Sign-Based Construction Grammarlingo.stanford.edu/sag/SBCG/sbcg.pdf · Sign-Based Construction Grammar Charles Fillmore U.C. Berkeley ... 1.3.5 CONTEXT 12 ... that feature structures

Sign-Based Construction Grammar�Charles FillmoreU.C. Berkeley

Paul KayU.C. Berkeley

Laura MichaelisU. Colorado, Boulder

Ivan A. SagStanford University �

February 10, 2007

CENTER FOR THE STUDYOF LANGUAGEAND INFORMATION

Contents

1 Signs and Constructions 11.1 Overview of the Chapter 11.2 Feature Structures 21.3 Signs 6

1.3.1 PHONOLOGY and FORM 6

1.3.2 ARGUMENT-STRUCTURE 6

1.3.3 SYNTAX 7

1.3.4 SEMANTICS 11

1.3.5 CONTEXT 12

1.3.6 Signs: A Synthesis 13

1.4 Constructions 171.4.1 Lexical Constructions and Lexical Items 19

1.4.2 Phrasal Constructions 35

1.5 Conclusion? 50

v

1

Signs and Constructions

1.1 Overview of the ChapterWe take a language to be an infinite set of signs and the job of the grammarian to provide arecursive enumeration of the signs that constitute the language (Chomsky 1955, 1957, 1965).We understand ‘sign’ in a sense close to that of Saussure (1916). However, whereas the Saus-saurian sign simply relates ‘form’ and ‘meaning’ (signifiant, signifie), we divide up Saussure’sdichotomy of sign properties into features of phonology, (morphological) form, syntax, seman-tics, and context.

Formally, we model signs as feature structures. Feature structures are defined either asfunctions – of a kind to be explained shortly, as atoms, like acc(usative), � , bumble-bee, etc. oras referential indices, such as i (for Isaac), j (for Jacob), and so on. The set of atoms is assumed tobe finite; the set of indices is not. The interesting feature structures are the functions. The domainof a feature structure function is a subset of an assumed finite set of features (or attributes), suchas SYN(TAX), CASE, VERBFORM (VFORM), [SEMANTIC]FRAMES, etc. Members of therange (codomain) of a feature structure are either feature structures, atoms or indices.

SBCG makes a strict distinction between model objects (formal representations of actualwords, phrases, and sentences), the most important of which are signs, and descriptions of modelobjects. Constructions are descriptions of model objects, but not of signs. The domain of a signis always the set of features PHON(OLOGY), FORM, SYN, SEM(ANTICS) and CONTEXT. Asign has no DAUGHTERS (DTRS) feature. Signs do not exhibit constituent structure. In informalterms, signs correspond to the nodes of a constituent structure tree.

Constructions are not descriptions of signs but of constructs. A construct is a feature struc-ture (model object) with a MOTHER (MTR) feature and a DAUGHTERS (DTRS) feature. Con-structs correspond intuitively to local trees (interpreted as bits of language, not of grammar). Thevalue of the MTR feature is a sign and the value of the DTRS feature is a list of signs, possiblyempty. (Values of features can be either feature structures or lists of feature structures). Signsand constructs, we have noted, are feature structures and belong to the model of the language.Constructions are descriptions of classes of linguistic objects and constitute part of the grammar,the description of the model of the language.

Since our representations of descriptions of feature structures naturally look a lot like ourrepresentations of the feature structures they describe, we need a notational convention to distin-

1

2 / SIGN-BASED CONSTRUCTION GRAMMAR

guish the two. We represent feature structures (model objects) in [[double square brackets]] anddescriptions (e.g. constructions) in [single square brackets]. The visual representation of featurestructures employed is attribute-value matrices, AVMs.

The grammar contains more than the constructions, notably certain prior limitations on pos-sible feature structures. These constitute the grammar’s signature, some aspects of which areimplied by things we have already noted. For example, the signature specifies a finite set of fea-tures (possible domain members of feature structures), a finite set of atoms, and a set of indices.It also contains a set of types of feature structures. The types jointly form a multiple inheritancehierarchy, which includes all the feature structures in the domain of discourse. For example thetypes phrase and word are both subtypes of the type expression, while the types word and lexemeare both subtypes of lex(ical)-sign. So word inherits constraints from both expression and lex-sign. Types are in effect bundles of constraints on feature structures. Each type is defined by theintersection of all the constraints of all its supertypes, plus any additional constraints it has on itsown. For example, a lexical-item (lex-item) is, as we will see, a kind of (lexical) construct, whichI would reword this.

The intersection is onthe level of the setsof fs-s that satisfy theconstraints, not theconstraintsthemselves.

means that it has both MTR and DTRS features. A constraint on the type lex-item that it doesnot inherit from any supertype is that its DTRS value is the empty list. The grammar signaturealso specifies for each type the features of its domain and the type of the value of each of thesefeatures. In effect, the grammar signature establishes a large space of feature structures out ofwhich the grammar selects the signs that constitute the language.

In order to provide for a recursive enumeration of the signs that license the sentences of thelanguage, a finite set of sentential sign descriptions must be provided. Assuming there is morethan one sentence-level construction in the grammar, the several members of this set representthe various major sentence types of the language.

The rest of this chapter will provide further details about the various kinds of signs andconstructs on the model side and types and constructions on the description side.

1.2 Feature StructuresWe will assume that grammatical objects of all kinds (including signs, case values, parts ofspeech, and constructions) are modeled as feature structures. We make the further assumptionthat feature structures are either atoms like pl(ural), acc(usative), � , etc.) or else functions fromfeatures to feature structures.1 This is a simple, but powerful way of modeling linguistic objects,one that is already familiar from much work in phonology, where speech segments are oftenmodeled in this way. For example the following:2

(1) ��CONTINUANT �VOICED �ANTERIOR �CORONAL �SONORANT �CONSONANTAL �VOCALIC �

��1Carpenter 1992. SWB 03.2This is a � � � in the feature system of Chomsky and Halle (1968).

SIGNS AND CONSTRUCTIONS / 3

Similarly, the fundamental tenet of ‘X-Bar Theory’3 is that familiar atomic categories like NP orVP are to be reanalyzed as functions, e.g. as in (2):4

(2) �� NOUN �VERB �BAR 2

��Note that the functional nature of this kind of analysis can be obscured by linguists’ tendency towrite the value of a feature before the feature’s name, e.g. [ � CORONAL] or [+VERB] or to useother notations, e.g. �� (Harris 1946) or X (Chomsky 1970). Yet it is clear that the analytic intentis preserved by regarding such objects as functions whose domain is a set of features and whoserange is a set of feature values (e.g. the set �� in the system of Chomsky and Halle 1968 orthat of Chomsky 1970). The use of functions to model linguistic objects is thus nothing out of theordinary, though notational idiosyncrasy and the failure of mainstream generative grammariansto make their modeling assumptions explicit often obscures this fact.

Building on the more explicit ideas pioneered by computational linguistic work of the late1970s (e.g. Martin Kay’s Functional Unification Grammar) and the extensive subsequent workin GPSG, LFG, and HPSG,5 we use functions to model all grammatical objects. Grammaticalcategories, for example, are here analyzed as complexes of various properties: nouns includespecifications for the features CASE, NUMBER, and GENDER, verbs are specified for theirinflection class (as [VFORM finite], [VFORM present-participle], etc.) and will have a ‘+’ or‘ � ’ value for the feature AUXILIARY.

The same is true for signs. Saussure talked of signs as ‘associative bonds’ of sound conceptsand semantic concepts. Adding in syntactic categories, which Saussure had little to say about,we arrive at a picture of signs like the one illustrated in (3)a-c (CN stands for common noun; Nfor non-common noun; V for verb):

3Ref Chomsky 1970, Jackendoff 1977, Pullum 1985: Assuming some version of X-bar theory. In William H. Eilfort,Paul D. Kroeber, and Karen L. Peterson, eds., CLS 21 Part I: Papers from the General Session at the Twenty-FirstRegional Meeting, 323-353. Chicago Linguistic Society, Chicago IL. The X-bar theory of phrase structure. [AndrasKornai and Geoffrey K. Pullum.] Language 66, 24-50. 1990.

4These are the distinctive features of the functional analysis of NPs proposed in Gazdar et al. 1985.5Bresnan et al. 1982. GKPS, P&S 87, 94, Carpenter 92, Dalrymple et al. 1995, Richter 2004.


(3) a. ��! #"$�CN‘being a table’

b. �&%('*)+�N‘a person named Kim’

c. �,".-0/1�2�V‘an event of laughing (in the past)’

Signs, which we take to be the central objects of linguistic description, are also to be individ-uated as functions that specify a form, a meaning, contextual connections, and relevant syntacticinformation (including syntactic category and combinatoric potential). These functions can bedescribed by attribute-value matrices, i.e. diagrams like the following:6

(4) a. �� PHONOLOGY ��%!')3�SYNTAX NPSEMANTICS ‘a person named Kim’

� �b. �� PHONOLOGY �,"�-4/1�$�SYNTAX V[fin]

SEMANTICS ‘a laughing event situatedprior to the time of utterance’

��And, following work in HPSG (Pollard and Sag 1987, 1994), we extend the notion of sign tophrases, recognizing feature structures like (5) for complex linguistic expressions:I’m not sure what to

do about Chuck’sissues with this. IAS

(5) a. �� PHONOLOGY ��5�68719!"�'*:(;#<='.>?�$�SYNTAX NPSEMANTICS ‘the set of properties common to all linguists’

��b. �� PHONOLOGY ��@!-A�B"�-4/1�$�SYNTAX S[fin]

SEMANTICS ‘the proposition that there was a laughing eventsituated prior to the time of utterance wheresomeone named Pat did the laughing’

� ��The non-atomic feature structures we use to model linguistic objects are total functions.

That is, once an appropriate domain (a set of features) is established for a particular kind of

6The semantics here is informal. The reader may think of the descriptions in single quotes as Saussure did – asconcepts, i.e. psychological objects.


feature structure, every feature structure of that kind assigns some appropriate value to everyfeature in that domain. The value assigned to any feature must also be a feature structure, hencethat value is either an atom, or else it too is a function that must assign a value to every featurein its appropriate domain. A feature structure is thus always ‘complete’ in a simple, intuitivesense: every feature in a function’s domain is assigned a value in the appropriate codomain.7

Only atomic feature structures lack the property of containing features which require values.It is important to notice that although feature structures themselves are complete, feature

structure descriptions may be as partial as you like. This is crucial because almost every dia-gram in this monograph employs feature structure descriptions and partiality will be rampant.Lexical entries will be formulated as partial feature structure descriptions (typically being trueof (or ‘satisfied by’) many feature structures), as will grammatical constructions of all kinds. Yetunderlying all our concerns will be the set of feature structure that are specified by the theory wepresent. If some aspect of our theory goes awry, we should be able to figure out why by isolatingcertain complete feature structures that don’t satisfy the constraints of our theory but should, orother feature structures that shouldn’t satisfy our theory but do.8 Expand at will. IAS

Our feature structures have one more property that isn’t part of the standard presentationof the basic theory of functions – we assume that feature structures are organized in terms of atheory of linguistic types. The benefit derived from assigning feature structures to types is thatwe can thereby better organize the properties that classes of grammatical objects have and sim-plify their description, as well. Intuitively, it makes no sense (in English, anyway) to ask whatcase a verb has or whether a noun is an auxiliary noun; certain grammatical feature specifica-tions are appropriate only for certain kinds of grammatical objects. This intuition is given formalexpression in terms of types that particular feature structures instantiate: each feature structureinstantiates a particular feature structure type and this type assignment guarantees that the featurestructure in question specifies values for a particular set of features and that each feature’s valueis a particular kind of feature structure (possibly, a function of a particular type; possibly an atom,e.g. nominative or � ).

