {and, both, but, NIL, neither, nor, or}. Goodall derives John drinks beer and Mary whiskey from the same source as John and Mary drinkbeerand whiskeyas alternativephonologicalinterpretations;see his examples (9a), (9b), and (9'). 136 IVAN A. SAG ET AL. The rationalefor including NIL will become apparentin the discussion below. Categories whose domain includes CONJ are expanded by the following rules: (49)a. b. X[CONJ NIL]-- H. X[CONJ a] -a, H where a e {and, both, but, either, neither,nor, or}. These rules spell out a value of CONJ as the appropriateconjunction. The fact that CONJ is default-absentwill ensure that no category that is an instantiationof the second daughterin (49b) will contain CONJ in its domain, hence blocking unwanted iteration of conjunctions. As noted above, this default guarantees that instantiated categories contain specifications for CONJ only when CONJ is mentioned in the preinstantiationID rule. Note that a structurecorrespondingto the second daughter in (49b) is an unspecified head daughter of a mother which is also unspecified for HEAD features and hence must agree with its mother vis a vis all HEAD features (includingBAR level). We distinguishtwo sorts of coordinate construction.In one, there can be only two conjuncts. In the other, there is no limit to the number of conjuncts permitted.We postulate two rule schemata, one for each type of coordination.These schemata are exhibited in (50a) and (51a). The first is for the arbitrary-lengthcoordinate structures,and the second for the binary ones. In (SOb)and (51b) we list the values that the CONJ feature has in English for the two constructions (actual occurrence of coordinationmorphemesbeing a highly parochialmatter).' (50) Iterating CoordinationSchema (CS') a. X-* b. aE{(and, NIL), (NIL, and), (neither, nor), (or, NIL), (NIL, or)}. (51) H[CONJ ao], H[CONJ al]l. Binary CoordinationSchema (CS2) a. X -H[CONJ ao], H[CONJ al]. b. a is in {(both, and), (either, or), (NIL, but)}. Among the possible values for CONJ mentioned in these schemata is NIL. Whereas NP[CONJ and] dominates terminal strings such as and theirdog, NP[CONJ NIL] dominates strings like theirdog. In English, this is reserved for nonfinal conjuncts but there are languages, as l l We use a0 and a, to designate the first and second membersof an ordered pair a. '+' designatespositive Kleene closure on multisets,as definedin GKPS chapter 3, section 4. COORDINATION AND CATEGORIES 137 mentioned above in footnote 8, in which all conjuncts take [CONJ NIL] in a coordinatestructure. Notice that the variables a0 and a1 here range only over the lists of specific morphemes that can mark pairs or sets of conjuncts. In all previous proposals for coordination schemata that we know of, coordination was stated on variablesranging over categories which enforced categorial identity across the conjuncts, by means of an implicit uniform substitution principle. For instance, there is some discussion of the semanticsfor coordinationschematain Keenan and Faltz (1978), Gazdar (1980), Cooper (1979), Partee and Rooth (1983), and Rooth and Partee (1982), and all of them, as far as we can tell, assume variables across categories. The present proposal does not: X is here, as before, not a variable over the set of categories, but a category - albeit a minimally specified one. In effect, if we abstractaway from the details concerning the values for the CONJ feature, all that (50a) and (51a) say is that the mother is a category, and that the daughtersare all heads. If completely free feature instantiationwere permitted(which of course it is not), then (Sla) would in principle be consistent with, for example, a coordinate structurein which a prepositionexpandedas an NP conjoined with a VP. The only informationthat the two schemata contain concerns the distributionof CONJ and the possibilityof iteration.A consequence of this is that the schema in (51a) collapses exactly three English coordination rules, namely those arrived at by substitutingin the tree possible value pairs for a0 and a,. These three rules will correspond to numerous distinct instantiationsdefined by feature instantiation,but they are not schemataover those instantiations.We are stressingthis point because it has importantempirical consequences to which attention will be drawn subsequently. In particular, the approach we have adopted, abjuring variables over fully specified categories, does not entail that every conjunct be categorially identical to each of its sisters, although nearidentityusuallyfollows as a consequence of the interactionof the various principlesof feature instantiation.This point is pursued in detail below. One further parochial component to our analysis of coordination is needed in order for us to be able to explore the claims it makes concerning the structureof English. We need LP statements to express the ordering constraints that hold across the various types of conjunct characterizedby distinct values for the feature CONJ. These LP statements can be collapsed into a single schema. (52) [CONJ ao] < [CONJ aJ] where a0 E {both, NIL, either, neither}, and a1 E {and, but, nor, or}. 138 IVAN A. SAG ET AL. Our three coordinationschematainteractto make a very wide range of detailed predictions concerning possible and impossible coordinate structuresin English. We will illustratethese predictionsby reference to examples involving coordinateVPs, and concern ourselves only with the predictions made with respect to iterability and choice of conjunction morpheme,since these are the issues that (50) through(52) address.The categorial identity, or lack of it, between mother and conjunct, and between conjunct and conjunct, is a topic that we leave to the next section. Here we will simply assume that constituents of the same category can conjoin to form a coordinate constituent of that category. Since there are eight distinct values for CONJ in English, it follows that there are 64 logically possible two-conjunct coordinate structures. However, only six of these 64 possibilities are, in fact, grammatical, namely the six illustratedin (53). (53)a. b. c. d. e. f. made a speech and stuttered made a speech or stuttered neither made a speech nor stuttered both made a speech and stuttered either made a speech or stuttered made a speech but stuttered The CS' and the LP-schema in (52) interact to generate the structures illustratedby (53a, b, c) and no others. The CS2 and (52) interact to generate the structuresillustratedby (53d, e, f) and no others. Thus the schematawe have given induce all and only the six grammaticalEnglish two-conjunctcoordinatestructures.12 We now turn our attention to three-conjunct examples with flat constituent structure.'3Here there are 512 logical possibilities. Of these, only five are grammatical.They are illustratedin (54). 12 But cannot be used to coordinate{[+N], [-VJ} categories (*Kim but Sandy stuttered), and bothcannotbe used to coordinatefull sentencesthat lack complementizers(*BothKim sang and Sandy danced).There are other such idiosyncraticfacts. As far as we can see, they can readilybe handledby meansof FCRs insofaras they do not follow from anything semantic,but we do not dwell on them here because they appearnot to illustrateanything interestingabout the principlesunderlyingcoordinationin English. 13 Obviouslythere are three-conjunctexamplesthat involve a two-conjunctstructurewith a furtherconstituentconjoined to it, and so on. To some extent semanticand intonational tests can be used to determinewhich structurewe would want to assumefor a given string used in a given context. The question we are addressingis how to account for multiconjunctcases that do not show any signs of being groupedhierarchicallyinto pairs. COORDINATION (54)a. b. c. d. e. AND CATEGORIES 139 whimpered,shouted, and screamed whimperedand shouted and screamed whimpered,shouted or screamed whimperedor shouted or screamed neitherwhimperednor shouted nor screamed The CS2 is irrelevantfor the flat structuresof these examples. However, the CS' interactswith the LP-schema in (52) to give us these five types of coordinateconstituent,but not any of the other 507 possibilities. Consider now flat four-conjunct constructions.Here there are 4096 logical possibilitiesfor assigningconjunctionsto the conjuncts, but again only five are grammatical,and just those five are legitimated by the interactionof the CS' and the LP-schema. (55)a. b. c. d. e. moaned, whimpered,shouted and screamed moaned and whimperedand shouted and screamed moaned, whimpered,shouted or screamed moaned or whimperedor shouted or screamed neither moaned nor whimperednor shouted nor screamed There is some variationamong speakersof English with respect to either and neither.More liberal varieties than our own allow the examples in (56). (56)a. either whimperedor shouted or screamed b. either moaned or whimperedor shouted or screamed That is, they flout the familiar prescriptive injunction not to use either.., or with more than two disjuncts.14There may also be people who are less liberal than us with respect to neither... nor, and are not prepared to use it iteratively. Such varieties are straightforwardlydescribed by makingminorchanges to the parochialcomponentsof the CS' and the CS2. Thus, to increase liberalitywith respect to either. . . or, one simply mentionsit in the value specificationsfor the CS' ratherthan the CS2, and to decrease liberalitywith neither... nor, one makes the relevant move in the opposite direction. Finally, we wish to comment on the semantic interpretationof the '4 As pointed out to us by Ed Keenan, there appearto be varieties of English in which exampleslike (i) and (ii) are used. (i) (ii) either whimpered,shoutedor screamed neitherwhimpered,shouted nor screamed The rules we have given do not take these possibilitiesinto account. 