What is Construction Grammar?

Construction Grammar
Paul Kay, U. C. Berkeley
(Entry for Handbook of Pragmatics: manual. edited by Jef Verschueren, Jan-Ola Ostman, Jan Blommaert. Amsterdam; Philadelphia: J. Benjamins, 1995.)

Construction Grammar (CG) is a non-modular, generative, non-derivational, monostratal, unification-based grammatical approach, which aims at full coverage of the facts of any language under study without loss of linguistic generalizations, within and across languages.
(Explanations of each of these terms are given below.)

Non-modularity

CG adopts from traditional grammar the idea that a grammar is composed of conventional associations of form and meaning, that is, grammatical constructions. The non-modular character of CG involves specifically the treatment of form and meaning as part of each grammatical element, i.e., rule or construction, rather than in separate components of the grammar. In this regard, CG is like many other contemporary approaches, including, among others, Montague Grammar, Categorial Grammar, Generalized Phrase Structure Grammar, and Head Driven Phrase Structure Grammar. CG and similar approaches contrast in this respect with the Government and Binding/Principles and Parameters (GB) approach.

Among current non-modular approaches to grammar, CG places great emphasis on the fact that probably any of the kinds of information that have been called 'pragmatic' by linguists may be conventionally associated with a particular linguistic form and therefore constitute part of a rule (construction) of a grammar. Examples (1-5) illustrate constructions of English which conventionally associate special pragmatic forces or effects with specific morpho-syntactic structures.

(1) Sit down.
(2) Him help an enemy?!
(3) Watch it not rain [now that I've bought an umbrella].
(4) What's it doing snowing in August?
(5) in her own right

Some of these forces or effects are very well known, such as the imperative force illustrated in (1). Others deserve careful study and description, for example, the incredulity expressed in (2), the conjuring of fate expressed in (3), the judgment that something is amiss conveyed by (4), or the complex presuppositional background required by (5), which includes not only an antecedent for her but also (very roughly) a particular accomplishment predicated of this person plus a parallel accomplishment predicated of another person, as well as a complex evaluative frame in which all these elements are set. From the CG perspective, the interest of examples such as (1-5) is that the pragmatic forces or effects resulting from utterances of such expressions are conveyed according to conventions of language rather than by a process of conversational reasoning and so must be accounted for by the grammar.

Generative Character of CG

Construction Grammar takes as a major goal the provision of an explicit account of the form and meaning of the sentences of each language, including their correct structures. CG is thus a generative approach in the sense of that term originally proposed, and later abandoned, by Chomsky (Gazdar, et al. 1985:6). (This traditional acceptation for the term generative (grammar) is not to be confused with a more recent use of that term to denote the extended family of GB-related approaches, few or none of which are generative in the original sense.) As a type of generative grammar, CG is also distinguishable from some current views of language which characterize themselves as cognitive and which, like GB, eschew the original goals of generative grammar.

Non-derivational, Monostratal and Unification-based Properties of CG

In its representational system, CG makes extensive use of feature structures and of the operation of unification of feature structures, or structure-sharing, for the representation of dependencies and co-construal. For example, subject-verb agreement in English consists in (1) part of the representation of each finite verb word containing within it a feature structure denoting the number of its subject and (2) each construction which provides a verb or verb-headed constituent with a subject unifying (Shieber 1986, Carpenter 1992) the appropriate substructures of the subject constituent and the verb word. Unification of feature structures is also the key formal ingredient in the constructions effecting coinstantiation (equi and raising) and in the various left isolation (extraction) phenomena, to be discussed further below.

There are no derivations in CG. A candidate sentence is licensed as a sentence of the language if and only if there exists in the grammar of that language a set of constructions which can be combined in such a way as to produce a representation of that sentence. Consider an example like

(6) What he saw was an elephant.