The properties that feature structures of particular types have are dictated by an importantpart of a Sign-Based Construction Grammar, namely its signature,9 which includes: Paul, we need some

back and forth hereabout how muchdetail toinclude...IAS

1. a set of grammatical types, organized into a hierarchy2. a set of features,3. appropriateness declarations, stating which features are appropriate for (feature structures

of) each type and what type of value each feature must have, and4. further constraints that feature structures must obey. We will assume that these are all

type constraints of the form:CEDGF ,

7Notice that these assumptions make it unnecessary to impose an additional requirement (as in Fillmore and Kay(1996) and Kay (2002)) that features must have values in linguistic models.

8Something condescending and moralistic about the confusion of description and modeling in modern linguistics.9A grammar signature is like a musical key or time signature, laying out the way in which the particulars of the

grammar (or the musical piece) are to be interpreted. For a more precise presentation, see Pollard and Sag 1994 or, for amuch more detailed and fully formalized presentation: Richter 2004. Sag, Wasow and Bender’s (2003) textbook is a veryaccessible introduction to an English grammar with these components.


where C is the name of some type of feature structure, and F is some featurestructure description.

All four components of a grammar signature may contain both universal and language particularelements. Together, these components define a space of well-formed feature structures. But onlya subset of these are licensed by the constructions of the language, as explained in section 1.4below.

1.3 SignsIn the following sections, we introduce the specific features whose values serve to distinguish thesigns of a language from one another.

1.3.1 PHONOLOGY and FORMWe will have little very little to say here about morphology, and nothing at all about phonology,but we fully intend for phonological and morphological entities to be part of our linguistic signs.Throughout this monograph, we will display the sequences of (morphological) objects that ouranalyses associate with the signs we discuss, leaving it to a largely autonomous set of constraintsto characterize the relation between the phonological and morphological aspects of signs.

We thus assume two distinct sign-level features: PHONOLOGY (PHON) and FORM:

(6) a. The value of the feature PHON is a phonological phrase ( H -phr); we assume that theseare modeled as feature structures of a particular type.

b. The value of the feature FORM is a sequence of morphological objects (formatives);these are the elements that will be phonologically realized within the sign’s PHONvalue.

Here we leave open the precise characterization of formatives, though we will assume thatthey include lemmas, and affixes. Our morphological functions will take as input both a for-mative and a list of lexeme identifiers (see the discussion of the feature LEXICAL-ID below),allowing us to accommodate morphological operations that make such distinctions as the follow-ing:

(7) a. lie/lay/lain ‘rest, recline’ vs. lie/lied/lied ‘tell falsehoods’b. can/could ‘be able to’ vs. can/canned ‘to put into cans’c. fly/flew (basic sense) vs. fly/flied (various derived senses)d. sell/sold vs. cell/cellede. write/wrote/written vs. right/righted/righted

For present purposes, we will simplify our presentation of signs by subsituting conventionalorthography for lists of formatives. We return to a discussion of related issues in Chapter 5.

1.3.2 ARGUMENT-STRUCTUREThe basic purpose of the ARGUMENT-STRUCTURE (ARG-ST) feature is to encode the com-binatoric potential of a lexical expression by listing its potential syntactico-semantic arguments.The order of elements on the ARG-ST list corresponds in the main to that of the ‘Accessibility Hi-erarchy’ of Keenan and Comrie (1977), where the first NP is the verb’s subject, the second NP (ifthere is one) is the verb’s direct object, etc. This ‘rank-based’ encoding of grammatical relations,


as shown by Keenan and Comrie and over a quarter century of research in relation-based syn-tactic theory, is independently motivated by cross-linguistic generalizations (e.g. relative clauseaccessibility), as well as by rank-based phenomena (binding, ‘advancements’, etc.) internal to thegrammars of many diverse languages. The rank-based encoding also eliminates the need for aninventory of features like SUBJECT, OBJECT, OBJ2 (SECOND-OBJECT), or COMP (COM-PLEMENT) to name particular grammatical ‘functions’. In a ‘nominative-accusative’ languagelike English, the verb’s subject is identified both as its XARG member (see the discussion ofthe feature XARG in section 1.3.3 below) and as the first member of its ARG-ST list.10 Othernominal elements on an ARG-ST list are objects.

Variable polyadicity of a given lexeme, e.g. active vs. passive vs. middle verbs, causative vs.inchoative verbs, or oblique-argument vs. ditransitive verbs, involves differences in the ARG-STlist. These differences can come about in two distinct ways in a SBCG: by derivational construc-tion (as in passivization) or by lexical underspecification (as in so-called spray/load alternations).

A lexical item like donate, which is a transitive verb, has an ARG-ST list with three mem-bers:11

(8) I NP , NP , PP JLexemes, especially verbal lexemes (see below), fall into diverse classes, as determined in part bythe length of their ARG-ST list and the constraints imposed on particular arguments. Only lexicalsigns (lexemes or words) specify a value for ARG-ST, as guaranteed by the appropriatenessdeclarations of the grammar signature.

ARG-ST lists are also the locus of the constraints of binding theory.12 For example, a reflex-ive on an ARG-ST list must be coindexed with a preceding element, if there is one (Principle A);personal pronominals must not be (Principle B). WILL WE DISCUSS

THIS MORESOMEWHERE RECLAUSES?

1.3.3 SYNTAXThe values of the feature SYNTAX are feature structures of type syntax-object (syn-obj). Theseare functions that specify values for the three features CATEGORY, VALENCE, and MARKING,which we discuss in turn.

CATEGORYThe values of the feature CATEGORY are complex grammatical categories, treated here as fea-ture structures of type category (cat).13 The various subtypes of cat will specify values for appro-priate features. For example, the signature of the grammar of English we assume here includes

10We assume, with Manning 1996 and Manning and Sag 1998, that in a syntactically ergative language, the verb’sXARG member is identified with the second member of its ARG-ST list.

11Some abbreviations:

NP = K SYN L CAT nounVAL MON+P?QMON = the empty list.

12This follows a tradition that begins with the Relational Grammar proposals of Johnson (1977). See also Pollard andSag (1992, 1994) and Manning and Sag (1998).

13Note that ‘CATEGORY’ denotes a feature and category denotes a type. Features are represented in CAPITALS andtypes in lower case italics. The common supertype of the types that can serve as values for the feature CATEGORY couldhave been named something different, e.g. part-of-speech.


the following information:

(9) a. The immediate subtypes of the type category are: noun, verb, preposition (prep), ad-jective(adj), R#R�R

b. CASE is used to distinguish the cases of nouns; the possible values of CASE (inEnglish) are nominative (nom) and accusative (acc).14

c. VFORM (VF) is used to specify the morphosyntactic category of a verb; the possiblevalues of VF are finite (fin), base, present-particple (prp), R#R�R .

d. AUXILIARY (AUX) is used to specify whether a verb is also an auxiliary; the possi-ble values of AUX are � and � .

This partial signature countenances complex grammatical categories like those shown in (10),but none like the ones pictured in (11):

(10) a. �� nounCASE nomR#R�R

�� b. �� verbVF finAUX �R�R#R

��

c. �� verbVF prpAUX �R�R#R

��

(11) a.

*

�� nounVF finR#R�R

�� b.

*

�� verbAUX �CASE nomR�R�R

� ��

c.

*

�� nounAUX �VF prpR�R�R

� ��

We should make clear that we use attribute-value matrix (AVM) notation to formulate ourfeature structure descriptions. The objects these formulas describe are functions of the appro-priate kind. For example, a nominal category is a function whose domain includes CASE, butnot AUX or VF, while a verbal category is a function whose domain includes AUX and VF,but not CASE. Note that when we mean to illustrate a particular feature structure, rather than afunctional description, we use double outermost brackets, as in (10)–(11).

Lexical descriptions are typically minimal, specifying perhaps a FORM value, a syntacticcategory and a meaning. But the set of possible feature structures that are licensed by any givenlexical description is circumscribed by the constraints of the grammar signature, which requirethat certain feature must have a value and that the value must have certain properties. For exam-ple, the lexical entry licensing the proper noun Dale says nothing about the value of the featureCASE. But any given occurrence of the proper noun Dale is modeled by a feature structure wherethe CASE value is resolved. In Dale left, it is resolved as nominative; in Find Dale!, it is resolvedas accusative.

14Note that genitive nominal expressions are not treated in terms of case. This is because case is a property of headnouns and the Modern English ’s is a phrasal clitic that appears in final position of a genitive NP, rather than as aninflection on the head noun:

(i) [[The man on the radio’s] voice] S?STS(ii) *[[The man’s on the radio] voice] S?STS

Genitive NPs are treated more fully in Chapter ??.


The fact that we model linguistic entities as total functions that can be underspecified bya linguistic description has further utility. If there is more than one model corresponding to agiven sentence description, this means that a sequence of formatives or a phonological structureis ambiguous. For example, the descriptions in (12) describe more than one feature structure, andhence predict the appropriate ambiguities:

(12) a. U FORM I I , forgot , how , good , beer , tastes J?V([[how good] [beer tastes]] vs. [how [[good beer] tastes]])

b. U FORM I flying , planes , can , be , dangerous J V(flying is an adjective modifying planes or else a gerund whose object is planes)

c. ��PHON K H -phrase��W�XY"Z�(9."�9$[ 6(\Z:]X]^([*� Q

�(I’ll believe in you vs. I’ll be leavin’ you)

There are three other CAT features that need to be introduced:15

(13) a. SELECT is used to let an expression select what it can modify or combine with as a‘marker’. The value of SELECT is a possibly empty list of signs. If an expression’sSELECT value is nonempty, then it is either a modifier (adjective, adverb, etc.) or a‘marker’ (complementizer, determiner, etc.).

b. EXTERNAL-ARGUMENT (XARG) is used to specify the argument of an argument-taking expression that is visible from outside its local domain (i.e. from outside thephrase it projects). The value of XARG is a possibly empty list of signs. The exter-nal argument of a clause is its subject; an NP’s external argument is its prenominalgenitive NP, if there is one (the XARG list of the NP is empty, otherwise).

c. LEXICAL-ID (LID) is used to individuate lexical items; the value of LID is a possiblyempty list of frames specifying the meaning of a lexeme, e.g. I book-frame J , �_� , R�R#R.

We will discuss SELECT and XARG in more detail in section 1.4 below; LID is discussed inChapter 5.

VALENCEThe basic function of the feature VAL(ENCE) is to encode the degree of saturation of any lin-guistic expression, i.e. which of its arguments it has yet to combine with syntactically. VAL isthus closely related to the feature ARG-ST – these features both take a possibly empty list ofsigns as their value.

In general, a verb’s VAL list, whose members we refer to as the verb’s valents, is exactly thesame as its ARG-ST list, as long as no covert realization takes place. Although phrases have no

15The feature SELECT was originally proposed by Van Eynde (1998) as a generalization of the two features MOD andSPEC that were employed by Pollard and Sag (1994). See also Van Eynde 2003, 2004, 2006. The fundamental insightsof the SELECT analysis here are indeed those of Van Eynde, despite minor differences of execution that might seem toindicate otherwise. For example, Van Eynde follows the basic feature inventory and more complex feature geometry ofPollard and Sag, which we have been concerned with streamlining, e.g. by eliminating the features HEAD and LOCAL.Similarly, the fact that we treat SELECT as list-valued is to provide more uniformity in the treatment of constraints thanPollard and Sag were able to achieve.