140 IVAN A. SAG ET AL. coordination schemata we have introduced. Without dwelling on the topic here, let it be made clear that a familiar Boolean semantics for coordinate conjunctionsis available for and entirely consistent with our syntactic proposal. The proposals made by Keenan and Faltz (1978), Gazdar(1980), Cooper (1979), Partee and Rooth (1983) and Rooth and Partee (1982) are all adaptable to our syntax. The one unfamiliar assumptionthat must be made is that features play a role in the semantic interpretationprocess. Thus it is the value of the feature pair a used in generating the coordinate structure that contributes the appropriate Boolean meaning to the interpretationof that structure. An analysisalong these lines is motivatedby the fact, noted by Carlson (1983), that it is often the case in languages that a single Boolean conjunction meaning is signalled by multiple occurrences of given conjunction. Indeed our analysisof English has just this property(e.g., in a "flat"structurewhere two or more conjuncts are markedwith and). The apparentproblemis solved once we assume that it is the value of a that contributesthe conjunctionmeaning,ratherthan each occurrence of the particularconjunction. 3.2. Featuresin Coordination The coordination schemata just introduced deal only with the distribution of the coordination morphemes and with the possibility of iteration of conjuncts. Nothing was said about the categorial status of conjuncts, or about the category of the mother given the categories of conjunct daughters. The conventional wisdom on this topic has it that conjunctsmust all be of the same category, say a, and that the motherof these conjuncts will also be of category a. But the conventional wisdom is wrong, or at best, seriously incomplete, for reasons that are fairly widely known. However, in the absence of any other candidate analyses, it has not so far been replaced. There are at least two classes of phenomena that show that the conventionalwisdom is wrong. First, the person, number,and gender of coordinated NPs do not behave in the manner it would lead one to expect. Thus, for example, a singularNP can conjoin with a pluralNP (or with another singular NP) to form a plural NP. We postpone our discussionof this puzzling,long-standingproblem until section 4. The second class of phenomenaincludes examples like those we noted in the introductorysection, all of which involve what one might call predicativeexpressions,as in the following examples: (57) We walked slowly and with great care. [Adv and PP] COORDINATION (58) AND CATEGORIES 141 Terry turnedout to be longwindedand a bully. [AP and NP] Two questions immediatelyarise: what is the category of the mother of the conjuncts in each case? And why is the coordinationof AP with NP seen in (58) not possible in (59)? (59)a. *The longwindedand a bully man was my brother. b. *Soon longwindedand a bully startedshouting again. It is questions of this sort that our theory of coordinationaddresses. We assume that the verb be is introducedby the following ID rule:15 (60) VP__ HO,X2[+PRD]. PRD is the feature mentioned earlier whose existence is presupposedin Bresnan (1973). X2[+PRD] may be realized as N2[+PRD], A2[+PRD], P2[+PRD], or V2[+PRD]. We assume that an FCR guarantees that all predicativeV2s are either passive or present participialin form (see FCR 14 in chapter 6, section 2, of GKPS). Hence the rule in (60) plays a role in the syntactic analysisof all the examples in (61): (61)a. b. c. d. e. Kim was a banker. Dana was quite competent. Leslie was in the flood zone. Ronnie was talking to Lou. Jean was given a prize. The verb becomeis introducedby the rule in (62). (62) Vl -' H?, X2{[+PRD], [+N]}. (62) requiresthat the complementof becomesbe a predicativeN2 or A2. (63) Pat has become a Republican. (64) Gerry became quite conservative. (65)a. *Connie has become of the opinion that we should get out. b. *Tracy became awardeda prize. c. *Chriswill become talking to colleagues. Given the rules and feature instantiationprinciplesdeveloped so far, it is now clear why the coordinationof unlike categories that we noted in the introduction is possible. In both coordination schemata we have presented, every conjunct is a head of the mother of the coordinate '5 We follow our earliernotationalconventions:X2[+PRD] standsfor {[BAR.2],[+PRDD. 142 IVAN A. SAG ET AL. structure(cf., Farkaset al., 1983). The HFC guaranteesthat the HEAD features of each conjunct will be a supersetof those of the mother of the coordinatestructurein any structureadmittedin virtue of either the CS' or the CS2. This in turn guaranteesthat in the most common case, where a verb takes a complement that is fully specified with respect to major features,e.g., a NP or an AP, then if such a complementis expandedinto a coordinate structure,each conjunct will have to contain all the major features of the coordinatemother, and will hence be of exactly the same major category as that mother. In the case of the rules just given for be, become,and consider,however, no such conclusion can be drawn.Since we allow partially-specifiedcategories, an X2[+PRD] introducedas the complement of be, for example, can be expanded into a coordinate structurewhere the conjuncts are of different major categories, as long as each such conjunct contains X2[+PRD]. Thus structuresof the sort sketched in (66) are allowed: (66) s VP NP Pat V( is X2 [+PRD] N2{[+PRD], N2 [+PRD] a republican [CONJ NIL1 and A2{[+PRD], [CONJ and]} A2 [+PRD] proud of it The coordinate structure in (66) satisfies the HFC because the intersection of the HEAD feature specificationsof the N2 and the A2, both of which are heads of the X2, is exactly the set of HEAD feature specifications of the partially specified X2 mother, namely {[BAR2], [+PRD], [+N]}. And because the rule in (62) specifies that the complementof become is X2{[+PRD],[+N]}, any conjunct of such a complement must contain X2{[+PRD],[+N]}. This explains such contrastsas those in (67): (67)a. Pat became a republicanand quite conservative. [N2 and A2] b. *Tracy has become a republican and of the opinion that we must place nuclear weapons in Europe. [N2 and p2] COORDINATION AND CATEGORIES 143 c. *Chris became quite conservative and trying to change their minds. [A2 and V2[PRP]] d. *Gerry became a republican and awarded a prize. [N2 and V2[PAS]] Similarfacts can be constructedfor all verbs which subcategorizefor less than the full range of X2[+PRD]s. Consider again examples like (57) [repeated here]: (57) We walked slowly and with great care. Coordinationof unlike categories here is predicted by our theory under either of two plausible analyses of such adverbials. If the grammar contains a rule like (68), (68) ADV2 __ p2 then examples like (57) are analyzableas coordinationof like category, namelyADV2. However, as suggested to us by Donka Farkas,there is an alternativeapproachto adverbials,consistent with our syntactic analysis of coordination, that eliminates the need for rules like (68). One need assume only that there are adverbial features like MANNER which cooccur with various combinations of major features, and that syntax rules introducingmanner adverbialsintroduce X2[+MANNER]. Such a constituent may then be expanded in such a way as to generate ADV2[+MANNER] as one conjunct, and P2[+MANNER] as the other, as in (57). A similar analysis of temporal adverbials is possible, which treats expressions like yesterday, the next time I see you, every chance they get, and the like, as instances of N2[+TEMP], and phrases like on Tuesday, in time, etc., as P2[+TEMP]s. This treatment of temporal adverbials, taken together with our coordination analysis, would explain the grammaticalityof exampleslike those in (69): (69)a. They wanted to leave tomorrowor on Tuesday. b. We are open Saturdays,any nationalholiday, and on alternate Sundays. As for exampleslike (59), repeatedhere, there is now a simple account of their deviance: (59)a. *The longwindedand a bully man was my brother. b. *Soon longwindedand a bully started shouting again. Neither the rule introducingprenominalAdjactive Phrases nor the one introducingsubject N2s introducesan underspecifiedcategory of the sort we have been discussing.Hence the featurescontained within A2[-PRD] 144 IVAN A. SAG ET AL. must be in any conjoined prenominalmodifier, and the feature of N2 must be in any conjoined subject. This accounts for facts like those in (59). Finite VPs provide another illustrationof the scope and power of the theory of coordinationwe have advanced. We admit tree fragments of the form shown in (70). (70)a. VP{[FIN],[-PAST1} VP{[FIN],[-PAST], [CONJNIL]} b. VP{[FIN],[-PAST], [CONJand]} VP [FIN] VP{[FIN],[-PAST], [CONJNIL} c. VP{[FIN],[+PAST], [CONJand} VPJ[FIN] VP{[FIN],[+PAST], [CONJNIL} d. VP{[FIN],[-PAST], [CONJand} VP{[FIN],[+PAST]} VP{[FIN],[+PAST], [CONJNIL1} VP{[FIN],[+PAST], [CONJand] Since the present framework treats partially specified categories in exactly the same was as fully specified categories, namelyjust as possible node labels in the structuraldescriptionof a sentence, there is nothing to stop any of the four tree fragmentsexhibited above from formingpart of a well-formed structuraldescription. In particular,the structuressketched in (70a, b, c, and d) will be responsiblefor admittingthe examples in (71a, b, c, and d) respectively. (71)a. b. c. d. Kim alienates cats and beats his dog. Kim alienates cats and beat his dog. Kim alienated cats and beats his dog. Kim alienatedcats and beat his dog. But none of the four structures allowed by our analysis will permit examples like (72).16 16 Nor can any of these examples be produced by an instantiationof diverse VFORM values on daughterswith no VFORM value at all on the mother. Every rule of grammar introducinga complementV2 introducessome VFORM value, which hence would have to be