The constituent an elephant is licensed by a construction which places a constituent which can play a determining role, such as an article, a demonstrative or a possessive noun phrase, before a nominal constituent not so determined. The constituent what he saw is licensed by a construction which structures headless relative clauses in general (whose properties are briefly noted below). The constituent was an elephant is licensed by the English verb phrase construction, which provides for a lexical verb to be followed by exactly as many constituents as are necessary to satisfy its non-subject valence requirements (including adjuncts, which are treated in CG as augmentations of minimal valences). The full sentence of (6) is licensed by the Subject-Predicate construction. This construction licenses as a sentence of English a structure consisting of a finite verb phrase preceded by a constituent capable of serving as its subject.

The objects of the representational system of CG are phrase structure trees (possibly incomplete in not having all linear precedence relations specified) whose nodes correspond to feature structures. Combination of structures consists of the superimposition of trees, subject to certain geometrical restrictions, with the provision that those feature structures which are brought together by the superimposition unify. Lexical items are treated as constructions with minimal constituent structures, consisting of a tree with a single node. Processes such as those linking grammatical functions with semantic roles are represented as templatic lexical constructions which unify with minimal lexical entries. Derivational and inflectional morphology are treated with constructions that simply unify with minimal entries in the case of "zero" morphology and with constructions providing a richer constituent structure skeleton in the case of phonologically expressed morphology.

Despite the unavoidably processual language in which the above description is cast, the formal system of CG is always in spirit, and almost always in practice, declarative and monotonic. That is, CG takes the notion that a grammar can be represented as a declarative and monotonic system to be a formal hypothesis worth pursuing as far as possible. A grammar is assumed to consist of a single module, consisting of a repertory of patterns (i.e., constructions). A subset of these is combined on each occasion of production or interpretation of a sentence, but the order in which they are combined is not determined by the grammar and need not be the same on different occasions of producing or interpreting sentence tokens of a given type. Under such a view of grammatical form, the commonplace virtuosity of normal first language acquisition does not conduce to the conclusion that the child's innate endowment of specific linguistic structure is enormously rich, as insisted upon by the usual Chomskian "poverty of the stimulus" argument. Positing a monotonic and declarative grammar of constructions -- as against, for example, the rich architecture of GB -- encourages substituting for the dramatic claim that human infants are endowed innately with extensive knowledge of linguistic structure the more cautious hypothesis that human infants are endowed innately with a special ability to induce linguistic structures from linguistic data, that is, to acquire linguistic constructions.

Full Coverage

The empirical commitment of construction grammar is that grammatical theory must in principle account for the totality of facts of any language, not recognizing a priori any theoretically privileged set of core grammatical phenomena. Of course in real life one can never describe with precision more than a small corner of any language, but what counts theoretically is whether or not the grammarian's methodological stance permits, or even dictates, that certain facts be ignored as peripheral to the UG core. The CG position is that linguists do not now have a sufficiently firm knowledge of universal grammar to sensibly exclude from the concern of theoretical, or descriptive, linguistics a significant portion of the data of any language.

Moreover, examination of the more marked constructions sheds a special light on how a grammar must be put together. First, in place of a a grammar divided neatly between idioms on the one hand and productive rules on the other, what the data appear to demand is a cline of constructions, from the relatively productive to the relatively frozen. Looking at the grammar of English, for example, from the point of view of constructions like those exemplified in (1-5), one finds no natural joint at which to carve a core/periphery distinction.

Secondly, accepting the marked constructions as legitimate data imposes on a grammatical framework the necessity to provide for a sentence which illustrates both relatively productive and relatively unproductive constructions an integrated analysis, hence the need for a representational system which treats productive and non-productive constructions as the same kind of formal object. The monostratal, monotonic and declarative framework described briefly above, which provides for the combination of structural representations that themselves may express linguistic information regarding any aspect of linguistic form or content, including of course "pragmatic" information, seems well suited to this task.

Consider briefly the construction illustrated in (4). In first asking what aspects of a sentence like (4) are attributable to this construction, we note that the clausal constituent structure of a main clause question can not be part of this construction, because the construction can appear in uninverted, embedded questions (Pullum 1973).