ARG-ST in SBCG, a verb phrase like persuaded me to go, which contains all but the first of theverb’s syntactico-semantic arguments (its subject), has the following singleton VAL list:

(14) I NP JSimilarly, the clause My dad persuaded me to go, which contains all the verb’s syntactico-semantic arguments, has an empty VAL list:

(15) I3JThe lexical head of the clause is the verb, and the phrases that it projects gradually ‘saturate’ theverb’s VAL list by ‘removing elements from it’.16 Clauses, NPs, pronouns, and proper nameshave an empty VAL list because they are already saturated, i.e. they need not, indeed cannotcombine with subjects or complements.

Discrepancies between a lexeme’s ARG-ST list and its VAL list can arise in several ways,e.g. via long-distance dependencies, which will not concern us until Chapter ??. Of direct rele-vance, however, is the phenomenon of null instantiation, which arises when a lexeme undergoesone of the derivational constructions discussed in Chapter ??.17

MARKINGThe feature MARKING (MRKG), introduced by Pollard and Sag (1994) and refined in crucialways by Van Eynde,18 is used to distinguish expressions like that Kim laughed, whether Morganwas awake, and Hillary’s from their respective ‘unmarked’ counterparts Kim laughed, Morganwas awake, and Hillary. The MRKG value is unmarked (unmk) in the case of all unmarked signs,but, we will assume various other MRKG values, such as those in (16):

(16) that that-clauses, e.g. that Kim laughedwhether whether-clauses, e.g. whether Morgan was awake, whether to leavethan compared phrases, e.g. than in Rome, than Patas equated phrases, e.g. as in Rome, as I couldof some of-phrases, e.g. of Romedet ‘determined’ nominal signs, i.e. the table, Prince, we

Some prepositions lead a double life as markers, as in the case of than, as, and of.19

16This way of looking at things, which has its origin in the argument cancellation of Categorial Grammar (Ajdukiewicz1935; see also http://en.wikipedia.org/wiki/Categorial grammar) involves a ‘bottom-up’ procedural metaphor where onestarts with a verb and builds successively larger phrases of which that verb is the lexical head. It is important to recognize,however, that a SBCG, like a Context-Free Phrase Structure Grammar, is a set of static constraints defining well-formedlocal tree configurations.

17In many varieties of the Romance languages, clitic pronouns have been reanalyzed as inflectional affixes (see PhilipMiller et Paola Monachesi 2003, Les pronoms clitiques dans les langues romanes” . In Godard, Danile. (d), LanguesRomanes, problmes de la phrase simple. Paris: Editions du CNRS, pp. 67-123.) This leads to another kind of null instan-tiation, where ARG-ST elements are realized morphologically, rather than syntactically. For a detailed analysis of thisphenomenon in French, broadly compatible with the framework developed here, see Miller and Sag 1997.

18See Van Eynde 2003, 2004, 2006. Van Eynde’s MARKING values are more complex than those assumed here, inlarge part because of his need to analyze complex morphological and agreement patterns absent in English. We leaveopen the possibility of modifying our theory of MARKING to incorporate further of Van Eynde’s insights.

19Add refs: Hankamer - CLS 73 on two thans in English. Ref Bonami et al. on French. Anne Abeille, Olivier Bonami,Daniele Godard et Jesse Tseng. 2006. The syntax of French a and de: an HPSG analysis. Dans P. Saint-Dizier (ed.),Syntax and semantics of prepositions, pp. 147–162. Dordrecht: Springer.


An element that specifies such a MRKG value other than unmk is called a ‘marker’; all suchelements also specify a nonempty value for the feature SELECT. The MRKG value of a markeris passed up to its mother via constraints on the Head-Functor Construction introduced in section1.4 below. See also Chapter ??, where MRKG values are discussed in more detail.

Note that the feature inventory introduced here allows variable-grain category description.That is, we may describe traditional categories (NP, VP, S, etc.), which correspond to the descrip-tions in (17):

(17) K CAT verbVAL I+J Q (S)K CAT verbVAL I NP J Q (VP)�� CAT nounVAL I+JMKG det

�� (NP)

But in addition, it is possible to describe very fine-grained categories, e.g. (18), which is thecategory of any determined NP whose head noun is the lexical item advantage:

(18) �� CAT U LID I advantage-frame J VVAL I+JMKG det

��Multigrain descriptions of this kind are a central feature of Construction Grammar, enablinggrammatical descriptions that ‘scale up’ to deal with large data sets that mainstream generativegrammar has failed to treat, indeed, has given the appearance of being uninterested in treating, inan internally consistent manner.

1.3.4 SEMANTICSChapter ?? is devoted to a more detailed presentation of the semantic framework employed inthis monograph. For present purposes, it will suffice to introduce the following two features ofour semantic objects which serve as the values of the feature SEM:

(19) a. INDEX is used to individuate the referent of an expression. Its value is an index, es-sentially a variable corresponding to an individual (in the case of an NP) or a situation(in the case of a VP, clause).20

b. FRAMES is used to specify the predications that together determine the meaning ofa sign The value of FRAMES is a (possibly empty) list of frames.

We individuate frames in terms of feature structure types21 that the grammar signature as-sociates with particular semantic role features, each of which takes an index as its value. In thecase of most (if not all) verbs, the frame will include a SITUATION (SIT) feature, whose value

20Event nominalizations may also require treatment in terms of situational indices.21Following the approach to relations explored first by Pollard and Sag (1994, sec. 8.??) and further developed by

Davis (2001). Linking by Types in the Hierarchical Lexicon Anthony R. Davis. OTHER REFS? FRAMENET REFS?


is a situational index, and an ACTOR feature, whose value is an individual index, as shown in(20):22

(20) ��sem-objINDEX `FRAMES a �� laugh-fr

ACTOR bSIT `

��dc��

And the SEM value of a proper noun like Pat will a naming frame, with specifications for thefeatures NAME and NAMED, as sketched in (21):

(21) ��sem-objINDEX bFRAMES a �� name-fr

NAME PatNAMED b

�� c� ��

Further abbreviatory conventions for semantic values will be introduced in the next chapter.

1.3.5 CONTEXTWe will have little to say about CONTEXT (CNTXT) values here. For the sake of completeness,the reader might want to envisage a theory of CONTEXT values (feature structures of type con-text) like the one developed by Pollard and Sag (1994),23 based on the features CONTEXTUAL-INDICES (C-INDS) and BACKGROUND (BCKGRND). The context-objects in their theorylook like (22):

(22) ��context

C-INDS

�� SPEAKER indexADDRESSEE indexLOCATION-OF-UTTERANCE indexR#R�R

��BCKGRND set(proposition)

� ��The various contextual indices specify contextual elements that ground an account of indexicaland deictic expressions formulated in the style of Kaplan’s seminal work (1989). The proposi-tions specified in the BACKGROUND value correspond to the set of utterance felicity conditions,which any part of the utterance sign may in principle contribute to. That is, as a mother sign isconstructed from a sequence of daughter signs, its BCKGRND must include all elements in the

22This corresponds to a Davidsonian event variable. Explain Davidsonian events in Finnish.23See also Georgia Green 1997. The structure of CONTEXT: The representation of pragmatic restrictions in HPSG.

Proceedings of the 5th annual meeting of the Formal Linguistics Society of the Midwest; Georgia M. Green ”The natureof pragmatic information.” Grammatical interfaces in HPSG, edited by Ronnie Cann, Claire Grover, and Philip Miller.Stanford, CSLI. (2000). Ginzburg and Sag 2000; Jonathan Ginzburg and Robin Cooper. 2004 ”Clarification, Ellipsis, andthe Nature of Contextual Updates” Linguistics and Philosophy 27(3): 297-365. OTHER REFS?


BACKGRND sets of the daughter signs.It is of course possible to augment these assumptions about contextual features, incorpo-

rating, for example, features like TOPIC and/or FOCUS, as in Lambrecht and Michaelis 1996.Engdahl and Vallduvi propose an analysis of ‘information packaging’ formulated in terms of theCONTEXT values structured as shown in (23):

(23) �� context

INFO-STRUCTURE

�� FOCUS R#R�RGROUND K LINK R#R�R

TAIL R#R�R Q� ��

We will not explore such elaborations here.24

1.3.6 Signs: A SynthesisSigns are analyzed as feature structures that specify values for the five features PHON, FORM,SYN, SEM, and CONTEXT, whose values have now all been introduced:

(24) ��signPHONOLOGY phonological-phraseFORM list(formative)SYNTAX syn-objSEMANTICS sem-objCONTEXT context

� ��The immediate subtypes of sign are lexical-sign (lex-sign) and expression. The immediate sub-types of expression are word and phrase, while those of lexical-sign are word and lexeme. Thesupertype relations of word thus reflect the fact that words share properties with phrases that lex-emes lack (e.g. the ability to be the daughter of a phrasal construct, discussed in the next section)and that words share properties with lexemes that phrases lack (e.g. having an ARG-ST list).These cross-cutting properties of words are accommodated by treating feature structures of thistype as both a kind of expression and a kind of lexical sign. The resulting multiple-inheritancehierarchy is sketched in (25): ADD STUFF

ABOUT ARG-STAND LEX-ITEMS?IAS DOESN”TKNOW WHAT TOADD

(25) sign

expression lex-sign

phrase word lexeme

We are now in a position to illustrate in more detail what the various signs introduced earlierin this chapter will look like in SBCG (Recall that double brackets indicate a function, rather

24For a thoughtful, discussion of issues relevant to CONTEXT values, see Georgia M. Green, The structure of CON-TEXT: The representation of pragmatic restrictions in HPSG. (click to download) Proceedings of the 5th annual meetingof the Formal Linguistics Society of the Midwest, edited by James Yoon. Studies in the Linguistic Sciences. (1996)


than a function description):25

(26) a. ��

��

word

PHON K H -phr�&@(-0�$� QFORM I Pat JARG-ST I+JSYN

�� CAT

�� nounSELECT I+JXARG I+JR#R�R

� ��VAL I+JMRKG det

��SEM

�� INDEX bFRAMES a �� name-fr

NAME PatNAMED b

�� c��

CNTXT R�R#R

��

��25A further abbreviation:

NP e = NP & f SEM g INDEX h.i


b. ��

��

word

PHON K H -phr��"�-0/j�$� QFORM I laughed JARG-ST I NP[nom] klJSYN

�� CAT

�� verbVF finSELECT I+JXARG I NP[nom... ] k JR#R�R

� ��MRKG unmkVAL I NP[nom... ] k J

��SEM

�� INDEX `FRAMES a �� laugh-fr

ACTOR bSIT `

� � , K past-frARG ` Q c

� ��CNTXT R�R#R

��

� ��c. ��

��

lexeme

PHON K H -phr��"�-0/?� QFORM I laugh JARG-ST I NP[... ] k JSYN

�� CAT

�� verbSELECT I+JXARG I NP[... ] k JR#R�R

� ��MRKG unmkVAL I NP[... ] kdJ

��SEM


ACTOR bSIT `

�� c��

CNTXT R�R#R

��

��


(27) ��

��

phrase

PHON K H -phr�&@(-A�("�-0/1�2� QFORM I Pat, laughed JSYN

�� CAT

�� verbVF finSELECT ImJXARG I NP[nom... ] kdJR�R�R

��VAL I3JMRKG unmk

� ��SEM

�� INDEX `FRAMES a �� name-fr

NAME PatNAMED b

� � ,�� laugh-frACTOR bSIT `


� ��R�R�R

��

��We will return to various syntactic and semantic details presupposed here as we present the

theory in which these signs are embedded.26 Note, however, that the objects of our grammaticaldescription are neither trees nor sets of trees put into correspondence by constraints or transfor-mations. Rather, signs are constructed by a single, recursive process that builds words from oneor more lexemes and phrases (phrasal signs) from one or more expressions. Each step of this pro-cess (which is representable as a derivation tree) involves a construct that consists of a motherand its daughters, as illustrated in (28):

26Note that since feature structures are total functions, it follows that all signs have a value for FORM, SYN, SEM,and CONTEXT. Similarly, all lexical signs must in addition specify a value for ARG-ST. Crucially, however, a lexical orphrasal description may specify minimal information.