contained in all conjunctsof a coordinationof such a complement.And wheneverS is introduced,some VFORM value is specified,either by a particularrule, or in the case of FIN, in the list of categories that can stand as independent utterances in discourse. Structuresderived from a coordinationschema which lacked VFORM in the V2 mother could never be utilizedin the syntacticanalysisof any usable sentence of the language. COORDINATION (72)a. b. c. d. e. f. AND CATEGORIES 145 *Kim alienatedcats and beating his dog. *Kim alienatedcats and to beat his dog. *Kim alienatedcats and beaten his dog. *Kim beating his dog and alienates cats. *Kim to beat his dog and alienatedcats. *Kim beaten his dog and alienates cats. It is perhaps not obvious that the principles we have outlined are successful in dealing with agreement between subject NPs and each of a number of coordinate VPs. We comment on this only in passing. A subject NP must be fully specifiedfor HEAD features and for agreement features. This follows from the assumptionthat lexical entries are fully specifiedand from the definitionwe have given of the HFC. Hence a VP with which a subject NP combines is marked (AGR, NP[a]), where NP[a] includes all the agreement features of the subject NP, by the Control Agreement Principle discussed in GKPS, chapter 5, section 3. Thus if such a VP is coordinated, it follows that each conjunct also contains (AGR, NP[a]) (since AGR is a HEAD feature). This in turn guaranteesthat the V? withineach conjunct also contains(AGR, NP[a]). Assuming finite verb forms are the result of a productive morphological component which creates verbal forms assigned to the appropriatecategories whose definitioninvolves the feature AGR, this has the effect of ensuring agreement of the desired sort between a subject NP and each verb in a coordinate VP (or V1, or V?) structure. 3.3. Coordination and 'Extraction' We turn now to the interactionof our syntax for coordinationand the analysis of unbounded dependency phenomena. As we shall show, the theory of coordinationthat we have outlined capturesRoss's Coordinate Structure Constraint and 'Across-the-Board' facts automatically. Nothing special has to be said about them. Before this can be demonstrated, however, we must recapitulate essential details of the GPSG analysisof unboundeddependencies. The category-valued feature SLASH signals the presence of a gap within a given constituent.Thus a VP containing an NP-gap belongs to the category VP[SLASHNP], an S containing a PP gap belongs to the category S[SLASHPP], and so forth. Various rules which we will not discuss in detail here introduce 'slashed' constituents and cause gaps to be realized at appropriateplaces within them in the mannerdescribedin GKPS and earlier work. For example we assume rules providing for structures like (73), and these provide the basis for the analysis of structureslike (74). 146 A. SAG ET AL. IVAN (73)a. S{[+INV], S{[+INV], [SLASH NP]} NP [WH NP] b. [WH NP]} VP [SLASH NP] NP{[+NULL], [SLASH NP]17 V" (74) S[WH NP] {[+INV], [SLASH NP] NP who g did NP VP [SLASH NP] Kim V( NP{[+NULL], [SLASH NP]} see What principles govern the inheritance of SLASH in feature instantiation? It is our view that a single general principle governs the inheritanceof all FOOT features: the FOOT FEATURE PRINCIPLE given informallyin (75). (75) Foot Feature Principle (FFP): The FOOT features instantiated on the mother in a tree fragmentare identical to the unificationof the FOOT features instantiatedon the daughters. In additionto providingfor the SLASH instantiationillustratedin a tree such as (74), the FFP also allows the featuresWH and RE to be inherited in such a way as to provide an analysisof the feature passing requiredby such examples as those in (76). 17 The rules responsiblefor structureslike (73b) are not basic ID rules; rather they are derived by a metarulesuch as that given in Sag (1982). Note that an FCR prevents the HFC copying SLASH onto the lexical head in (73b) and (74) - see GKPS, chapter 2, section 3, for the formulationof such an FCR. Structure(73a) is sanctionedby the same rule utilizedin the analysisof topicalizedsentences. COORDINATION (76)a. b. c. d. AND CATEGORIES 147 A student whose teacher was out of town ... Which student'sgrades went unreported? They found pictures of themselves. They knew that picturesof each other would be on sale. For more detailed discussionof this last point, see Pollardand Sag (1983) and GKPS, chapters6 and 7. Note that because SLASH is a HEAD feature, as well as a FOOT feature, SLASH-specificationsare always to be found on the phrasal head of a mother category instantiated with a SLASH specification, though not on a lexical head, thanks to the FCR mentioned above. The case for treating SLASH as a HEAD feature was first developed by Flickinger(1983) who showed that it provided an explanationfor a wide range of island phenomena.Subsequently,Sells (1983) was able to show that the same requirementon SLASH was able to explain a numberof otherwise puzzling minimal pairs involving parasitic gaps. From the interaction of the HFC and the FFP it follows that SLASH can be instantiatedon the sister of a nonlexicalhead just in case it has also been instantiated(with identical value) on that head. This fact provides an immediateaccount of the data in (77) noted by Engdahl (1983). (77)a. b. c. d. *Who did you say [[my talking to-] [would bother Hilary]]? Who did you say [[my talking to -] [would bother ]]? *Whicharticle did Terry [[file papers][withoutreading_]]? Which article did Terry [[file -] [withoutreading_]]? This inclusion of SLASH in HEAD is thus motivated by a range of considerationshaving nothing to do with coordination. However, an importantconsequence follows from the dual membership of SLASH in HEAD and FOOT: all categories in a coordinate structure have identical SLASH specifications. Consider the following informal demonstration of this proposition. Suppose we instantiate (SLASH, NP{[+XSP], [+THP], [+SING]}) onto any phrasal conjunct. The mother of the coordinate structure must also contain the specification (SLASH, NP{[+XSP], [+THP], [+SING]}), in accordance with the FFP. But every conjunct is a head of the mother and hence, by the HFC, must include all HEAD feature specificationsof the mother. Thus any instantiated specification of a FOOT feature that is also a HEAD feature must be instantiatedon all conjuncts. This consequence, taken together with the independentlymotivated requirementthat the feature NULL be default absent, provides a correct account of the deviance of the following examplesdiscussed by Sag (1982) and Gazdar et al. (1982): 148 IVAN (78)a. b. c. d. A. SAG ET AL. Which books did Robin read-and hate-? *Whichbooks did Robin talk to Chris and read-? *Whichbooks did Robin read-and talk to Chris? *Who did Robin visitL and-? And finally,since the features RE and WH (which mark the presence of reflexive and reciprocal pronouns, and of interrogative and relative elements, respectively) are FOOT features, but not HEAD features, conjuncts need not agree on specifications for these features. This correctly accounts for the possibilityof exampleslike those in (79): They talked to Kim and to each other. He hated himself and his friends. They were wary of themselves and (of) each other. ?They asked which student and Lee could get along together. They asked which students and which teachers would get along together. f. ?We called up every man whose father and Sandy had played on the same team. g. We called up every man whose father and whose mother had played on the team. (79)a. b. c. d. e. The examplesin (79a)-(79c) were pointed out independentlyby Elisabet Engdahl and by Paul Schachter as problematic for the analysis of coordinatestructuresdeveloped in Gazdar et al. (1982)18. We will conclude this section by briefly considering a recent transformationalaccount of coordination that purportsto explain the CSC and ATB facts we have been considering.Pesetsky (1982) introducesthe notion of PATH (essentiallya set of adjacent connected nodes in a tree) into grammaticaltheory, and allows various constructionsto induce such paths. He then proposes a principle which legislates against structures which contain distinct paths such that one path is not contained within the other. Coordinate structuresand unboundeddependency constructions both give rise to paths, and it turns out that CSC-violations are associatedwith pairs of paths that show overlap but not containment. However, Pesetsky provides no serious underlying theory of coordinationitself, and he is inexplicitabout how one is supposed to arriveat such constructions(pp. 439-440). For example, he gives no hint what18 See Engdahl (1983b) and Schachter(1983). The present account differs from that of Gazdar et al. (1982) in two respects: (i) it deals with a wider range of data, and (ii) it eliminatesthe need for any principlesthat specificallymention coordinatestructures:the ConjunctRealizationPrinciplehas been eliminatedin favor of the HFC. COORDINATION AND CATEGORIES 149 soever as to how the very basic facts illustrated in (53) through (55) above might be captured in the grammarof English. He does assume that coordinateconstructionsare multiplyheaded, but this claim, instead of playing a crucial explanatoryrole as it does in the theory outlined here, appearsto be independentof his other claims. Thus it plays no part in his descriptionof the ATB/CSC facts, nor even in the transmissionof informationabout tense between a sentence and its main verb (p. 462), and he is forced to specify that conjunctshave the same bar level as their mother, althoughhe claims that "nothingimportantseems to follow from this stipulation"(p. 440). More seriously, his analysis of the ATB/CSC facts depends on two furtherstipulationswhich make crucial reference to coordinate structure itself. In the first of these, conjunctions are required to theta-mark conjuncts in order to induce the uncontained paths needed by the analysis (p. 440). No independent motivation is provided for this new kind of theta-marking. In the second coordinate-specific stipulation, 'sisterhood'is redefinedso that the aunt of a conjunct counts as a sister for all grammaticalpurposes(p. 569). This surprisingmove calls, in turn, for a special 'relativized' version of the theta-criterion, and a novel definitionof 'subject'.