Head-Driven Phrase Structure Grammar

Head-driven phrase structure grammar (HPSG) is a lexicalist, constraint-based family of theories of GENERATIVE GRAMMAR to which Sag and Wasow (1998) offers an elementary introduction. Two assumptions underlie the theory of head-driven phrase structure grammars. The first is that languages are systems of types of linguistic objects like word, phrase, clause, person, index, form-type, content, rather than collections of sentences. The other is that grammars are best represented as process-neutral systems of declarative constraints (as opposed to constraints defined in terms of operations on objects as in transformational grammar). Representations are structurally uniform: all objects of a particular type have all and only the attributes defined for that type. What attributes are defined for an object type is restricted empirically, not by a priori conditions; they cover phonological, semantic, structural, contextual, formal and selectional (subcategorizational) properties.

A grammar (and for that matter, a theory of universal grammar) is thus seen as consisting of an inheritance hierarchy of such types (an "is-a" hierarchy similar to familiar semantic networks of the sort that have "creature" as a root and progressively more specific nodes on a branch leading to a particular canary "Tweety"). The types are interrelated in two ways. First, some types are defined in terms of other types. Second, the hierarchy allows for multiple inheritance, in that linguistic objects can belong to multiple categories at the same time, just as other conceptual objects do. The constraints in the linguistic hierarchy are all local, so that well-formedness is determined exclusively with reference to a given structure, and not by comparison to any other candidate structures. The LEXICON is a rich subhierarchy within the larger hierarchy constituting the grammar. Having declarative constraints on a hierarchy of interrelated types of linguistic objects is seen as enabling an account of language processing which is incremental and pervasively integrative. Thus, as long as information about grammatical number is consistent, it does not matter whether it comes from a verb or its subject, as shown by the fact that (1-3) are acceptable, whereas (4) is not.

  1. The dogs slept in the barn.
  2. The sheep which was mine stayed in the pen.
  3. The sheep which stayed in the pen were mine.
  4. *The sheep which was mine are in the pen.

Linguistic objects are modeled as feature structures. Feature structures are complete specifications of values for all the attributes that are appropriate for the particular sort of object that they model, and they are the entities constrained by the grammar. Feature structure descriptions describe classes of feature structures, by means of familiar attribute-and-value matrices (AVMs) that (partially) describe them. A partial description constrains all the members of whatever class of feature structures it describes, while a total description is a constraint that limits the class to a single member. For the most part, grammar specification deals with generalizations over classes of objects like words and phrase-types, and therefore with (partial) feature structure descriptions. Feature-based unification grammar formalisms like HPSG are thus conceptually lean and computationally tractable, and are being used in increasing numbers of NATURAL LANGUAGE PROCESSING systems.

A feature's value is of one of four possible types: atom, feature structure, set of feature structures, or list of feature structures. (Set values are represented as sequences within curly brackets: SLASH , . The empty set is denoted: { }while { [ ] } denotes a singleton set. List values are represented as sequences within angled brackets: COMPS < NP, VP[inf ] >. The empty list is denoted: < >, and < [ ] > denotes a singleton list.) Values that are not specified in a feature-structure description are still constrained to be among the legitimate values for the features that the constraints on the types to which it belongs require.

Like other linguistic objects, categories that figure in the SYNTAX have rich internal structure and constituency descriptions. But HPSG is a "WYSIWYG" theory; empty categories are avoided rather than exploited.

The general outlines of the HPSG approach to constituent order derive from the theory of linear precedence rules sketched in GPSG (Gazdar and Pullum 1981; Gazdar et al. 1985), and discussed at some length in Pollard and Sag (1987). As in GPSG, so-called free word order (i.e., free phrase order) is a consequence of not constraining the order of constituents at all. (Genuinely free word order, where (any) words of one phrase can precede (any) words of any other phrase requires a word-order function that allows constituents of one phrase to be recursively interleaved with constituents of another; see Gazdar and Pullum 1981; Pollard and Sag 1987; Dowty 1996; Reape 1994).

As grammar-writing research on a number of languages (especially notably, German and French) has made abundantly clear, word order constraints are not always compatible with the semantic and syntactic evidence for constituency, and the exact form of the resolution to this dilemma constitutes a lively topic in current research.

Constraints on phrase types project meanings, subcategorization requirements, and head properties from subconstituents. The HEAD-feature principle, for example, represented in figure 1, constrains HEAD properties of a phrase (i.e., category information like person, number, case, inflection) to be the same as that of its head daughter.

Figure 1

Constraints on phrase types also provide COMPOSITIONALITY in the semantics by specifying how the semantics of a phrase type is a function of the semantics of its daughter constituents.

Equi and raising structures (like Kim tried to run and Kim seemed to run, respectively) are both projections of heads that subcategorize for an unsaturated predicative complement, and have the same sorts of constituent structure. Equi verbs like try, however, systematically assign one more semantic role than raising verbs like seem do. Pollard and Sag (1994) represent this difference by saying that an equi verb subcategorizes for an NP with a referential index (i.e., one that is not an expletive), which is the same as the index of the subject element that its complement subcategorizes for, and assigns a semantic role to that index, whereas a raising verb just subcategorizes for whatever its infinitive VP complement subcategorizes for, and assigns no semantic role to the index of that element.