(28)

�� phrase

FORM n Kim , laughed op&p�p��

�� word

FORM n Kim op&p�p�� word

FORM n laughed op&p�p��

�� lexeme

FORM n laugh op�p&p� � � ��

The constraints that license which constructs are legitimate in a given language are called con-structions.

1.4 ConstructionsAs emphasized above, a grammar signature defines a space of well-formed feature structuresby placing general constraints on the domain and range of each type of (functional) featurestructure. This space includes an infinite set of signs. But only a subset of the signs consistentwith the constraints of a grammar signature are well-formed signs of the language in question.For example, we want to ensure that (29) is not a sign of English:


(29) ��

��

phrase

PHON K H -phr�&@(-A�!��q![� QFORM I Kim, the JSYN

�� CAT

�� verbVF finSELECT ImJXARG I NP[nom... ] rsJR�R�R

��VAL I3JMRKG that

� ��SEM

�� INDEX `FRAMES a �� name-fr

NAME BoNAMED b

� � ,�� sneeze-frACTOR bSIT `


� ��R�R�R

��

��Given what we have said so far, (29) is a well-formed feature structure of type sign. Each

feature (PHON, FORM, SYN, SEM, and CNTXT) has a value of an appropriate type and eachof these values is a function that also specifies well-formed values for all and only the appropri-ate features, and so forth. However, even though this is the case, the problems with the featurestructure in (29) are numerous:

(30) a. This is a finite clause whose FORM value is I Kim , the J , yet Kim the is not a well-formed English clause..

b. The FORM value I Kim , the J has been phonologically realized as ��@!-A�!�.q([*� .c. The meaning of the sentence is (roughly) that a person named Bo sneezed at sometime in the past, yet that meaning cannot be expressed by uttering Kim the.R�R#R

Clearly, we need more grammar than just the inventory of types and the feature declarations thatare specified in the grammar’s signature.

We rule out unwanted signs like (29) by refining our grammar in two ways. First, we intro-duce an inventory of constructions – a constructicon – and a general principle requiring well-formed signs to be licensed by some construction. Second, we introduce further type constraints,which impose significant general constraints on the space of signs.

A construction in SBCG is not simply “any conventionalized pairing of form and meaning”,as assumed by Goldberg (1995, p. 4) and much previous work in Construction Grammar. Rather,a construction in SBCG is a constraint that licenses a particular class of feature structures byspecifying certain properties that they must have. To be sure, these constructions will induceconventionalized pairings of form and meaning, but the mechanism by which they achieve this


is circumscribed with particular reference to mothers and their daughters, in a way that we nowdescribe.

Signs in SBCG are licensed by a general grammatical principle which we formulate as fol-lows:

(31) The Principle of Construction:Every sign must be the mother of a construct licensed by some construction of thegrammar.

This principle presupposes that the constructions of the grammar license well-formed constructs.These constructs, even lexical items, are uniformly structured as a configuration of a motherand its daughters. Only licensed constructs give rise to well-formed signs. The goal of the nexttwo sections is to lay out the assumptions that underlie this pristine formulation of grammaticaltheory.

1.4.1 Lexical Constructions and Lexical ItemsThe properties of a word – a verb, say – are partly determined by a lexical entry, partly byconstraints on particular lexeme types, and partly by one of a set of inflectional constructionsthat build words from lexemes.27 The morphological base, semantic frame, and valence list oflaughed, for example, are specified in its lexical entry (on which, see below), but its inflectedform and the added constraint that the CASE value of its subject (its first, and only, ARG-STmember) must be nom, is determined by the preterite construction, one of a handful of inflectionalconstructions that survive in the grammar of Present-Day English.

We fit lexical constructions uniformly into a two-level mode, one that is articulated in termsof a mother and its daughter(s).28 For example, the verbal word whose form is laughed is con-structed from the verbal lexeme whose form is laugh, in accordance with the Preterite Construc-tion. We will refer to a mother-daughter configuration of this sort as a construct.

In order to express constructional generalizations in a systematic way, it is useful to modelconstructs as feature structures of the form sketched in (32):

(32) K MTR signDTRS list(sign)Q

MOTHER (MTR) is used to specify the sign that is constructed in a given construct; the value ofMTR is a sign. The feature DAUGHTERS (DTRS) specifies the more basic sign(s) which mustbe licensed in order for the mother to be; the value of DTRS is a (possibly empty) list of signs.It is feature structures of this kind that are licensed by particular constructions. These licensedconstructs in turn give rise to the set of well-formed signs, as per the Principle of Construction in(31) above.29

27Our approach to morphology here is realizational, perhaps closest in spirit to the realizational approach developedby Stump (refs) and others (refs). Lexical affixes are not signs; rather affixation (as well as more complex morphologicalprocesses) are as dictated by the morphological functions associated with specific lexical constructions.

28The approach to lexical constructions adopted here is based on ideas developed originally (to the best of our knowl-edge) by Copestake 1992. See also Sag et al. 2003, ch. 16, Sag to appear, Koenig 1999, and OTHER REFS??

29The terms ‘construct’ and ‘construction’ thus have a meaning here that is distinct from the way these terms have beenused in the Berkeley Construction Grammar (BCG) tradition exemplified by Fillmore, Kay and O’Connor 1988, Fillmoreand Kay 1996, Kay and Fillmore 1999, Kay 2002a[tags], Michaelis and Lambrecht (000), Michaelis (000), others??? In


What then is a construction? In SBCG, a construction is a constraint defining the propertiesthat are common to all members of a family of constructs. That is, a construction is a constraintof the form shown in (33), where x-cxt is the name of some construct type, i.e. some subtype ofthe type construct:

(33) K x-cxtR�R#R QEach construction licenses a grammatically distinct class of constructs.

The immediate subtypes of construct are lexical-construct (lex-cxt) and phrasal-construct(phr-cxt). Lexical constructs (following Sag et al. (2003, ch. 16)) are further classified in terms ofthe subtypes derivational-construct (deriv-cxt), inflectional-construct (infl-cxt), and postinflectional-construct (pinfl-cxt).30 In addition, we introduce lexical-item (lex-item) as a subtype of lex-cxt.This hierarchy of construct types is sketched in (34):

(34) construct

lex-cxt

deriv-cxt infl-cxt pinfl-cxt lex-item

phr-cxt

Lexical constructs, which we discuss in the remainder of this section, are constrained by thefollowing type declaration, specified in the grammar’s signature:

(35) lex-cxt D U DTRS list(lex-sign) V(The daughters (if any) of a lexical construct are all of type lex-sign, i.e. are words orlexemes.)Paul worries that this

is all too much forthe Finns.Not for the FinnsIvan knows...:-)

Lexical ItemsLet us begin with lexical items. In most feature-based theories of grammar, these are specifiedin terms of lexical entries that provide a partial description of an item’s form, syntactic category,and meaning. In SBCG, by contrast, lexical items are like all constructs – they have a MTR andDTRS value. However, lexical items are subject to a constraint requiring the DTRS value to bethe empty list, as illustrated in (36):31

(36) lex-item D [DTRS I3J ](Lexical items have no daughters.)

the BCG tradition, a construct was any fleshed-out ‘feature structure tree’ (See Kay 2002b[inf. sketch]) – of any degreeof configurational complexity, including single node feature structure trees. In the present approach, feature structuretrees have been recast as feature structures specifing values for the features MTR (a sign) and DTRS (a list of signs). Inaddition, the notion ‘construct’ is restricted to the intuitive equivalent of a fully determinate local tree. A constructionis the grammatical constraint (analogous to a ‘rule’) that licenses a particular set of constructs. The major theoreticalchanges, are thus: (1) replacing trees-with-feature-structures-at-the-nodes (feature structure trees) with feature structuressimpliciter and (2) imposing locality on constructions and constructs by making each in its respective domain correspondintuitively to a local tree.

30Add refs. Runner and Aronovich.31This formulation is motivated by the simplicity it imparts to the formulation of the Principle of Construction in (31)

above, putting all licensing on an equal footing.


A lexical entry like (37) licenses license a construct like (38) (pn-wd stands for proper-noun-wd.):32

(37) ��lex-item

MTR

��pn-wdFORM I Kim JSEM


NAME KimNAMED b

� �tc� ��

��(38) ��

��

lex-item

MTR

��

pn-wdFORM I Kim JARG-ST I+JSYN

�� CAT

�� nounSELECT I3JCASE nomXARG I3JR#R�R

��VAL ImJMRKG det

� ��SEM


NAME KimNAMED b

�� c��R�R�R

��DTRS I+J

��

��Since a lexical entry is a construction, i.e. a licensing constraint listed in the constructicon, a

lexical entry licenses a class of lexical items (feature structures of type lex-item), each of whichis a function specifying appropriate values for all appropriate features. Notice that there must bea value for CASE in (38) (either nom or acc), since (38) is a total function (as are all the functionswithin it), not a constraint on functions.

32We assume here that the predications of proper names are part of their semantic content. Alternate views are possible;see, for example, Pollard and Sag (1987, 1994), who treat these predications as part of the set of background conditionsspecified within CNTXT values.


The grammar signature contains various type constraints that impose further conditions onthese feature structures. The value of the MTR feature in (38), for example, obeys the type con-straint sketched in (39):

(39)

pn-wd D �� SYN

�� CAT�� nounSELECT I+JXARG I+J

��VAL I+JMRKG det

��

And since (38) is a well-formed feature structure licensed by a construction in the constructicon,it follows from the Principle of Construction in (31) that a sign like the following is well-formed:

(40) ��

��

pn-wdFORM I Kim JARG-ST I+JSYN

�� CAT

�� nounSELECT ImJCASE nomXARG ImJR#R�R

��VAL I+JMRKG det

��SEM


NAME KimNAMED b

��uc� ��R�R�R

� ��

� ��A lexeme must thus always be the mother of some licensed construct, be it a lexical item or

a derivational construct (see below). An intuition behind the theoretical and terminological inno-vations presented here (which distinguish SBCG from earlier work in Construction Grammar),is that constructions build complex objects out of simpler ones and that constructs can be viewedas the result of such ‘acts’ of construction, except in the case of lexical items, which serve toground the sign recursion, like terminal nodes in a Context-Free Grammar. Crucially, however,constructions, like the rules in a Context-Free Grammar, are static constraints on mother-daughterconfigurations and needn’t be thought of in a directional way.I’ve woven the

comparison withBCG into the prose,but it might still beforeshadowed inChapter 1, as Paulsuggested. IAS

For various reasons, the class of lexical items that satisfy any lexical entry is infinite. This istrue even in the case of the lexical entry for Kim given in (37) above, because there are infinitelymany indices that could serve as the value of the feature INDEX in (38). This in turn meansthat our grammar licenses infinitely many feature structures like (40), differing from one another


only in terms of the indices they contain. Various of the features specified within the CNTXTvalue (e.g. SPEAKER, ADDRESSEE) are also index-valued, further expanding the space offeature structures. However, all feature structures that differ only in this way are equivalent forgrammatical purposes; the only significant distinction among these functions is the value for thefeature CASE.