` Even with all these additional devices, Pesetsky's analysis, as he himself admits, is unable to account for CSC violations involving the coordinationof non-maximalprojections (p. 566), and he has to "retain the A/A conduction to prevent extraction of an entire conjunct" (p. 443). 4. FURTHER ISSUES In this section we discuss four furtherissues in the theory of coordination which are not covered by the analysispresentedso far. These are issues which, to the best of our knowledge, present problemsfor every existing theory of coordination.20Our proposalsin respect of the last two problems we consider are somewhatspeculative. 4.1. Asymmetric Conjunction Ross (1967, pp. 93-94) pointed out that not every English sentence '9 Essentially: "sister of tensed VP". This definition, unlike the one it replaces ("daughter of tensed S"), entails a rather abstract analysis of VSO structures. 20) For another issue of this kind, one not considered here, see Browne (1972) and Grosu (in press). 150 IVAN A. SAG ET AL. containing and instances a coordinate structure.Three of his examples are shown in (80), below. (80)a. b. c. I went to the store and bought some whiskey. She's gone and ruined her dress now. I've got to try and find that screw. We deal with such cases by means of the ID rules in (81). (81)a. VP-* Ho, (PP),H[CONJ and]. b. VP-- H[BSE], H[CONJ and]. At least come and go belong to the class of verbs admittedby (81a) and at least try belongs to the class admittedby (81b). There are two heads in these rules, hence both will be realizedas [+V] and [-N] in accordance with the HFC. The second heads will also be realized as {[-SUBJ], [BAR 2]} since these HEAD features appear on the mothersand are not inconsistentwith the features specified on these heads in the rule. Note that [BAR 2] will not be forced onto the lexical heads by the HFC since this would result in the latter no longer being a category (BAR would have two values). In instantiationsof (81b), the VFORM value BSE will be forced onto the mother (by the HFC) and, hence, onto the VP complement (also thanks to the HFC). And in instantiationsof (81a), the HFC will ensure that the mother, the lexical head, and the VP complement all share the same value for VFORM. Thus the two rules we have given predict the grammaticalitydistribution shown in (82) and (83). (82)a. b. c. d. e. f. She goes and buys some whiskey. I have gone and bought some whiskey. Going and buying whiskey is not the solution to your problems. I will go and buy some whiskey. I will try and buy some whiskey. I want to try and buy some whiskey. (83)a. *She goes and buying some whiskey. b. *1 have gone and buys some whiskey. c. *To go and buying whiskey is not the solution to your problems. d. *1 will go and bought some whiskey. e. *I tried and buy some whiskey. f. *1 was trying and buying some whiskey. COORDINATION AND CATEGORIES 151 Our present analysis(in contrastto that proposedin Gazdaret al., (1982) does not provide a syntactic account of the unacceptability of the following examples: (84)a. *Whatdid you say I went and get? b. *Whatdid you say I go and got? However, the HFC will not allow the structures that legitimate these strings to have a tense specification of the mother VP (since the intersection of {[+PAST]}with {[-PAST]} is the empty set). If tense has to be interpretedsemanticallyat VP or S nodes, rather than at V nodes, as scope facts might lead one to propose, then there is an independently motivated semantic explanationfor the anomalyevident in these examples. The rules we have given lead us to expect that the semantics of sentences induced by them will not turn out to be identical to the semantics of sentences containing genuine coordinate phrases. Any standard cross-categorial semantics for the latter will predict that, modulo quantified subjects, sentences involving coordinate VPs are synonymouswith correspondingsentences containingcoordinateSs. But, as Schmerling (1975, p. 220) has pointed out, this synonymy does not manifest itself when we consider the putative sentential counterpartsof the examples in (80). (85)a. I went to the store and I bought some whiskey. b. ?She's gone and she's ruined her dress now. c. I've got to try and I've got to find that screw. Schmerling notes that none of these sentences "can be understood as equivalent to the 'reduced' versions given earlier .., the 'extra' informationconveyed in the 'reduced'versions in simplynot present here" (1975, pp. 220-221). Recall that CS2 permits an initial conjunct in a genuine binary coordinate constructionto be markedwith both. But the rules in (81) do not make provision for the lexical heads to be markedwith both. Thus our analysisleads us to expect that inserting bothin the examples in (80) will force them to be interpreted as genuine VP-coordinations, rather instances of the verb + complements structure that the rules in (81) deal with. As Schmerlinghas demonstrated,this indeed turns out to be the case. Her examples (1975, p. 222) are shown in (86). (86)a. I both went to the store and bought some whiskey. b. ?She's both gone and ruined her dress now. c. I've got to both try and find that screw. 152 IVAN A. SAG ET AL. She points out that these examples are interpreted as paraphrasesof those in (85), not those in (80). Both rules in (81) may undergo feature instantiationin such a fashion that [SLASH NP] is instantiatedon the VP[CONJ and] daughterand on the VP mother (in accordance with the HFC and FFP). Note that the presence of the SLASH specification on the head VP complement satisfies the requirements of the FFP: the FFP does not require its presence on any of the other daughters. The HFC likewise does not require (or indeed permit) the presence of the SLASH on the lexical head, in view of the FCR alluded to above. Thus a grammarwhich includes the rules shown in (81) will automaticallygenerate the examples shown in (87) (from Ross 1967, pp. 93-94) without any extra provision having to be made. (87)a. Here's the whiskey which I went to the store and bought. b. Which dress has she gone and ruined now? c. The screw which I've got to try and find holds the frammisto the myolator. Bever et al. (1975) argued that examples such as those in (80) and (82) entailed a choice between (i) a position in which one claimed that they were grammatical and invoked global and transderivationalrules to handle them, or (ii) a position in which one claimed that they were ungrammatical,but acceptable for processing reasons. They opted for position (ii). However, as we have shown, the present analysis is not impaledon either of the horns of their dilemma. 4.2. Noun Phrase Coordination We have so far said nothing about the person and numberof coordinated NPs, a long-standingsyntactic problem that to our knowledge has never received a precise formal treatment.Partof the difficultyin providingan account of NP coordinationis that it is somewhat unclear what generalizations are to be expressed. Our analysispresupposesa set of generalizationsthat may not be valid for all varietiesof English, but which seem to us to closely approximatethe facts of many speakers whose judgements we have consulted. Let us begin with the observation that the 'person' of NPs coordinated with either and or or is as indicatedin (88). COORDINATION (88) NPa 1st 1st 1st 2nd 2nd 3rd AND CATEGORIES NPb NPa CONJ NPb 1st 2nd 3rd 2nd 3rd 3rd 1st 1st 1st 2nd 2nd 3rd 153 This observationis illustratedby the sentences in (89), where we rely on the overtly expressed person of a direct object reflexive pronoun to indicate the person of the coordinatesubject NP. (89)a. b. c. d. e. f. g. h. Either we Americansor I myself will get ourselves in trouble. Either you or I will perjureourselves. You and I may perjureourselves. We Americansand the Britishpamperourselves. You Britishand you Americanspamperyourselves. You Britishor you Americanswill get yourselvesin trouble. You and Kerry have outdone yourselves. You or Kerry have perjuredyourselves. The facts of (88) are symmetric:permutingthe coordinatedNPs in the examples of (89) does not alter the person of the coordinateNP subject, as the readermay verify.2'The generalizationwe may extractfrom these data then, if we make the traditionalassumptionsthat the proper categories to distinguish within NPs are 1st person, 2nd person and 3rd person, is that the personof a coordinateNP structureis the minimumof the persons of the conjuncts. If, however, we make certain different assumptionsabout the categories involved, we can explain the observed generalization.As noted in section 2.11, we analyze person in terms of the features XSP (for 'eXcluding the SPeaker') and THP (for 'THird Person'). These features have only '+' as their value and are hence unspecifiedin certain instances, rather than being negatively specified. These assumptions allow us to replace the traditionalperson categories with the following feature specifications: (90)a. b. c. 21 1st Person:{ }. 2nd Person:{[+XSPI}. 3rd Person:{[+THP], [+XSP]}. Althoughother factors may interfereto make such permutationsless acceptable. 