(7) I wonder what the flags are doing at half mast

Other morpho-syntactic peculiarities (some also noted by Pullum) include:
(a) the lexeme do in the lexical form doing is required (*I wonder what act the flags are performing, *I wonder what the flags did at half mast yesterday.),
(b) the form doing in this construction is not a vehicle of progressive aspect (What is he doing knowing the answer?, Cf. *He is knowing the answer),
(c) the construction does not accept else as a modifier of what (what (*else) are you doing eating cold pizza?),
(d) the main verb must be be (*What does that dust keep doing on the desk?, Cf. What is that dust still doing on the desk?),
(e) neither be nor doing can be negated, although the complement of doing may be (*What aren't my brushes doing in the right place?, *What are my brushes not doing in the right place?, Cf. What are my brushes doing not soaking in water?)

Kay and Fillmore (1994) call this construction the What's X Doing Y? construction (WXDY) and argue for a syntactic analysis along the following lines. The WXDY construction consists of a particular lexical entry specifying the lexeme be, and stipulating a valence requirement for a complement VP headed by a special, copula-like verb doing which itself has a valence requirement for a complement belonging to the wide (and unexplained) range of syntactic forms which elsewhere occur as secondary predicates. These notably include such predicates as with shoes on and without shoes on, which cannot complement ordinary copula be (*He is without shoes on). The subject of the secondary predicate undergoes (the CG version of) raising twice to become the subject of be, wherever that constituent occurs, i.e., inverted or not. A special interrogative what, deprived of reference, is specified to be the object valence requirement of doing; this constituent appears according to where the rest of the grammar of the sentence causes it to appear, possibly illustrating a long distance dependency as in

(7) What did you say his books were doing in your locker?

The point of this somewhat extended example is to illustrate how a CG analysis of a construction must show which aspects of sentences containing this construction are due to the construction itself and which are due to other constructions with which it combines (such as those providing for main clause questions, embedded questions, ordinary verb phrases, and so on) and precisely how these various grammatical objects fit together in the sentences illustrating the construction. The point will have been made if the reader notes that no part of the constituent structure of any sentence containing WXDY is directly contributed by WXDY itself. All aspects of the constituent structure in any sentence containing this construction come from other constructions which have been combined with it.

Two general points of construction grammar may be illustrated by considering the notional content of the WXDY construction. First, it is clear that sentences like (4) and (7) convey someone's judgment that something is amiss. In these two cases, and in many others, that judgment is attributed by an utterance of a WXDY sentence to the speaker of that sentence, but such need not be the case. For example, in plotting a surprise party someone might say

(8) But what will we tell her the lights are doing off?

In the utterance of (8) the implicit judge of something's being amiss is neither the speaker nor the addressee but the person to be surprised. (Fillmore et al. 1988: 533 make a similar observation with regard to the LET ALONE construction.) Sorting out the exact pragmatic and other notional implications of constructions specifying imputed judgments can be a delicate affair.

Secondly, when a grammatical construction with a particular pragmatic force or effect is proposed, it is important to demonstrate that the pragmatic value is indeed conveyed by a convention of grammar and not by the application of some form of conversational reasoning to the prior and independent communication of a proposition innocent of the pragmatic value. Many WXDY sentences are in fact ambiguous between the constructional "non-canonicity" reading and a pragmatically innocent propositional reading which might in context trigger a process of conversational reasoning leading up to the non-canonicity judgment. Thus, if you catch me going through your papers and say

(9) What are you doing in my office?

it could be argued that you have simply asked a normal question whose answer is obvious and thereby implicated something further. But this kind of analysis will not work for examples such as (4,7, and 8), where there is no corresponding literal question to be asked (What are the lights doing off? *Resting.). Moreover, the particular formal properties of WXDY sentences (the requirements regarding doing and be, and so on) argue compellingly that the peculiar pragmatic force of expressing a judgment of non-canonicity is here tied to the particular grammatical form. Together the associated form and content constitute a single conventional unit of the grammar of English, a grammatical construction of English. (Griceans may consider the seldom exploited notion of conventional implicature to be relevant here.)