In other circumstances, a given construction will license infinitely many feature structuresthat differ from one another in grammatically significant ways. One place where this arises hasto do with the feature structures licensed by verbal, adjectival, and prepositional lexical entriesthat specify a nonempty ARG-ST (and VAL) list. For example, the lexical entry for the lexemelaugh is sketched in (41) (siv-lexeme abbreviates strict-intransitive-verb-lexeme.):

(41) ��

siv-lexemeFORM I laugh JSYN

��CAT K verb

SELECT I+J Q �ARG-ST I NP k JSEM


ACTOR bSIT `

�� c��

��This lexical entry licenses infinitely many feature structures like (42):

(42) ��

��

siv-lexemeFORM I laugh JARG-ST I NP k [ R�R#R ] JSYN

�� CAT

�� verbSELECT I+JXARG I NP k [ R#R�R ] JR�R#R

��MRKG unmkR#R�R

��SEM


ACTOR bSIT `

�� c��R�R�R

��

��The key thing to see here is that the NP k [ R�R#R ] on the ARG-ST list (and the identical feature

structure on the XARG list; see below) must be fully specified in (42), even though neither the


lexical entry in (41) nor any of the constraints that (42) inherits places any constraints on whatthis NP feature structure should have as its FORM or SEM value, for instance. This is as itshould be, since this NP feature structure will be identified with the subject of laugh and thereare infinitely many NP signs that could perform that function, corresponding to infinitely manysentences of the form: NP laugh/laughed/laughs.

As noted earlier, the semantic and ARG-ST properties of lexeme classes are organized bythe hierarchy of lexeme types, i.e. the subtypes of the type lexeme. This method is illustrated bythe partial lexeme hierarchy in (43):33

(43) verb-lxm

intr-verb-lxm

siv-lxm srv-lxm scv-lxm p�p�p trans-verb-lxm

stv-lxm orv-lxm ocv-lxm dtv-lxm p�p&pWe assume all verbal lexemes obey the constraint in (44), which says that a verb’s first argumentis its external argument, that verbs are unmarked and that they do not use the SELECT feature toselect their dependents:34 v 35

(44)

verb-lxm D �� ARG-ST I w , R#R�R=JSYN

�� CAT K verbXARG I w J Q

MRKG unmk

��

But constraints are stated at diverse levels, so as to affect, for example, all lexemes, all verballexemes, all intransitive verb lexemes, or all lexemes of a particular maximal (leaf) type. A givenlexeme must obey the constraints specified in the lexical entry that license it, but also all theconstraints that affect the types that it belongs to. A ditransitive lexeme, for example, must obeywhatever constraints affect dtv-lexeme, trans-verb-lexeme, verb-lexeme, and feature-structure. Agiven lexical entry can thus be streamlined so as to include minimal stipulation, leaving it to thetheory of lexical classes (embodied in the lexical type constraints) to determine which lexicalproperties are compatible with it.36

This approach allows for underspecification of a kind that would permit a given lexical entryto license lexemes of more than one lexical type. We presume this is the right way to analyzea variety of lexical problems, including the analysis of ‘spray/load’ alternations. That is, we

33We abbreviate as follows: strict-intransitive-verb-lexeme (siv-lxm, e.g. die), subject-raising-verb-lexeme (srv-lxm,e.g. seem), subject-control-verb-lexeme (scv-lxm, e.g. try), strict-transitive-verb-lexeme (stv-lxm, e.g. devour), object-raising-verb-lexeme (orv-lxm, e.g. believe), object-control-verb-lexeme (ocv-lxm, e.g. persuade), ditransitive-verb-lexeme (dtv-lxm, e.g. hand).

34We put aside here proposals such as that of Sag (1997), where verbs heading relative clauses bear non-empty valuesfor the MOD feature, which could be recast here in terms of non-empty SELECT specifications.

35In the formulation of constraints, boxed integers are used as variables ranging over feature structures and boxedletters as variables ranging over lists of feature structures. Variables appear twice (or more) when they appear in aconstraint, since they are used to impose identity on the values of two distinct features.

36This kind of simplification is typical of object-oriented analyses of complex data.


assume that a verb like spray has one lexical entry whose MTR value is compatible with twodistinct subtypes of the type lexeme. CAN WE MAKE

THIS MOREPRECISE, MAYBEWITH A REALANALYSIS THATGIVESCONSTRAINTSFOR EACHSPRAY-LOADTYPE?

Morphological FunctionsIn the next two sections we discuss inflectional and derivational constructions. A key part ofsuch word-building constructions are the morphological functions. As part of that discussionwe present, as an example of an inflectional construction, the Preterite Construction. This con-struction builds, from a verbal lexeme daughter, the preterite form of that lexeme (which is oftype word) as mother. The mother displays the appropriate semantics. The form of the mother isthe image of the form of the daughter under the morphological function F xzy&{1| .

Morphological functions allow us to model ‘elsewhere’ phenomena in SBSG morphologywithout changing the overall logic of the architecture, as well as to deal with other problemsposed by various kinds of irregularity.

We take the FORM value of a lexeme (of type form) to be the inflectional stem if there is one,and any stem if there are more than one. (Morphological functions will effect stem alternationsas necessary. It is not clear that more than one stem is necessary for any English lexeme, althoughmultiple stems for different tenses, cases, etc. are commonplace in languages generally.)

A member of the domain of a morphological function is an ordered pair giving the FORMvalue of the lexeme and its LID value; the range (codomain) consists of FORM values. Bothelements of a domain member are necessary because in some cases it is the FORM value thatdetermines the inflected form (e.g. have D had for all of the distinct lexemes have) and some-times the LID value (lie ’recline’ D lay, lie ‘prevaricate’ D lied). F x}y�{1| might be defined alongthe following lines:

(45) I$~0��J F xzy�{1|�� ~4�T��I be, ��J undefinedI have, �BJ hadI lie,‘recline’ J layI dream,‘dream’ J dreamtI swim, ‘swim’ J swamI buy, ‘buy’ J boughtI keep, ‘keep’ J kep+edR�R#ROtherwiseI$~0��J I�~ +ed J

Special constructions will be needed to specify the preterite forms of BE for various personand number pairings. All lexemes with FORM have will be inflected the same for preterite ashad. lie ‘recline’ will be inflected as lay by the third line of the function and lie ‘prevaricate’ willbe inflected as lie+ed by the ‘Otherwise’ clause. As shown here for illustrative purposes, F x}y�{1|will only produce the irregular dreamt preterite form of dream. A special construction would beposited to get the dream,-ed form. This formulation is consistent with the view that both formsare memorized. On the alternate view, that the regular forms are produced by the regular processand only the exceptional forms of the doublets memorized, dream would be omitted from the


exception list at the beginning of F xzy&{1| . Then dream+ed would be licensed by the Otherwiseportion of F xzy&{�| and a separate construction would be posited for dreamt. swam and bought areunproblematical irregular forms without doublets. The inflection of I keep, ‘keep’ J as kep+ed isintended to illustrate the ability of morphological functions to deal with stem changes, althoughit is arguable that this inflection, and analogous ones in English preterites, should be viewed asanything other than suppletive. 37

Inflectional ConstructionsIn addition to the general constraints on lexical constructs, inflectional constructs are subjectto the following constraints, specified as part of the grammar signature (nelist(T) stands for anonempty list, each of whose members is a feature structure of (some subtype of) type T.):

(46)infl-cxt D K MTR word

DTRS nelist(lexeme)Q(The mother of an inflectional construct is of type word; the daughters are all lex-emes.)

This constraint embodies the traditional intuition that inflectional constructions are resources forbuilding words from lexemes.38

An inflected word like laughed is modeled via feature structures like the one sketched inFigure 1.

Because this is a well-formed construct, the feature structure in (47) is licensed:

37Non-past-tense uses of the preterite morphological form, such as counterfactual conditional protases (If I had myway,...) could in principle be licensed by a separate inflectional construction that also avails itself of F �&��j� . Alternatively,one might pursue a semantically bleaching post-inflectional ‘pumping’ construction, whose mother and unique daughterdo not differ in FORM. In either case, special arrangements must be made to distinguish, for example, counterfactual Iwere... from its past tense analogue I was.... We will not resolve these issues here.

38There is usually, if not always, a single daughter in an inflectional construct. For convenience, we here ignorelanguages with layered inflection.


��

��

infl-cxt

MTR

��

wordFORM I laughed JARG-ST I NP k [nom R#R�R ] JSYN

�� CAT

�� verbVF finSELECT I+JXARG I NP k [nom... ] JR�R#R

��MRKG unmkVAL I NP k [nom R�R#R ] J

��SEM

�� IND `FRAMES a K exist-fr

BV ` Q ,�� laugh-frACTOR bSIT `

�� , K past-frARG ` Q c

��R#R�R

��

DTRS a

��

lexemeFORM I laugh JARG-ST I NP k [nom R�R#R ] JSYN

�� CAT

�� verbVF finSELECT I+JXARG I NP k [nom... ] JR�R�R

��MRKG unmkVAL I NP k [nom R�R�R ] J

��SEM


ACTOR bSIT `

�� c��R�R#R

� ��

c

��

��FIGURE 1 A laughed Construct


(47) ��

��

wordFORM I laughed JARG-ST a NP k K nomR#R�R Q cSYN

��CAT

�� verbVF finSELECT I+JXARG I NP k [nom... ] J

��VAL a NP k K nomR�R#R Q cMRKG unmk

� ��SEM

�� IND `FRAMES a K exist-fr

BV ` Q ,�� laugh-frACTOR bSIT `


� ��R�R�R

��

� ��Several observations are in order here. First, the SYN values of the mother and the daugh-

ter in Figure 1 are identical. Second, we have provided a Davidsonian analysis of the preteritethat introduces a tense frame taking the situation of the lexeme’s frame as its argument and anexistential quantifier binding the situation variable.39 This ignores many interesting issues in thesemantics of tense in English; but however the semantic analysis of past tense is articulated, thepast tense semantics will be absent from the lexeme daughter in Figure 1, but present in the se-mantics of the preterite word. Third, preterite constructs like the one in Figure 1 belong to thetype infl-cxt, and hence must obey all constraints affecting feature structures of that type. Fourth,the feature structure illustrated here involves the default assumption that the VAL list of a wordis identical to its ARG-ST list (i.e. that no argument has been extracted or null instantiated).Hence, given the constraint in (44) above, all three shaded NPs in (47) are identical. Similarly,the SELECT value is here realized as the empty list, reflecting the fact that this clause (unlikeits homophonous relative counterpart) cannot function as a modifier. Finally, the information en-coded in Figure 1 is exactly the same as what is presented in a more familiar diagram, the unary(non-branching) local tree in Figure 2. Because of their familiarity, we will use trees wheneverpossible to illustrate feature structures of type construct.

The Preterite Construction can now be formulated as follows:

(48) Preterite Construction (preliminary formulation):

39In the next chapter, we will refine this analysis in terms of restricted (generalized) quantification.