154 IVAN A. SAG ET AL. The HFC, as formulatedearlier, correctly predicts the person of coordinate NPs without further stipulation. XSP and THP are both HEAD features, and hence the person feature specificationsof coordinate NPs must be the intersectionof the person feature specificationsinstantiated onto the NP conjuncts, all of which are heads. In the present case, this is equivalent to a requirementthat the set of person feature specifications of a coordinateNP is the intersectionof the person feature specifications of the conjuncts. The HFC thus correctly predicts the facts of (88), as illustratedin (91).22 (91) NPa NPt NPaCONJ NPh {} {} {} {} {[+XSP] {} {} {[+XSP] {} {[+XSP]} {[+XSP], {[+XSP] {[+XSP],[+THP} {[+XSP] {[+XSP], [+THP]} {[+XSP], [+THP} {[+XSP], [+THP]} [+THP]} {[+XSPJ The basic observations to be made about the number of coordinate NPs appear to be the following: (1) the coordinationof two (or more) plural NPs is always plural; (2) the coordination of a plural NP and a singular NP is always plural; (3) the coordination of two singular NPs with or is either singular or plural; and (4) the coordination of two singular NPs with and is always plural.23These observations are illustrated in (92). (92)a. The boys and the girls seem/*seems happy. b. Either the boys or the girls are/*is going to be there. c. The students and Professor Swansong are/*is meeting in the park. d. Either Professor Swansong or the graduate students are/*is going to invigilate the exam. 22 Our analysisis thus in the spiritof Farkasand Ojeda (in press),but makes no use of any ancillaryfunction, as theirs does, to predict the agreement features of coordinate NPs. Karttunen(1984, p. 32) finds our feature system "counterintuitive"in that it assigns 3rd person categories the most feature specifications,whereas, traditionally,3rd person is regarded as the unmarkedform. We are unclear as to the status of such intuitionsand suspect that the notion 'marked'is an equivocal one. Karttunen'sown analysis,which is closely relatedto that presentedhere, employs additionalfeaturalmachinerynot available withinthe versionof GPSG assumedin the presentpaper. 23 We assumethat whatevermechanismis responsiblefor the acceptabilityof exampleslike (i) is a mechanismthat permitsmorphologicallypluralNPs to appearas singularNPs, as illustratedby the acceptabilityof (ii). (i) (ii) Ham and eggs is my favorite breakfast. Flapjacksis my favorite breakfast. Hence there is nothing idiosyncraticabout coordinationto be consideredin the face of exampleslike (i). Cf. Hoeksema 1983, pp. 71-72. COORDINATION e. f. 155 AND CATEGORIES Either Dana or Lee is/are going to lead the parade. Kim and Terry are/*is happy. Our analysis of NP number makes use of the feature SING which, like the person features, has only the single value '+'. Note that if nothing more is said, the HFC predicts that the number of a coordinate NP will be as indicated in (93). (93) NPa {[+SING]} NPb NPaCONJ NPb {} {} {[+SING]} {[+SING]} {} {} {[+SING]} {[+SING]} This is the correct result for all cases, except when both NPa and NPb are singular, where we obtain different results, depending on the choice of conjunction, as we have noted. We can account for the remaining data if we introduce a minor modification to the analysis we have put forth. We modify the rule we presented in (49b) to expand elements specified for CONJ in the fashion shown in (94). (94) X[CONJ a]-- a, H([+SING]), where a E {and, both, but, either,neither,nor, or}. This rule now optionally introduces the specification [+SING] on the head daughter. To ensure that the coordination of two singular NPs with and is always plural, we need only add an FCR stating that an NP[CONJ and] must be unspecified for [+SING], i.e., must be plural. This has the effect of ruling out any structures resulting from (94) which involve instantiation of [+SING] onto mother and head when the mother is an NP and a is and. However, (94) will permit the head daughter to be [+SING] and this will be tolerated by the HFC even though the NP mother is prevented from being [+SING] by the FCR just mentioned. The very general 'default-like' character of the HFC can thus sometimes permit a plural mother to immediately dominate a singular head, just as it can sometimes permit a [BAR 2] or [BAR 1] mother to dominate a lexical category. The observations made above are now all correctly predicted by the HFC.24 24 The English NP coordination facts discussed in this section are but the tip of an interesting iceberg. For enlightening discussion of related facts in a range of other languages,see Corbett(1983) and Schwartz(1984). 156 IVAN A. SAG ET AL. 4.3. Non-constituent Coordination The analysis developed in section 3 deals with conditions on the coordinationof constituents.It is well-known, however, that English exhibits constructionswhich appearto involve the coordinationof stringsthat are not constituents.Typical cases of non-constituentcoordination,familiar from the recent literature,are illustratedin (95)-(98): Kim likes Sandy,and Lee Leslie. (95) to try to go to Rome. Pat wanted to try to go to Berne, and Chris to go to Rome. to Rome. Kim gave a dollar to Bobbie and a dime to Jean. Kim went to the store, and (then) Lou. (96) (97) (98) These examples have been analyzed in terms of such transformationsas Gapping ((95) and (96)), Conjunction Reduction ((97)), and Stripping ((98)). Are these examples to be analyzed in terms of distinct grammatical processes as the transformationalliteraturehas implied?In our view, the best theory of such examples would be one which unified their account, treating each kind of example as a variation of a single general phenomenon.After arguingfor this view, we will offer a tentative sketch of such a unifiedanalysis.25 Hudson (1982) argues against any unifiedtreatmentof these examples. Hudson suggests that there is an isolable phenomenonof Gappingwhich must be given an account unrelated to that of examples (97) and (98). We find his argumentsfor this a priori undesirableconclusion unconvincing.26 Hudson claims, for example, that Gapping, but not Conjunction Reduction (which he takes to be a distinct grammaticalprocess involved in the derivation of (95) but not (97) or (98)), is blocked if the conjunction in question is but. Although the observation is often made that examples like (99) are deviant, (99) ?Jean likes Lou, but Rene, Dominique. the explanationfor this deviance need not be grammaticalin nature, as many examples of Gapping with but are contextualizable,as noted by Neijt (1979).27 Cf. Schachterand Mordechai(1983). Russell (1983) has independentlyprovideda detailed rebuttalof Hudson'sarguments. 27 Perhapsthe reason that exampleslike (i) (cited by Hudson) are easier to contextualize has to do with the usage conditionsassociatedwith but. 25 26 COORDINATION AND CATEGORIES 157 (100)a. Some people go by car, but others by bike. b. Some people like bagels, but others cream cheese. [Neijt 1979, p. 59] (101)a. Fourteen-year-oldslike punk and new wave, but sixteen-yearolds, only punk. b. On weekdays, Terry eats meat and vegetables, but on weekends, only vegetables. Hudson also argues that Gapping is distinctfrom ConjunctionReduction in that the former,but not the latter, is restrictedto allow only two constituents in the elliptical conjunct. In support of this claim, he cites contrastslike the following: (102) [ = Hudson's (5a), (7)] a. *Johndrinkscoffee at 11, and Mary,tea at 10:30. b. John gave the books to Maryat Christmas,and the records to Sue for her birthday. But this claim (advanced earlier by Stillings (1975)) is known to be incorrect. Examples like the following are noted in Sag (1976), Ross (1976), and elsewhere: (103)a. Peter talked to his boss on Tuesday, and Betsy to her supervisor on Wednesday. b. John talked to his supervisorabout his thesis, and Erich to the dean about departmentpolitics. It appearsthat whateverdeviance inheresin exampleslike (102a) is more properly explained by appealing to the processing difficultyassociated with sequences of NPs found in ellipsis contexts. Note that single-word NPs appear to be particularlyconfusing. Acceptability increases when the NPs are more contentful, as in (104). (104) A businessmanwill drink a martinito relax, and a health nut, a glass of wine, just to remain healthy. On the basis of these facts, we can conclude that the grammarof both (i) John gave the books to Mary, but the records to Sue. But-coordination requires that the second conjunct convey a message that contrasts in some fashion with that conveyed by the first conjunct. Yet the two messages must share a common topic, and subjects are the most unmarked indicator of topic. Hence the appropriateness conditions for 'Gapped' examples like (100), where the subject is retained within the second conjunct are complex and highly marked. The presence of only in the examples in (101), however, eases the burden of constructing an appropriate context. 158 IVAN A. SAG ET AL. the 'Gapping' and 'Conjunction Reduction' constructionsshould allow multipleellipsis 'remnants'. A third argumentmade by Hudson is based on contrastslike (105). (105)a. ?