Generalizations

Construction Grammar is devoted to the extraction of all the generalizations potentially available to the speaker of a language. No claim is intended that the internal representation of the language in the mind of each speaker contains every generalization inherent in the data. There seems no evidence available to support such a strong claim outside of cases where the data covered by the generalization are infinite and so cannot be learned or stored individually. Nevertheless, a proper charge to the grammarian would seem to be that the grammar express every generalization available to the native speaker. Variability among speakers with regard to which speakers extract which generalizations are appropriate questions for research in psycholinguistics and variation studies. It is the grammarian's job to lay out the initial possibilities by identifying the full range of candidates.

An important formal device of CG for the expression of generalizations is that of inheritance. This is exemplified by the Left Isolation construction, which is the abstract construction corresponding in CG to the property seen in common in the following examples.

(10)
a Where do you think he said she was hiding?
b [the man] who she sold the car to
c [John,] who she sold the car to,
d [I wonder] what he cooked
e [I ate] what he cooked
f You, I can trust
g A black eye you'll get
h [And] save him she did
i The longer you sleep the louder you snore

Example (10)a illustrates a matrix, non-subject, wh-question, (10)b a restrictive relative clause, (10)c a non-restrictive relative clause, (10)d an embedded question, (10)e a headless relative clause (Recall what he saw in (6)), (10)f a topicalized structure, (10)g a focus-fronted or Y-moved structure, (10)h a VP-fronted structure, (10)i a comparative conditional structure.

As the names suggest, the type of structure illustrated by each example possesses a distinct array of properties of form, interpretation or both. That is, a construction grammar of English will contain distinct constructions for wh-questions, restrictive relative clauses, non-restrictive relative clauses, embedded questions, headless relative clauses, topicalization, and so on. But also, there is a common set of properties shared among of these patterns (and others), and the grammar will express this generalization as an abstract construction which is inherited by the several constructions licensing the relevant parts of the examples in (10). If distinct constructions X and Y inherit construction A, then X and Y each share all the properties of A and each of X and Y possesses some additional properties it does not share with the other (or with A).

We will call the set of properties shared in common by the constructions illustrated in (10) the Left Isolation (LI) construction. (This corresponds roughly to phenomena known in the generative tradition by the label "extraction".) The abstract LI construction consists of two sisters; the left sister represents a valence requirement of some predicator occurring in the right sister. (This property is defined formally by giving precise expression to the intuitive notion "X is a valence element of a valence element ... of a valence element of Y". Kay and Fillmore 1994, Kay 1994.) Additionally, the right sister of the Left Isolation construction is a maximal verb-headed constituent, that is, either a clause or a VP. (Some additional details are omitted here.)

The LI construction and its inheritance by the less abstract constructions involved in each example of (10) represents but one instance out of many that could be cited as examples of inheritance in CG. Inheritance is also used extensively in lexical constructions, for example in representing the alternate assignments of grammatical functions to semantic arguments contained in linking constructions, fully formed verb words inheriting both linking template constructions and morphological constructions (Koenig and Jurafsky 1994).

The notation of CG itself contains a set of commitments to the form in which linguistic knowledge is represented and so makes universal commitments regarding a variety of the formal properties of natural languages. The relevance of phrase structure, a certain array of syntactic categories, an array of syntactic and semantic relational categories such as grammatical functions and semantic roles, and the role of inheritance in the grammars of all languages are a few of the more important universal commitments of this type made by Construction Grammar. Some of the formal devices which CG does not make available to itself also constitute universal claims. These include the absence of empty of categories and the absence of movement. The structures CG permits itself are all quite "surfacey" and concrete, except as inheritance allows abstraction over structures of this type.


Back to Ling 220 home page