��

��

word

FORM n laughed oARG-ST n NP e [nom p�p�p ] oSYN

�� CAT

�� verb

VF fin

SELECT n�oXARG n NP e [nom... ] op�p�p

��VAL n NP e [nom... ] oMRKG unmk

��SEM

�� IND �FRAMES a K exist-fr

BV � Q ,

�� laugh-fr

ACTOR �SIT �

�� , K past-frARG � Q c

��p�p�p

��

��

��

lexeme

FORM n laugh oARG-ST n NP e [nom p�p�p ] oSYN

�� CAT

�� verb

VF fin

SELECT n�oXARG n NP e [nom... ] op�p�p

��VAL n NP e [nom... ] oMRKG unmk

��SEM

�� IND �FRAMES a �� laugh-fr

ACTOR �SIT �

� �uc� ��p�p�p

� ��

� ��FIGURE 2 A laughed Construct in Tree Notation

30 / SIGN-BASED CONSTRUCTION GRAMMAR��

infl-cxt

MTR

��FORM I F x}y�{1| � � �0JARG-ST � I NP[nom] , R�R#R�JSYN w L CAT U VF fin V PSEM

��FRAMES a K exist-fr

BV ` Q , � [SIT ` ] , K past-frARG ` Q c �

��DTRS a �� FORM I � J

ARG-ST �SYN wSEM U FRAMES I � J V

� �� c

��One way of paraphrasing (48) is as follows: Given a verbal lexeme40 one can construct a verbalword meeting the following four conditions:

(49) a. the word’s VF value is finiteb. the word’s FORM value is related to that of the lexeme via the morphological function

F xzy&{1| ,41

c. the word’s SYN and ARG-ST values are identified with those of the lexeme daughter,thus requiring that everything in the lexical entry that licensed the lexeme be consis-tent with the constraints introduced by this construction, e.g. that the subject valent’sCASE value be nominative, and

d. the word’s FRAMES list adds a ‘pastness’ frame and an existential quantifier to thelexeme’s FRAMES list, identifying the argument of this frame with the situation spec-ified in the lexeme’s original frame and with the variable bound by the existentialquantifier.

The morphological function F x}y�{1| is defined in such a way as to include exceptions thatpreempt the general affixation of the past-tense affix to the base formative supplied by the verballexeme. This is illustrated in (50):

(50) F xzy&{1|,�j�(� ��s�}�u� ,F xzy&{1|,�&� �� is undefined,F xzy&{1| �1� b1�� b1� ,R�R#Rotherwise, F x}y�{1|,� ��#�=�}� .

For further discussion of related lexical issues, see Chapter 5.

40It must be a verb because its CAT value must be compatible with a VF specification, and the grammar signatureensures that VF is appropriate only for feature structures of type verb.

41Recall that these morphological entities are distinct from (and more ‘abstract’ than) the phonological entities thatrealize them. See the discussion in section 1.3.1 above.


Notice that it would be redundant for the construction in (48) to explicitly stipulate that theMTR value be of type word or that the daughter must be of type lexeme. Because the constructslicensed by the Preterite Construction are all instances of the type infl-cxt, they must obey allconstraints on feature structures of this type imposed by the grammar signature. As we havealready seen (46), requires that the MTR value of all inflectional constructs be of type word andthat all daughters of inflectional constructs be of type lexeme. Hence this is true of all preteriteconstructs as well, by the process of constraint inheritance (see section 1.2 above).

Moreover, as we scale up our analysis of English morphology, this construction can be fur-ther simplified. Since all finite forms in English require nominative subjects, as shown in (51),

(51) a. She/*her walked home.b. They/*them walk home after work.c. I suggested that he/*him walk home.d. I/*me am walking home today.

we will surely want to posit a subtype of infl-cxt and preterite-cxt to express this generalization.Let us call this new type finite-cxt and posit the following type constraint:

(52)

finite-cxt D �� infl-cxt

MTR�� ARG-ST I NP[nom] , R#R�R]JSYN L CAT U VF finV P

��

Once this ‘finite words assign nominative case to their first argument’ constraint is inherited,rather, than being stipulated as part of the Preterite Construction, the latter (and all its sisterconstructions) can be simplified further:


(53) Preterite Construction (final formulation):��

finite-cxt

MTR

�� FORM I F xzy&{1| � � �0JSYN wSEM

��FRAMES a K exist-fr

BV ` Q , � [SIT ` ] , K past-frARG ` Q c �


SYN wSEM U FRAMES I � J V

� �� c

��The simplification achieved here may appear slight, but it contributes to the overall goal of type-based constraint inheritance, which is to eliminate all unmotivated redundancy from grammar.

Derivational ConstructionsDerivational constructions are governed by the following type constraint:

(54)deriv-cxt D K MTR lexeme

DTRS nelist(lex-sign)Q(The mother of a derivational construct is of type lexeme; the daughters of a deriva-tional construct are lexical signs (words or lexemes).)

Derivational constructions thus allow new lexemes to be built from one or more lexical signs. Forexample, we assume that there is an un-prefixation construction, sketched in (55), which allowsun-verb lexemes to be derived from a semantically specifiable class of verb lexemes:

(55) Un-Verb Construction:��deriv-cxt

MTR

�� FORM I F �¡ � � �0JSYN wSEM U FRAMES � ¢ R�R#R V


SYN w [CAT verb]

SEM U FRAMES � V� �� c

� ��Obviously, there are further semantic (and possibly syntactic) constraints that must be added to(55).

Since inflectional constructions are constrained so as to require a daughter of type lexeme,there is a natural relation that exists between the two types of construction: derivational con-structions feed inflectional constructions. That is, a derived lexeme, one that is the mother of aconstruct licensed by some derivational construction, can then serve as the daughter of a constructlicensed by some inflectional construction, as illustrated in Figure 3, where the two constructs are


conflated, with the shared lexeme serving simultaneously as mother of the derivational constructand daughter of the inflectional construct.

Derivational constructions, which we will have more to say about in Chapter 5, include: SOMEONEWROTE: DISCUSSCRITERIA FORSAYING WE HAVEACONSTRUCTION.BUT I THINK THISSHOULD GO INCH. 2 - IAS

(56) a. passivization, which feeds overt inflectional constructions in many languages andword-formation processes in English (see Bresnan 2001),

b. agentive noun formation,c. denominal verb formation (ref. Clark and Clark),d. various kinds of nominalization,e. the Un-Adjective Construction, andf. the -Able Adjective Construction.

An example of a binary derivational construction is English noun-noun compounding. In fact,here we posit a number of constructions, reflecting the various lexically and semantically cir-cumsribed subpatterns that are to be found in noun-noun compounding.

The general compounding construction, which appeals to a contextually salient (but other-wise arbitrary) property to relate the interpretations of two nouns, accounts for compounds likethe following:42

(57) a. pumpkin bus: ‘a bus that was used in some previous excursion to a pumpkin patchfamiliar to the relevant interlocutors’

b. Jaeger potato: ‘potato of the kind that the speaker once used for something whenspending an evening with someone named Jaeger’

It is also possible to incorporate proposals like that of Copestake and Lascarides’ (1997), whichposits a number of more specific constructions specifying patterns that fit particular classes ofnouns together in conventionalized ways. We will not undertake that task here, however. As afinal remark about compounding, note that the first member of most noun-noun compounds is alexeme (computer screen, pumpkin bus, etc.), but in others it is a word: algorithms course, salestax, etc.43

Postinflectional ConstructionsPostinflectional constructs are subject to the following type constraint:

(58)pinfl-cxt D K MTR word

DTRS list(word)Q(The mother and daughters of a postinflectional construct are of type word.)

Postinflectional constructions thus allow for words to be derived from other words. Sag et al.(2003) introduce this category as a way of incorporating a number of proposals that have beenmade (by Warner (1993), Kim and Sag (2002), Kim (2000) and others) in terms of lexical rulesthat create adverb-selecting auxiliary verbs (e.g. did (not), will (not)), as well as not-contractedwords (e.g. didn’t, couldn’t) and related elements.

42Ref. Kay and Zimmer 1975. Downing 1977. Copestake, A. and A. Lascarides [1997] Integrating Symbolic and Sta-tistical Representations: The Lexicon Pragmatics Interface, Proceedings of the 35th Annual Meeting of the Associationfor Computational Linguistics (ACL97), Madrid, July 7th–12th 1997, pp136–143.

43Pinker (REF), Kiparsky (REF), Ramscar (REF) OTHER REFS? on the supposed innateness of this not being aword...


��

��

word

FORM n untied oSYN

�� CAT

�� verb

VF fin

SELECT n�op�p&p� ��

VAL n NP e [nom... ] , NP[acc... ] oMRKG unmk

��p&p�p

��

��(Preterite Cx)

��

��

lexeme

FORM n untie oSYN

�� CAT

�� verb

VF finSELECT n�op&p�p

��VAL n NP e [nom... ] , NP[acc... ] oMRKG unmk

��p�p�p

��

��(Un-Verb Cx)

��

��

lexeme

FORM n tie oSYN

�� CAT

�� verb

VF fin

SELECT n�op&p�p��

VAL n NP e [nom... ] , NP[acc... ] oMRKG unmk

��p�p�p

��

��FIGURE 3 Un-Verb Construction Feeding Preterite Construction


Various other lexical regularities can be analyzed in terms of a postinflectional construction.For example, Sag et al. (2003) present a postinflectional analysis of it-extraposition. This can berecast in the present framework in the manner sketched in (59):

(59) Extraposition Construction:��

pinfl-cxt

MTR

��FORM I � JSYN

�� CAT wMRKG �VAL I � NP[it] J ¢ � ¢ I £ CP J

��SEM ¤

��DTRS a ��

FORM I � JSYN

�� CAT wMRKG �VAL I £ J ¢ �

� �SEM ¤

�� c

� ��This construction licenses words that take it-subjects and CP complements on the basis of theexistence of phonologically indistinguishable counterparts that take CP subjects. The words li-censed by this construction are fully equipped to project head-complement phrases of the sortdiscussed below, with the extraposed clause appearing as a complement.44 Verbs or adjectiveswhose cooccurring clauses occur only in extraposed position, e.g. the instances of (non-raising)seem and appear illustrated in (60), are simply listed in the constructicon with the same ARG-STvalue as elements that are constructed by (59):

(60) a. It seems that you’ve been called for jury duty.b.*That you’ve been called for jury duty seems.

Finally, it should be noted that it is sometimes quite difficult to see the differing conse-quences that distinguish a postinflectional analysis from a derivational one. These often have todo with the feeding relations that exist between the construction in question and other derivationalconstructions. For example, a word licensed by a postinflectional construction cannot in generalserve as the daughter of a derivational construct, because most derivational constructions requirea daughter of type lexeme. Hence, treating a given alternation via a postinflectional constructionimmediately predicts that it cannot feed a derivational construction.

1.4.2 Phrasal ConstructionsPhrasal (syntactic) constructs work in the same way as lexical constructs, except that they obeythe following general constraint:

(61)phr-cxt D K MTR phrase

DTRS list(expression)Q44For a more comprehensive analysis of English extraposition, where extraposed clauses are treated as non-

complement dependents, see Kim and Sag (2005, in press).


(The mother of a phrasal construct must be a phrase and the daughters must be ex-pressions, i.e. words or phrases.)

The Subject-Predicate ConstructionSimple declarative clauses are licensed by the subject-predicate construction sketched in (62):

(62) Subject-Predicate Construction (preliminary version):��phr-cxt

MTR��SYN K VAL I+J

MRKG � Q �DTRS a w ,

�� SYN

�� CAT K verbVF finQ

VAL I w JMRKG � unmk

�� c

��This construction says that a [VAL I+J ] phrase can be built from two daughters, as long as thesecond is a finite (and hence verbal) sign that selects for the first via the VAL feature. Independentprinciples (that is, linear precedence constraints) require that the FORM value of the mother bethe result of adding the members of the second daughter’s FORM value to the FORM value of thefirst daughter.45 Similarly, the Principle of Compositionality introduced in the next chapter re-quires that the FRAMES list of the two daughters be merged to form the mother’s FRAMES list.With the interaction of these principles, (62) licenses phrasal constructs like the one in Figure 4.Notice that the mother of this construct is just the phrasal sign illustrated in (27) above.