*Johnleft at 11 and at 12, Bill. b. John left his office at 11 and at 12, the library. Gapping, but not ConjunctionReduction, it is claimed, requiresthat the order of elements in the second conjunct parallel the order of the correspondingelements in the first conjunct. This contrastseems dubiousto us; (106), for example, appearsto have about the same degree of acceptabilityas (105b): (106) A policemanwalked in at 11, and at 12, a fireman. These examples,which we returnto brieflyin a moment, are counterexamples to all analysesof coordinateellipsis that we are familiarwith. Hudson also contends that the first remnant in a gapped clause must always be a subject. However, as Kuno (1976 p. 307, n.11) notes, this contention is false. Exampleslike the following illustratethis point: (107)a. Two days ago, we went out to dinner, and this afternoon, to the movies. b. On this table, they put a lamp, and on that table, a radio. Here again, there is no reason to assume that 'Gapping' and 'Conjunction Reduction' are distinct grammaticalconstructions subject to differentsyntactic restrictions. Finally, Hudson notes the contrastin (108): (108)a. *Johndidn't see Maryand Bill Sue. b. John didn't give the books to Mary and the papers to Sue. The fact that the 'Gapping'constructiondoes not occur with and if the first conjunct contains a negated main verb was noted first by Ross (1967). Below we will suggest that (108a) is a case of sentence-level ellipsis, and (108b) is a case of VP-level ellipsis. We conjecture that this distinctionwill play a central role in the analysisof such contrasts. In sum, we have examined Hudson's arguments for the a priori undesirableposition of distinguishingmultiple grammaticalprocesses in the analysisof ellipsis phenomena, and found them to be unconvincing. We now sketch the outlines of a unifiedaccount of the various instances of coordinateellipsis illustratedearlier in (95)-(98) [repeated here]. (95) Kim likes Sandy, and Lee Leslie. COORDINATION AND CATEGORIES (96) Pat wanted to try to go to Berne, to try to go to Rome. and Chris to go to Rome. to Rome. (97) Kim gave a dollar to Bobbie and a dime to Jean. (98) Kim went to the store, and (then) Lou. 159 In each of these examples there is a non-initial conjoined element that contains a sequence of phrases which is not generated freely as such elsewhere in the grammar. Moreover, there is a dependency between the form of these phrases and the presence of appropriate elements in the relevant positions in the preceding conjoined clause. Consider the examples in (109): (109)a. *Kim likes Sandy, and Lee to Leslie. b. *Pat wanted to go to Berne, and Chris going to Rome. c. *Kim gave a dollar to Bobbie and a dime into his pocket. d. *Kim likes Lee, and to Ronnie. Previous analyses of this data have by and large dealt with this discrepancy indirectly. The Gapping transformation, for example, was formulated roughly as in (110) [adapted from Sag 1976, Chapter 3]: (110) SD: XPI-WI-XP2W2-CXP3-W3XP4- W4 1 2 4 3 6 7 5 8 9 SC: 1, 2, 3, 4, 5, 6, 0, 8, 0 conditions: 2 = 7, 4 = 9, and C E {and, or, but}. The pre-deletion structures had to satisfy the identity conditions in (110); hence an example like (109a) would have been derived from the same structure as that underlying (111): (111) *Kim likes Sandy and Lee likes to Leslie. The deviance of this structure was predicted by subcategorization mechanisms, in consequence of which (109a) was also blocked. In purely interpretive analyses such as that developed by Stump (1978), no account is provided of the dependency just illustrated. Presumably, an adherent of this approach is committed to a theory wherein all subcategorization dependencies preserved under Gapping are reanalyzed as purely semantic dependencies. Alternatively, one might supplement the rule which assigns interpretations to Gapped structures with something much like the structural description in (110). This con- 160 IVAN A. SAG ET AL. dition would require that each Gapping remnant match in syntactic category some element within the preceding conjoined clause. But such a requirement would be too strong. As the examples in (1 12) show, the predicative Gapped 'remnants' need not agree totally with the syntactic features of the corresponding element in the left conjunct. (112)a. Leslie is rather foolish, and Lou a complete idiot. b. Kim seems to be just surviving, and Terry in dire need of our help. c. We consider Leslie rather foolish, and Lou a complete idiot. Rather, it seems that Gapped structures, as well as the related structures illustrated in (95)-(98) that have been treated as instances of unrelated phenomena, all obey a substitutional generalization. If the result of substituting the remnant for the corresponding element in the preceding conjunct is well-formed, then the Gapping structure is well-formed as well. It seems that any 'surface-based' account of this phenomenon must involve this notion of substitution. Further support for a substitutional treatment is provided by examples like (113): ( 13)a. Pat has become crazy, and Chris depressed. b. Pat has become crazy, and Chris an incredible bore. c. *Pat has become crazy, and Chris in good spirits. Such contrasts further illustrate the point that syntactic subcategorization dependencies hold into Gapped constituents, and provide further evidence that purely semantic analyses like Stump's are untenable. The idea that the Gapping phenomenon should be handled by a sentence grammar mechanism, such as a transformational rule, is curious, As Hankamer and Sag (1976) show, Gapping can occur across speakers in discourse, as in ( 14). (I 14) SPEAKER A: I shall miss you. SPEAKER B: And I you. Note that the shift in (first person singular) deixis in such examples renders implausible the suggestion that it should be regarded as a peculiar kind of two-speaker collaboration on a single sentence. Gapping is a discourse anaphoric process, not a rule of sentence grammar, despite claims to the contrary made by Williams (1977) and Neijt (1979). The rules of sentence grammar thus have only to generate structures consisting of a conjunction followed by any number of X2 phrases, and a rule of discourse must be employed to predict the interpretation of the COORDINATION AND CATEGORIES 161 sequence of phrases from the preceding, directly juxtaposed linguistic context. This rule will presumably need to make reference to the fact that the phrases are focused constituents, typically bearing contrastive accent. Notice, in this connection, the unacceptability of unstressed pronouns in examples like those under discussion. (1 15)a. *You talked to John's mother, and I him. [him unstressed] b. *I gave a book to John's mother and a magazine to him. [him unstressed] In the absence of any formal theory of discourse rules and focus, it is impossible to provide any explicit analysis of the phenomena under discussion. It is, moreover, clearly beyond the scope of this paper to attempt to develop such a theory. Instead, we will informally offer a very tentative proposal, without committing ourselves on a number of important details. First, we will need an ID rule to generate the sequence of phrases. Something along the lines of ( 116) will do this, though presumably additional features will have to be stipulated to indicate that this is an elliptical construction whose constituents must be focussed. (116) V2[CONJ a]-- a, X2+ where a E {and, but, nor, or}. Note that the category V2 includes both sentences and VPs, which are distinguished in virtue of the feature SUBJ, as noted earlier. (116) says that a coordinate elliptical sentence or verb phrase can consist of any number of phrasal constituents. The rule in (116) will give rise to structures like those in (117) and (118): S[CONJ and] (117) and NP NP Lee Terry 162 IVAN A. SAG ET AL. VP [CONJ or] (118) or NP PP a book to Kerry These structures must be assigned interpretations by substituting the phrases for the appropriate corresponding elements in a preceding directly juxtaposed structure.28 Thus if the result of substituting a remnant in for a constituent within a prior structure is a structure not analyzable by the grammar, i.e., not independently generable, then no interpretation is obtained for the elliptical V2. The semantic interpretation for structures admitted by ( 16) may be given by the rule informally stated in (119): (119) The interpretation of an elliptical construction is obtained by uniformly substituting its immediate constituents into some immediately preceding structure, and computing the interpretation of the results.29 This rule may be applied to a conjunct of a coordinate structure (though as we have seen, this need not be the case), in which case the immediately preceding structure used to compute its interpretation must be a preceding sister conjunct. Thus the semantic interpretation for a structure like (120) is obtained by interpreting the result of substituting the circled elements for the boxed elements in the fashion illustrated. We simplifythe discussionhere by ignoringellipsis when more than two conjunctsare involved. The proposal made here is not unlike the idea of assigning interpretationsto quantifiedsentencesof predicatelogic by consideringvarious'substitutioninstances'of the formulasto which the quantifiersare prefixed. 29 One of the refereespointedout an interestingdifferencebetween an approachlike ours, which interpretsthe resultsof substituting,versus an approachlike that of Stump(1978), which substitutesinterpretations.Considera sentence like (i): 2x (i) Pat is looking for a piece of paper,and Chris, a pencil. On Stump's analysis, if a piece of paper is interpreted de re, then a pencil must be interpretedde re as well. Our approachcorrectlypermitsone object to be interpretedde re and the other de dicto. COORDINATION (120) 163 AND CATEGORIES S /O \ Ta Terry e S (121) NP gave VP NP PP the book to Lee te records to Kim The interpretations that result are those shown in (122): (122)a and'(like'(Stacy*)(Terry*), likeS(Lee*)(Tracy*)). b. [andl(give'(Lee*)(the'(bookI)), giveF(Kim*)(the'(record')))] (Sandy*). 