The general outlines of the HPSG treatment of unbounded extractions (WH-MOVEMENT) follow the three-part strategy developed in GPSG (Gazdar 1981; Gazdar et al. 1985). An extra constituent is licensed just in case it matches a missing constituent. Something must ensure that the missing constituent is missing. The correspondence between the gap and the extra constituent (the filler) is recorded via constraints on local (i.e., depth one) constituency relations over an indefinitely large array of structure.

In HPSG, the extra constituent is licensed in strong (topicalization -type) extractions by the schema or sort declaration that defines head-filler clauses (topicalization structures), and for weak extraction phenomena such as tough-constructions, by subcategorization and sort specifications that require a complement daughter to not be lexically realized. Gaps (or traces) are licensed in phrases by constraints or rules that allow dependents to be unrealized when the lexical head that selects them inherits information that a matching element should be missing. As in GPSG, "a linked series of local mother-daughter feature correspondences" (Gazdar et al. 1985: 138), embodied as constraints on phrase-types, entail that the extra constituent and the missing constituent match.

The HPSG account of the binding of indexical elements like her and themselves is stated in terms of the relative obliqueness of the GRAMMATICAL RELATIONS of the indexical and its antecedent relative to a predicate. Considering its nonconfigurational approach, the HPSG binding theory nonetheless resembles familiar configurational accounts:

A. A locally-commanded anaphor must be locally o-bound.
B. A personal pronoun must be locally o-free.
C. A non-pronoun must be o-free.

However, it differs crucially from typical configurational accounts in that it has an inherently narrower scope. Principle A does not constrain all anaphors to be locally o-bound (coindexed to something before them on a predicate's argument-structure list); it constrains only those that are locally o-commanded (i.e., the ones that are noninitial on the list). This makes strong, vulnerable, and apparently correct claims. First, pronouns that are initial elements on argument-structure lists are unconstrained -- free to be anaphors, coindexed to anything, and vacuously satisfying principle A, or to be pronouns, substantively satisfying principle B. Thus, the theory predicts that phrases in these "exempt" conditions, which are coindexed to anything anywhere in a higher clause, or even outside the sentence altogether, can be either anaphors or pronouns. This is correct; the reflexive pronouns that contradict the naive versions of principle A are generally replaceable with personal pronouns with the same reference.

Unification-based, declarative models of grammar like HPSG are attractive for natural language processing applications (e.g., as interfaces to expert systems) precisely because they are nondirectional and suited to the construction of application-neutral systems serving NATURAL LANGUAGE GENERATION as well as parsing and interpretation.

See also

Additional links

-- Georgia M. Green


Dowty, D. (1996). Towards a minimalist theory of syntactic structure. In H. Bunt and A. van Horck, Eds., Discontinuous Constituency. Berlin: Mouton de Gruyter.

Gazdar, G. (1981). Unbounded dependencies and coordinate structure. Linguistic Inquiry 12:155-184.

Gazdar, G., and G. K. Pullum. (1981). Subcategorization, constituent order, and the notion "head." In M. Moortgat, H. v. D. Hulst, and T. Hoekstra, Eds., The Scope of Lexical Rules. Dordrecht: Foris, pp. 107-123.

Gazdar, G., E. Klein, G. K. Pullum, and I. A. Sag. (1985). Generalized Phrase Structure Grammar. Cambridge, MA: Harvard University Press.

Pollard, C., and I. Sag. (1987). Information-based Syntax and Semantics, vol. 1. Stanford: CSLI.

Pollard, C., and I. Sag. (1994). Head-driven Phrase Structure Grammar. Chicago: University of Chicago Press.

Reape, M. (1994). Domain union and word order variation in German. In J. Nerbonne, K. Netter, and C. Pollard, Eds., German in Head-driven Phrase Structure Grammar. CSLI Lecture Notes No. 46. Stanford: CSLI.

Sag, I. A., and T. Wasow. (1998). Syntactic Theory: A Formal Introduction. Stanford: CSLI.

Further Readings

Carpenter, B. (1992). The logic of typed feature structures. Cambridge Tracts in Theoretical Computer Science 32. New York: Cambridge University Press.

Copestake, A., D. Flickinger, R. Malouf, S. Riehemann, and I. A. Sag. (1995). Translation using Minimal Recursion Semantics. Proceedings of the Sixth International Conference on Theoretical and Methodological Issues in Machine Translation (TMI-95) Leuven, Belgium.

Kay, M., J. M. Gawron, and P. Norvig. (1994). Verbmobil: A Translation System for Face-to-Face Dialog. CSLI Lecture Notes No. 33. Stanford: CSLI.

Lappin, S., and H. Gregory. (1997). A computational model of ellipsis resolution. Master's thesis, School of Oriental and African Studies, University of London. Available from website http://semantics.soas.ac.uk/ellip/.

Meurers, W. D., and G. Minnen. (1995). A computational treatment of HPSG lexical rules as covariation in lexical entries. Proceedings of the Fifth International Workshop on Natural Language Understanding and Logic Programming. Lisbon, Portugal.

Pollard, C. (1996). The nature of constraint-based grammar. Talk presented at Pacific Asia Conference on Language, Information, and Computation. Seoul, Korea: Kyung Hee University.

Pollard, C., and D. Moshier. (1990). Unifying partial descriptions of sets. Information, Language and Cognition. Vancouver Studies in Cognitive Science, vol. 1. Vancouver: University of British Columbia Press, pp. 285-322.

Shieber, Stuart. (1986). An Introduction to Unification-based Approaches to Grammar. CSLI Lecture Notes Series. Stanford: Center for the Study of Language and Information. (Distributed by University of Chicago Press.)