Although our theory of constructions is strictly localist, i.e. our constructions – like rulesin a context-free grammar – mention only a mother and its daughters, we can nevertheless ac-commodate grammatical dependencies that are non-local. In particular, we build on work in theGPSG/HPSG tradition that has used feature specifications to locally encode information aboutlong-distance dependencies. Just as the featural representation of a category like ‘NP’ encodesthe fact there is a head word within whose category is ‘noun’, other feature specifications canencode key grammatical information about an element present in (or absent from) a phrase. Forexample, the VF value of a verbal phrase (VP or S) encodes a morphosyntactic property of thehead word within that phrase. Similarly, the feature GAP46 is used to encode the absence of an‘extracted’ element (or, as mainstream generative grammarians often put it: the ‘presence of agap’) within a given phrase. As we develop a theory of such feature specifications and the princi-ples that govern their distribution throughout constructs, we will be developing a general theory

45For convenience, we will henceforth omit discussion of linear ordering, assuming that the order of elements on theDTRS list determines the order of elements on the mother’s FORM list. This is a gross simplification of a complex setof issues that have motivated ID-LP format (the separation of constructions and the principles that order their daughters)and ‘Linearization Theory’, the augmentation of sign-based grammar to allow interleaving of daughters as an accountof word order freedom. On ID-LP grammars, see Gazdar and Pullum 1983, Gazdar et al. 1985, Sag 1987 and ???. OnLinearization Theory, see Reape 1994, Mueller 1995, 1999, 2002, 2004, Kathol 2000, Donohue and Sag to appear, otherrefs?.

46Gazdar 1981, Sag 1982, Gazdar et al 1985, Pollard and Sag 1987, 1994, Sag et al. 2003. Explain history - SLASHto GAP).

SIGNS AND CONSTRUCTIONS / 37��

phrase

FORM n Pat, laughed o

SYN

��CAT

��

verb

VF fin

XARG a��word

FORM n Pat oSYN

�� CAT

�� noun

CASE nomp&p�p��

VAL n�oMRKG unmk

� ��p�p�p

� ��c

p�p�p

� ��VAL n�oMRKG unmk

��SEM

�� INDEX �FRAMES a �� name-fr

NAME PatNAMED �

� � , K exist-fr

BV � Q ,

�� laugh-fr

ACTOR �SIT �

� � , K past-fr

ARG � Q c��

��

wordFORM n Pat oSYN

�� CAT

�� noun

CASE nom

XARG n�op�p&p��

VAL n�oMRKG unmk

��SEM

�� INDEX �FRAMES a �� name-fr

NAME PatNAMED �

� �4c��p&p�p

��

��

word

FORM n laughed oSYN

��CAT

�� verb

VF fin

XARG n K FORM n Pat op�p�p Q o��

VAL n K FORM n Pat op�p�p Q oMRKG unmk

��SEM

��INDEX �FRAMES a K exist-fr

BV � Q ,

�� laugh-fr

ACTOR �SIT �

�� ,

K past-fr

ARG � Q c��

� ��FIGURE 4 A Subject-Predicate Construct


of what nonlocal information can be lexically selected at a higher level of structure or referencedby a construction higher in a phrasal derivation.

As has been recognized at least since Chomsky (1965),47 lexical restrictions are circum-scibed, i.e. they are localized in a fashion that must be made precise. Behind the search for theprecise characterization of the relevant notion of locality of selection is the clear intuition that nolanguage has, for example, a verb that requires a clausal complement that must contain an overtsubject that is feminine, or singular, etc. Early accounts of locality excluded subjects, but sinceidiosyncratic case assignment in numerous languages (perhaps most famously in Icelandic48)clearly involves the subjects of verbs, the most likely first approximation of the relevant localitydomain is: a lexical element’s grammatical dependents. We may formulate the relevant hypothe-sis as in (63):

(63) Selectional LocalityFor purposes of selection (subcategorization), (nonanaphoric) agreement, semantic roleassignment, case assignment, etc., a lexical element has access only to those elements thatit is in a grammatical relation with (subject of, complement of, etc.).

Our various features and the particular choices made about the nature of their values, takentogether with general constraints on how information is percolated as phrasal signs are con-structed, constitute a precise formulation of the basic idea embodied in (63). In particular, theinformation lexically specified on an element’s ARG-ST and SELECT lists constrain the natureof the elements it combines with, providing access to grammatical dependents, but not to ele-ments within them. And by adding a specific feature like XARG to systematically encode certaininformation about elements embedded within dependents, we in effect localize certain nonlocalinformation, making it available to an element selecting that dependent.

External ArgumentsBefore proceeding to semantic matters, there are two more features and two more constructionsthat we must explicate. Our head complement construction, which is used to build VPs, APs,PPs, and NPs, makes use of the feature XARG, which was introduced briefly in section 1.3.3above. In this section, we review some of the basic motivation for this feature.

As a number of researchers have recently shown, there are phenomena in diverse languageswhose analysis, for example, requires that a verb selecting a sentential complement, be able toplace constraints on the subject within that complement. It is interesting to examine some of thespecific seemingly nonlocal phenomena that have led to such conclusions and the proposals thatthey have given rise to in various languages.

Levine (2000) argues that the constituency of English APs headed by tough is as indicatedin (64):

47See also Kajita 1968 and Sag to appear.48See Thraınsson 1975, Andrews 1982, 19??


(64) AP

A

tough

CP

for Kim to talk to

Despite being included within the for-to clause, the subject of the complement clause is nonethe-less a semantic argument of tough, as illustrated in (65):

(65) a. Sandy is tough for Kim to talk to.b. tough(kim,talk-to(kim,sandy))

That is, as Levine argues, Kim is a semantic argument of the tough predication, even though itis not a syntactic argument of tough (i.e. not on tough’s ARG-ST or VAL list).49 This kind ofanalysis is supported by various pieces of evidence, including the ungrammaticality of exampleslike (66), where the subject of the for/to-clause is a nonreferential NP:

(66) a.*Pat is [tough [for there to be a conversation with ]].b.*Bo is [easy [for it to bother that it’s raining ]].c.*Dana is [hard [for it to rain on ]].

Assuming Levine’s structures are correct, an English tough adjective appears to violate locality– it imposes restrictions on an element within its complement clause and assigns that element asemantic role. The selection in question is rendered local, however, if we make use of the XARGspecification illustrated in (67):

(67) ��FORM I tough JARG-ST a NP k , �� CAT K verb

XARG I NP r}J QVAL I+JGAP I NP k [acc] JSEM w

�� cSEM tough(j, w )

��Similarly, Meurers (1999) discusses German examples where fronted constituents like those

in (68) and (69) are licensed by the case assigning properties of the finite verb that governs thegap associated with the dislocated element: Stefan Mueller

cautions againstreasoning thissimplistically withthese data. I haven’tseen Meurers’ mostrecent analysis,which is supposedlycompatible, so thiswill be revised.-IAS

49A more refined analysis should also include a treatment of the somewhat elusive predications that seem to hold ofthe referent of the subject of tough (cf. the contrast between (i) and (ii)), as well as the generic (iii) or pragmatic (iv)interpretation of an unexpressed second argument of tough:

(i) The sonatas are easy to play on this violin.(ii) This violin is easy to play the sonatas on.

(iii) The test is probably difficult. [generically](iv) Was the test difficult? [for the addressee]


(68) a. [Ein Außenseiter gewonnen] hat hier noch nieAn[nom] outsider won has here still never‘An outsider has never won here yet.’

b. [Einen Außenseiter gewinnen] laßt Gott hier nieAn[acc] outsider win lets god here never‘God never lets an outsider win here.’

As Meurers argues at length, the category of the fronted phrase in these examples (which isidentified with the category of the gap) must include information about the subject NP withinthe fronted phrase, so that the verb governing the gap can assign the appropriate case to it. Foranother proposal that converges with the approach developed below, see Meurers 2003.

In a similar vein, Bender and Flickinger (1999) analyze agreement in English tag questionsby allowing the subject’s agreement information to percolate up to the level of the clause. Whena clause is combined with the tag of a tag question, this agreement information is then identi-fied with that the pronoun in the tag. This induces the familiar tag question agreement patternillustrated in (69):

(69)

[They left,] didn’t

¥¦¦¦¦¦¦§ ¦¦¦¦¦¦¨they

*(s)he*we*you*I

© ¦¦¦¦¦¦ª¦¦¦¦¦¦« ?

The problem here is not selectional locality, but rather the issue of constructional locality,about which we may formulate the following hypothesis:

(70) Constructional Locality:Constructions other than lexical items make reference to properties of the daughtersigns, never (directly) to properties of elements within the daughter signs.

Notice that by defining constructions in terms of constructs, where the latter are well-formedsigns with certain specified properties – no matter how they have been constructed, we deriveconstructional locality as a direct consequence of the architecture of SBCG. If the properties ofsomething used to construct a given sign (i.e. something ‘inside of it’) determine what constructsthat sign can be part of, those properties must be specified as part of the sign itself.

Bender and Flickinger assume that the agreement between the two subjects is syntactic andhence that the two verbs and the two subjects in any tag question must all agree. This view,however, is inconsistent with well known data like (71):

(71) a. Sears is having a sale, aren’t they?b. At least one of us is sure to win, aren’t we?c. The crowd is getting agitated, arent’ they?

Following Oehrle (1987) and Culicover (1992), Kay (2002) argues that the agreement betweenthe two subjects here is semantic in nature, whereas the agreement between each verb and itssubject is syntactic in nature. Notice, however, that in any analysis positing a structure for tagsalong the lines shown in (72), the agreement relation between the two subjects appears to benon-local, i.e. it involves agreement between two elements that are not sisters:


(72) S

SU XARG n ¬ NP e o V¬Sears

VP

is having a sale

V

aren’t

NP ethey

By positing an analysis wherein a clausal sign includes an XARG value reflecting the agreementproperties of the clause’s subject, we make it possible to treat the agreement in tag questionslocally, i.e. via a constraint requiring the relevant identity between the clause’s value for XARGand that of the pronoun in the tag. We develop this analysis further in Chapter 9.

In sum, many researchers have noted that there are phenomena in various languages thatmotivate modifying grammatical theory to allow external access to subjects realized within aclause. These include English ‘copy raising’ (Rogers 1974, Potsdam and Runner 2001, Asudehto appear) illustrated in (73):50

(73)There looks like

¥¦§ ¦¨ there’s going to be a storm*it’s going to rain*Kim’s going to win

© ¦ª¦« .

Assuming, following Pollard and Sag (1994) that there are three subtypes of the type index – ref(referential-index), it (expletive-it-index), and there (expletive-there-index) – contrasts like thesecan be treated simply by associating the relevant looks like construction with the ARG-ST list in(74):

(74) U ARG-ST I NP k , S[XARG I NP k [pro] J ] J VAlso relevant are controlled pronominal subjects in Serbo-Croatian (Zec 1987), HalkomelemSalish (Gerdts and Hukari 2000) and other languages, where control verbs also include the ARG-ST specification in (74). The problems of raising across Polish prepositions (Przepiorkowski1999, Dickinson to appear), and complementizer agreement in Eastern Dutch dialects (Hohle1997) are similar, and submit to similar analysis.