164 IVAN A. SAG ET AL. Note that nothing said so far requires that the phrases occur in the same order as their corresponding elements within the leftmost constituent. A further condition would have to be imposed to block the examples discussed earlier [see (105) and (106)] which Hudson argues should be regarded as ungrammatical. Additional constraints should perhaps be placed on the substitution operation formulated in (119) to avoid unwanted interpretations. It seems, however, that the majority of constraints on the interpretation of elliptical constructions should be regarded as extra-syntactic in nature, as suggested by Hankamer (1973). Kuno (1976) and Sag (1976). Processing strategies in all likelihood play a primary role in the explanation of the numerous preferences for certain interpretations. We are thus proposing to unify the account of Gapping phenomena (where two or more phrases are involved) and coordinate structure instances of Stripping phenomena (where there is only one phrase), thus accounting for their shared properties [noted in Hankamer (1971)]. Our intended analysis would allow all the examples in (95)-(98) to be generated, but no interpretations would result for the examples in (109), as substitution of the Gapped phrases for any elements of the leftmost constituent would produce a structure that was not analyzable by rules of the grammar, and which would hence not be assigned any interpretation. Note further that the examples in ( 11) would be correctly permitted, as substitution of, for example, N2 for A2 would yield a structure that is analyzable by our rule for be in (50) above, and hence is interpretable. Our analysis would also allow more than two post-conjunction phrases, though examples of this sort quickly become unintelligible, as we have noted. We emphasize again that the treatment of verbal ellipsis sketched above is highly tentative and incomplete. This is unavoidable, given the rudimentary state of current understanding of the sorts of discourse factors which play such a central role in these phenomena. Moreover, many further factors may be involved in the explanation of the full range of judgements about sentences of the sort under discussion. Given the arguably extragrammatical character of such factors, however, the very general syntactic substitution operation we have suggested is a plausible candidate for the grammtatical part of that explanaition. 4.4. Embedded Clauses and NPs Many speakers permit NPs to be conjoined with embedded clauses in COORDINATION AND CATEGORIES 165 certain environments:3" (123)a. Pat remembered the appointment and that it was important to be on time. b. That Himmier appointed Heydrich and the implications thereof frightened many observers. An obvious idea is to handle these examples along the same lines as those in (2) and (3), that is by positing a set of features common to both N2s and that-clauses31, and assuming that the rules introducing the coordinate nodes in (123) mention only those features. However this suggestion would fail to explain the fact that, although embedded clauses can appear coordinated with NP objects of prepositions, they cannot appear as objects of prepositions alone: (124)a. We talked about Mr. Colson and that he had worked at the White House. b. You can depend on my assistant and that he will be on time. c. Pat was annoyed by the children's noise and that their parents did nothing to stop it. (125)a. *We talked about that he had worked at the White House. b. *You can depend on that he will be on time. c. *Pat was annoyed by that their parents did nothing to stop it. These examples are apparent counterexamples to our theory of coordination, since that theory predicts that each conjunct of a coordinate construction should be able to appear alone in place of the entire coordinate structure. Our solution to this dilemma presupposes the following rule:12 (126) N2[NFORM a]-- a, where a E {S[COMP that], S[COMP for]}. NFORM is a HEAD feature which distinguishes sentential NPs from other NPs.33 The default for the feature NFORM is essentially one which 3" The intuitionsof speakerswho uniformlyreject these examples could be handled by eliminatingNFORMfrom the HEAD features.This would have the effect of ensuringthat only the default value could be instantiatedon conjuncts. 31 Although for-to clauses seem to also allow coordination with NPs under certain circumstances,we will here confine our discussionto finite clauses. 32 See Weisler(1982) for a defense of a rule along these lines, pace Koster (1978). 33 This analysisis simplifiedin inessentialrespects. In Sag and Klein (1982) and GKPS, chapter6, section 2, the NFORManalysisis generalizedto providean analysisof 'dummy' pronounsas well. 166 IVAN A. SAG ET AL. requires NFORM to be unspecified, and this situation characterizes non-sentential NPs. We could formalize such a default as (127a), given the GKPS notation for representing defaults introduced in section 2.2, above. It would follow from (127a), that it is in general the case that sentential NPs occur only when sanctioned by a rule that specifically mentions them, and overrides the default assignment. But, as the data above suggests, this is not exactly what we want. Rather, we need a restriction on NFORM that is, in effect, suspended in conjuncts. Accordingly, the default we actually need is that shown in (127b). (1 27)a. FSD: - [NFORM]. b. FSD: [NFORM] v [CONJ]. Here, (127b) says that NFORM can only be (freely) instantiated on conjuncts. The following rule expands P': (128) P'-+H(', NP. Since this rule does not mention NFORM on the object NP it introduces, the object must be unspecified for NFORM. This guarantees that in general the objects of prepositions are 'ordinary' NPs, rather than sentential ones. But if a prepositional object is coordinated, then (I 27b) will permit a sentential NP to appear as a conjunct. As a result, coordinate prepositional objects may have the form of (129): NP (129) NP NP{NFORM S[that], [CONJ and]} And this will give rise to examples like those in (124), but not those in (125). There is an order asymmetry in this kind of coordination. Thus we find contrasts like these: (1330)a. We talked about the issues we had worked on as students and that our perspectives had changed over the years. b.*We talked about that our perspectives had changed over the years and the issues we had worked on as students. This asymmetry is not manifest in subject position, as evidenced by examples like ( 131 ): COORDINATION (131) AND CATEGORIES 167 That our prespectives had changed over the years and the issues we had worked on as students were the topics of discussion. These puzzling facts can be accounted for if we adopt the following LP rule: (132) [ACC] < [NFORM S]. The LP rule in (132) says that sentential NPs always follow their non-subject sisters.: Thus the peculiar distribution of data involving coordinated NPs and embedded clauses can be fairly straightforwardly described in the present framework. It remains without any analysis in all the other accounts of coordination that we are familiar with. 5. CONCLUSION In this paper we have presented a detailed treatment of key problems in the syntax of coordination in English which goes well beyond previous treatments in the breadth of its coverage. The separation of immediate dominance rules from linear precedence rules had played an essential role in our analysis. It is this aspect of Generalized Phrase Structure Grammar that allows the full range of conjunctions in English to be treated in a unified manner using a small 34 Notice that the LP rule in (132) maintains that the asymmetry between coordinated sentential and non-sentential NPs holds not only for prepositional objects, but in all non-nominative positions. Example (ii) appears to contradict this prediction. (i) (ii) I didn't remember until it was too late John's inability to get along with Pat, and that he had no background in logic. I didn't remember until it was too late that John had no background in logic, and his inability to get along with Pat. Although (ii) looks like a violation of (132), it in fact has another possible analysis, namely via the ellipsis mechanism discussed in section 4.3. Under this latter analysis, the sentential and non-sentential NPs do not form a constituent; hence it should be impossible to topicalize them together. This treatment would therefore predict a contrast between (iii) and (iv): (iii) (iv) John's inability to get along with Pat and that he had no background in logic, I didn't remember until it was too late. ?That John had no background in logic and his inability to get along with Pat, I didn't remember until it was too late. Readers may assess for themselves the accuracy of this prediction. 168 IVAN A. SAG ET AL. set of constructs. This same factoring of dominance and ordering information is what allows us to account for such problems as the peculiar properties of the coordination of embedded clauses and NPs, as we have shown. In addition, it is the interplay of various independently motivated principles in GPSG, such as the Head Feature Convention and the Foot Feature Principle, that enable one to derive, rather than stipulate, a solution to such long-standing problems as the facts commonly discussed in terms of the Coordinate Structure Constraint and the Across-theBoard Convention. Over twenty years ago, the syntax of coordination was a key topic in the discussions that led to the widespread acceptance of transformational grammar. It is curious, then, that even today no version of transformational grammar has succeeded in explaining, and often not even in describing, well-known and very basic facts about coordination (e.g., the fact that arbitrary tensed VPs can coordinate with each other). Moreover, the various instances of coordination of unlike categories, which we have provided an account of without appeal to any ancillary devices or ad hoc principles, have received no serious analysis within the transformational tradition. Of course, much remains to be done on the grammar of coordinate constructions. Among the problems we have addressed insufficiently or not at all are the precise formulation of the syntax and semantics of non-constituent ellipsis, the treatment of 'right node raising' constructions, and the semantic peculiarities of N'-coordination discussed by Bergmann (1982). Nevertheless, the present paper improves on earlier generative treatments of coordination by broadening the coverage while at the same time stipulating less. REFERENCES Abraham, R. C.: 1941, A Modem Grammar of Spoken Hausa, Crown Agents for the Colonies, London. Bear, John: 1981, Gaps as SyntacticFeatures,MA dissertation,Universityof Texas, Austin [also availablefrom IndianaUniversityLinguisticsClub]. Bever, Thomas G., J. M. Carroll, and R. Hurtig: 1975, 'Analogy, or ungrammatical sequences that are utterable and comprehensibleare the origins of new grammarin languageacquisitionand linguisticevolution'. in Thomas G. Bever, JerroldJ. Katz, and D. Terence Langendoen (eds.), An Integrated Theory of Linguistic Ability. New York, ThomasCrowell. Bergmann, Merrie: 1982, 'Cross-categorialsemantics for conjoined common nouns', Linguistics and Philosophy 5, 399-402. Borsley, Robert: 1983, 'A Welsh agreementprocess and the status of VP and S', in G. Gazdar,E. H. Klein, and G. K. Pullum(eds.), Order,Concordand Constituency,Foris Publications,Dordrecht,pp 57-74. COORDINATION AND CATEGORIES 169 Bresnan, Joan: 1973, 'Syntax of the comparative clause construction in English', Linguistic Inquiry 4(3), 275-344. Browne, E. Wayles: 1972, 'Conjoined question words and a limitation on English surface structures', Linguistic Inquiry 3, 223-226. Carlson, Greg: 1983, 'Marking constituents', in F. Heny and B. Richards (eds.), Linguistic Categories: Auxiliaries and Related Puzzles, Vol. 1, pp. 69-98. Chomsky, Noam: 1957, Syntactic Structures, Mouton, The Hague. : 1970, 'Remarks on nominalization', in R. Jacobs and P. Rosenbaum (eds.), Readings in English Transformational Grammar, pp. 184-221. Cooper, Robin: 1979, 'Model theory for a fragment of English', unpublished paper, University of Wisconsin at Madison. Corbett, Greville: 1983, 'Resolution rules: agreement in person, number, and gender', in Gerald Gazdar, Ewan Klein, and Geoffrey K. Pullum (eds.), Order, Concord and Constituency, Foris Publications, Dordrecht. Dixon, Robert: 1972, The Dyirbal Language of North Queensland, Cambridge University Press, Cambridge. Engdahl, Elisabet: 1983a, 'Parasitic gaps', Linguistics and Philosophy 6, 5-34. : 1983b, 'Some remarks on the treatment of unbounded dependencies and parasitic gaps in Gazdar, Klein, Pullum, and Sag, "Coordinate structure and unbounded dependencies",' unpublished manuscript, Linguistics Institute, UCLA. Falk, Yehuda: 1983, 'Constituency, word order, and phrase structure rules', Linguistic Analysis 11, 331-360. Farkas, Donka, Daniel P. Flickinger, Gerald Gazdar, William A. Ladusaw, Almerindo Ojeda, Jessie Pinkham, Geoffrey K. Pullum, and Peter Sells: 1983, 'Some revisions to the theory of features and feature instantiation', unpublished manuscript, Linguistics Institute, UCLA. Farkas, Donka and Almerindo Ojeda: in press, 'Agreement and Coordinate NPs', Linguistics. Flickinger, Daniel P.: 'Lexical heads and phrasal gaps', in Michael Barlow, Daniel Flickinger, and Michael T. Wescoat (eds.), Proceedings of West Coast Conference on Formal Linguistics, vol. 2, Stanford Linguistics Association, Department of Linguistics, Stanford University, pp. 89-101. Gazdar, Gerald: 1980, 'A cross-categorial semantics for coordination', Linguistics and Philosophy 3, 407-409. Gazdar, Gerald, Ewan Klein, Geoffrey Pullum, and Ivan Sag: 1982, 'Coordinate structure and unbounded dependencies', in M. Barlow, D. Flickinger, and I. A. Sag (eds.), Developments in Generalized Phrase Structure Grammar: Stanford Working Papers in Grammatical Theory, vol. 2, Indiana University Linguistics Club, Bloomington, pp. 38-68. : 1985, Generalized Phrase Structure Grammar, Harvard University Press, Cambridege, MA, and Blackwell, Oxford. Gazdar, Gerald and Geoffrey K. Pullum: 1981, 'Subcategorization, constituent order and the notion "head"', in M. Moortgat, H.v.d. Hulst and T. Hoekstra (eds.), The Scope of Lexical Rules, Foris Publications, Dordrecht, pp. 107-123. : 1982, 'Generalized phrase structure grammar: a theoretical synopsis', mimeo, Indiana University Linguistics Club, August 1982. Gazdar, Gerald; Geoffrey K. Pullum, and Ivan A. Sag: 1982, 'Auxiliaries and related phenomena in a restrictive theory of grammar', Language 58, 591-638. Gazdar, Gerald, Geoffrey K. Pullum, Ivan A. Sag, and Tom Wasow: 1982, 'Coordination and transformational grammar', Linguistic Inquiry 13, 663-676. Goodall, Grant: 1983, 'A three-dimensional analysis of coordination', in Amy Chukerman, Mitchell Marks, and John F. Richardson (eds.), Papers from the 19th Regional Meeting of the Chicago Linguistic Society, pp. 146-154. 170 IVAN A. SAG ET AL. Grosu, Alexander: in press, 'On acceptable violations of parallelism constraints', in R. Dirven (ed.), Functionalism in Linguistics, Benjamin, Berlin. Hankamer, Jorge: 1971, Constraints on Deletion in Syntax, Ph.D. dissertation, Yale University [published in book form by Garland Publishers, 1979]. :1973, 'Unacceptable Ambiguity', Linguistic Inquiry 4, 17-88. Hankamer, Jorge and Ivan Sag: 1976, 'Deep and surface anaphora', Linguistic Inquiry 7, 391-428. Hoeksema, Jack: 1983, 'Plurality and conjunction', in Alice ter Meulen (ed.), Studies in Model- Theoretic Semantics, Foris, Dordrecht. Hudson, Richard: 1982, 'Incomplete conjuncts', Linguistic Inquiry 13, 547-550. Kay, Martin: 1979, 'Functional grammar', in C. Chiarello et al. (eds.), Proceedings of the Fifth Annual Meeting of the Berkeley Linguistics Society, Berkeley Linguistics Society, Berkeley, pp. 142-158. Karttunen, Lauri: 1984, 'Features and values', Proceedings of COLING84, Association for Computational Linguistics, Stanford. Keenan, Edward and Leonard Faltz: 1978, Logical Types for Natural Language, UCLA Occasional Papers in Linguistics, no. 3, Department of Linguistics, UCLA. Koster, Jan: 1978, 'Why sentential subjects don't exist', in S. Keyser (ed.), Recent Transformational Studies in European Languages, Linguistic Inquiry Monograph no. 3. Cambridge, MIT Press. Kraft, Charles H. and Kraft, Marguerite G.: 1973, Introductory Hausa, University of California Press, Berkeley. Kuno, Susumo: 1976, 'Gapping: a functional analysis', Linguistic Inquiry 7, 300-3 18. Neijt, A.: 1979, Gapping: a Contribution to Sentence Grammar, Foris Publications, Dordrecht. Partee, Barbara, and Mats Rooth: 1983, 'Generalized conjunction and type ambiguity', in R. Bauerle et al. (eds.), Meaning, Use and Interpretation of Language, de Gruyter, Berlin. Pesetsky, David: 1982, Paths and Categories, unpublished Ph.D. dissertation, MIT. Peterson, Peter: 1981, 'Problems with constraints on coordination', Linguistic Analysis 8, 449-460. Pollard, Carl: 1982, 'Generalized grammar - toward the formalization of some concepts on the syntax-semantics frontier', unpublished manuscript, Stanford University. Pollard, Carl J. and Ivan A. Sag: 1983, 'Reflexives and reciprocals in English: an alternative to the binding theory', in Michael Barlow, Daniel Flickinger, and Michael T. Wescoat (eds.), Proceedings of the Second West Coast Conference on Formal Linguistics, Stanford Linguistics Association, Department of Linguistics, Stanford University, pp. 189-202. Rooth, Max and Barbara Partee: 1982, 'Conjunction, type ambiguity, and wide scope "or"', in Daniel Flickinger, Marlys Macken, and Nancy Wiegand (eds.), Proceedings of the First West Coast Conference on Formal Linguistics, Linguistics Department, Stanford University, pp. 353-362. Ross, John: 1967, Constraints on Variables in Syntax, unpublished Ph.D. dissertation, MIT. :1976, 'Clausematiness', in Edward Keenan (ed.), Formal Semantics of Natural Language, Cambridge University Press, Cambridge. Sag, Ivan: 1976, Deletion and Logical Form, Ph.D. dissertation, MIT [published in book form by Garland Publishers, 1980]. 1982, 'Coordination, extraction, and generalized phrase structure', Linguistic Inquiry 13, 329-336. Sag, Ivan and Ewan Klein: 1982, 'The syntax and semantics of English expletive pronoun constructions', in M. Barlow, D. Flickinger, and I. A. Sag (eds.), Developments in COORDINATION AND CATEGORIES 171 Generalized Phrase StructureGrammar: Stanford Working Papers in Grammatical Theory vol. 2, Indiana University Linguistics Club, Bloomington, pp. 92-136. Schachter, Paul: 1983. A Note on Syntactic Categories and Coordination'. NLLT 2, 269-281. Schachter, Paul and Susan Mordechai: 1983, 'A phrase structure account of "nonconstituent" conjunctions', in M. Barlow, D. Flickinger, and M. Wescoat (eds.). Proceedings of the Second West Coast Conference on Formal Linguistics. Department of Linguistics, Stanford. Schmerling, Susan: 1975, 'Asymmetric conjunction and rules of conversation', in Peter Cole and Jerry Morgan (eds.), Syntax and Semantics 9: Speech Acts, Academic Press, New York. Schwartz, Linda: 1984, 'Asymmetric feature distribution in pronominal coordinations', paper presented to the Workshop on Agreement, Stanford, October 1984. Sells, Peter: 1983, 'Thinking about foot features', unpublished manuscript, Linguistics Institute, UCLA. Stillings, Justine: 1975, 'The formulation of gapping in English as evidence for variable types in syntactic transformations', Linguistic Analysis 1. 247-273. Stump, Gregory: 1978, Interpretive gapping in Montague grammar, in D. Farkas et al. (eds), Papers From the 14th Regional Meeting of the Chicago Linguistic Society, Chicago Linguistics Society, Chicago, pp. 472-481. Weisler, Steven: 1982, 'Coordination and the syntax of that-clauses', in Alan Prince and Steven Weisler (eds.), U. Mass. Amherst Occasional Working Papers in Cognitive Science, University of Massachusetts, Amherst, pp. 113-134. Williams, Edwin: 1977, 'Discourse and logical form', Linguistic Inquiry 8(1). 101-139. :1978, 'Across the board rule application', Linguistic Inquiry 9, 31-43. :1981, 'Transformationless grammar', Linguistic Inquiry 12, 645-654. Received 17 January 1984 Revised 5 November 1984 Dept. of Linguistics Stanford University Stanford, CA 94305 U.S.A.
© Copyright 2024