Finally, as discussed further in Chapter 5, there are many English idioms that require refer-ential and agreement identity between a subject and a possessor within an object NP, or whichassign a semantic role to the object’s possessor. These are illustrated in (75)–(76):

(75) a. He k lost [his k /*her r marbles].b. They k kept [their k /*our r cool].

(76) a. That k made [her r hair] stand on end.b. That k tickled [your r fancy].

50Ash Asudeh. 2002. Richard III. In Mary Andronis, Erin Debenport, Anne Pycha and Keiko Yoshimura (eds.), CLS38: The main session. Chicago, IL: Chicago Linguistic Society. Presented April 26, 2002


If an object NP includes information about its (prenominal) possessor in its XARG value, thenan idiomatic verb like lose can be specified as in (77):

(77) U ARG-ST I NP k , NP[XARG I NP k [pro] J ] J VAnd, similarly, a verb like tickle can assign a semantic role to its object’s possessor. In bothcases, all that is required is that the NP’s XARG member be identified with the NP’s possessor,as sketched in (78):

(78) NPU XARG n ¬ NP e o V¬your

N

fancy

All of the phenomena just enumerated provide motivation for the XARG feature we havealready introduced, specified as a property of clausal and NP signs. Note that XARG lists, unlikeVAL lists, do not shrink as larger phrases are constructed. That is, elements are not ‘cancelledoff’ the XARG list. Analyses along these lines have been suggested independently by many ofthe researchers just mentioned;51 our XARG proposal simply synthesizes these ideas within theoverall framework of SBCG. For further discussion of this kind of analysis, in relation to theissue of ‘locality of selection’, see Meurers and Levine (eds.) to appear, and in particular Sag toappear.

As noted above [(44)], we assume that the XARG value of verbal lexemes is a singleton listwhose member is also the first member of the verb’s ARG-ST list and its VAL list, as shown inthe following lexeme licensed by the lexical entry for love:52SHOULD WE ADD

SOMETHINGABOUT PASSIVEAND MIDDLE,WHICH CHANGEVAL?? BUT NOTXARG?? IDISAGREE ABOUTPASS ANDMIDDLE- IAS

51Pollard’s (1994) ERG feature is an early proposal of a noncancelling feature coding a dependency relation, based onunpublished ideas of Andreas Kathol’s. Kiss (1996) introduced a feature for the subject of German verbal clauses andcalled it SUBJECT; this is the feature used by Levine and by Meurers. However, since we also use this in our analysisof NPs to make possessor NPs available for external selection, we have adopted the more neutral term ‘EXTERNAL-ARGUMENT’, which was originally introduced in a similar context by Sag and Pollard (1991).

52The XARG value is a list because it simplifies the formulation of the Head-Complement Construction in the nextsection.


(79) ��

��

stv-lxmFORM I love JSYN

�� CAT

�� XARG I NP[ R�R�R ] JSELECT I+JR#R�R

��MRKG unmk

VAL I NP k [ R�R�R ] , NPr [ R�R#R ] J��

ARG-ST I NP k [ R�R#R ] , NP r [ R#R�R ] JSEM

��INDEX `FRAMES a �� love-fr

ACTOR bUNDGR SIT `

�� c��R�R�R

��

��By contrast, so-called ‘case-marking’ prepositions like to or of have an empty XARG list:

(80) ��

��

FORM I of JSYN

�� CAT

�� prep

XARG I+JSELECT I+J

� ��VAL I NP[ R�R#R ] JMRKG unmk

� ��ARG-ST I NP[ R�R#R ] JSEM U FRAMES I+J VR�R�R

��

��The Head-Complement ConstructionWith these lexical contrasts in place, we can now analyze the different properties of VP and PPswithout proliferating head-complement constructions. We posit the general head-complementconstruction sketched in (81):


(81) Head-Complement Construction (preliminary version):��MTR

��SYN K VAL �

MRKG w Q �DTRS a �� word

SYN�� CAT [XARG � ]VAL � ¢ ® nelistMRKG w

�� c ¢ ®

��What (81) says is that a head-complement construct must involve a (phrasal) mother whoseMARKING value matches that of the first daughter (the head daughter). The first daughter, inaddition, must be followed by all of the valents that it selects, except for the XARG, if there isone. The mother’s VAL value is the head daughter’s XARG list, which will be singleton in thecase of a verb, but empty when the head daughter is a case-marking preposition. A grammar thatincludes this construction licenses constructs like the following:53

(82) ��FORM n loves , Pat oSYN

�� CAT

�� verb

XARG n NP[ p&p�p ] oVF fin

SELECT n�o��

VAL n NP[ p�p�p ] oMRKG unmk

��p�p�p

��

��FORM n loves oSYN

�� CAT

�� verb

XARG n NP[ p�p�p ] oVF fin

SELECT n�o��

VAL n NP[ p&p�p ] , NP[ p�p�p ] oMRKG unmk

��p&p�p

��

��

�� FORM n Pat oSYN

��CAT

�� noun

CASE acc

SELECT n�o�� p�p�p��

53Here, we omit SEMANTICS as well as PHONOLOGY and CONTEXT.


(83) ��FORM n of , Pat oSYN

�� CAT K XARG n�oSELECT n�o Q

VAL n�oMRKG unmk

��p�p�p��

��FORM n of oSYN

�� CAT K XARG n�oSELECT n�o Q

VAL n NP[ p&p�p ] oMRKG unmk

� ��p&p�p� ��

��FORM n Pat oSYN

�� CAT

�� noun

CASE acc

SELECT n�o��

VAL n�oMRKG unmk

� ��p�p�p

� ��

Headed ConstructsThese constructs illustrate another familiar property of headed constructions: that category prop-erties of the head daughter are shared by its mother. This property is guaranteed by a constraintoften referred to as the Head Feature Principle. This principle is an important component of allwork in X-Bar Theory. Our use builds directly on Pollard and Sag’s reformulation (1987, 1994)of the Head Feature Convention of Gazdar et al. 1985. In order to state the Head Feature Princi-ple, we need some way of identifying the head daughter in phrasal constructs. We accommodatethis need by introducing headed-construct as a subtype of phr-cxt:

(84) a. headed-construct (hd-cxt) is an immediate subtype of phr-cxt.b. HD-DTR is used to specify the head daughter of a headed construct; the value of

HD-DTR is of type expression.

The two types of phrasal construct that we have considered thus far – subj-pred-cxt and hd-comp-cxt – are both headed constructs. In the former case, the second daughter (the VP) is thehead daughter; in the latter case, the head is the first daughter (of type word). The final versionof these two constructions is given in (85):


(85) a. Head-Complement Construction (final version):��

hd-cxt

MTR��SYN K VAL �

MRKG � Q �DTRS a w �� word

SYN�� CAT [XARG � ]VAL � ¢ ® nelistMRKG �

� �� c ¢ ®

HD-DTR w

��b. Subject-Predicate Construction (final version):��

hd-cxt

MTR

�� phrase

SYN K VAL I+JMRKG � Q

��DTRS a w , � �� SYN

�� CAT K verbVF finQ

VAL I w JMRKG �

�� c

HD-DTR �

� ��The Head Feature Principle can now be stated succinctly as a type constraint restricting the

well-formedness of headed constructs:

(86) Head Feature Principle:

hd-cxt D �� MTR U SYN [CAT w ] VHD-DTR U SYN [CAT w ] V ��

(The category of a phrase and its head daughter are identical.)

Notice that MRKG and VAL are syntactic features, but are not part of the CAT value, and henceare not covered by the Head Feature Principle (HFP). The resulting analysis is illustrated by thederivation tree in Figure 5, where the effects of the Head Feature Principle are highlighted.

The Head-Functor ConstructionWe now turn to the Head-Functor Construction. We follow the essential insights of Van Eynde(2006a,b), who argues that significant generalizations are missed by analyses based on so-called‘functional categories’.54 In their place, he offers a unified analysis of determiners, markers and

54Other critiques. Newmeyer. Hudson. Who else?


��

��

phrase

FORM n Leslie, loves, Pat oSYN

�� CAT

�� verb

VF fin

XARG n NP[ p&p�p ] oSELECT n¯o

� ��VAL n�oMRKG unmkd

��p�p�p

��

��(Subj-Predicate Cx)

��wordFORM n Leslie oSYN

�� CAT K nounp�p&p Qp�p&p��p�p&p� ��

��

��

phraseFORM n loves, Pat oSYN

�� CAT

�� verb

VF fin

XARG n NP[ p�p�p ] oSELECT n�o

��MRKG unmkd

VAL n NP[ p�p�p ] o��p�p�p

��

��(Head-Comp Cx)

��

��

word

FORM n loves oSYN

�� CAT

�� verb

VF finXARG n NP[ p&p�p ] oSELECT n�o

� ��MRKG unmkd

VAL n NP[ p&p�p ] , NP[ p&p�p ] o� ��p�p�p

��

��

��word

FORM n Pat oSYN

�� CAT K nounp�p�p Qp�p�p��p�p�p��

FIGURE 5 Derivation Tree for Leslie loves Pat


modifiers in terms of a simple, direct combination of a ‘functor’ expression and the head it selects,based on the SELECT feature, discussed in section 1.3.3 above.

All major categories specify values for SELECT in Van Eynde’s theory: nouns, adjectives,adverbs, prepositions, and verbs. For some of these, e.g. finite verbs, the value is I�J . Adjectives,by contrast, select nominal heads and complementizers (whose category assignment Van Eyndedoes not discuss) select verbal heads, as illustrated in (87):

(87) ��FORM I happy JSYN

�� CAT

�� adj

SELECT��SYN K CAT noun

MRKG unmkQ ��

MRKG unmk

��

FORM I that JSYN

�� CAT

��SELECT

��SYN

�� CAT verbVAL I+JMRKG unmk

�� MRKG that

��

Given these lexical specifications, we can formulate the Head-Functor Construction as follows:

(88) Head-Functor Construction:��hd-cxt

MTR��SYN K VAL �

MRKG w Q �DTRS a ��

SYN K CAT [SELECT � ]MRKG w Q � , � c

HD-DTR � [SYN [VAL � ]]

� ��This will allow us to construct both marked clauses and modified nominal phrases, as shown in(89)–(90):


(89) ��FORM n that , Kim , left oSYN

�� CAT

�� verb

SELECT n¯oVF fin

��MRKG that

VAL n�o��p&p�p

� ��

� ��FORM n that oSYN

�� CAT K SELECT n [ ] op&p�p QMRKG thatVAL n¯o

��p�p�p

��

��FORM n Kim , left oSYN

�� CAT

�� verb

SELECT n�oVF finp�p�p

��MRKG unmk

VAL n�o��p�p�p

� ��

� ��(90) ��

��FORM n happy , puppy oSYN

�� CAT

�� noun

SELECT n�op�p�p��

MRKG unmk

VAL n�o� ��p�p�p

��

��FORM n happy oSYN

�� CAT

�� adj

SELECT n [ p&p�p ] op�p&p��

MRKG unmkp&p�p��p&p�p

��

��FORM n puppy oSYN

�� CAT

�� nounSELECT n�op�p�p

��MRKG unmkVAL n�o

��p�p�p

� ��

Note that in each of these constructs, the mother’s SELECT specification is inherited from thehead daughter, in accordance with the Head Feature Principle ((86) above).


1.5 Conclusion?

Sign-Based Construction Grammarlingo.stanford.edu/sag/SBCG/sbcg.pdf · Sign-Based Construction Grammar Charles Fillmore U.C. Berkeley ... 1.3.5 CONTEXT 12 ... that feature structures

Documents