A Minimalist Implementation of Verb Subcategorization

Traditional accounts of verb subcategorization, from the classic work of Fillmore on, require either a considerable number of syntactic rules to account for diverse sentence constructions, including crosslanguage variation, or else complex linking rules mapping the thematic roles of semantic event templates with possible syntactic forms. In this paper we exhibit a third approach: we implement, via an explicit parser and lexicon, the incorporation theory of Hale and Keyser (1993, 1998) to systematically cover most patterns in English Verb Classes and Alternations (Levin 1993), typically using only 1 or 2 lexical entries per verb to subsume a large number of syntactic constructions and also most information typically contained in semantic event templates, and, further, replacing the notion of “thematic roles” with precise structural configurations. The implemented parser uses the merge and move operations formalized by Stabler (1997) in the minimalist framework of Chomsky (2001). As a side benefit, we extend the minimalist recognizer of Harkema (2000) to a full parsing implementation. We summarize the current compactness and coverage of our account and provide this minimalist lexicon and parser online at http://web.mit.edu/niyogi/www/minimal.htm 1 The Problem of Verb Subcategorization Why do certain verbs undergo particular certain alternations and not others? On some accounts, e.g. Levin (1993), referred to hereafter as EVCA, alternations provide insight into verb subcategorization and hence hooks to parsing, cross-language variation, machine translation, and class based verb learning. However, fully implemented accounts of the phenomena remains an open problem, with at least three alternative models, shown in Figure 1. Accounts may be solely descriptive – for example, classifying verbs as having an intransitive, a transitive, and/or ditransitive form, as is familiar. Traditional computational accounts (see 1) map these forms into individual grammar rules, (perhaps by macro expansion-like techniques) adding as many rules as necessary to account for naturally’ occurring constructions (wh-movement, passive forms, etc.) For each grammatical rule, a separate semantic decomposition is required, typically labeling component phrases with one of several “thematic roles.” A richer account provided by lexical semantics (see 2), exemplified in Jackendoff (1983, 1990) and Rappaport Hovav and Levin (1998), is one that hypothesizes semantic templates, but requires linking rules mapping syntactic frames with semantic templates governed by a particular verb. Often these semantic templates are constructed in an ad hoc manner, and the corresponding linking rules are consquently a collection of difficult-toimplement heuristics. In this paper we implement a rather different formalism (Hale and Keyser’s Incorporation theory, see 3), wherein fewer lexical entries govern syntactic and semantic behavior, with no appeal to thematic roles or complex linking rules. 0. Verb Subcategorization Phenomena * Bob put. Butter was put on the bread. * Bob put butter. What was put on the bread? Bob put butter on the bread. Where was the butter put? 1. Traditional Account VP → V0 NP PPloc V0 → put VP → was VPass VPass → V0 PPloc VP/NP → V0 NP/NP PPloc VP/NP → V0 NP PPloc/NP PPloc → Ploc NP Ploc → on | in | ... PPloc/NP → Ploc NP/NP Exhaustive modelling with a considerable number of grammatical rules. Semantics separate, otherwise unspecified. 2. Lexical Semantics Account   put V NPjPPk CAUSE ( [BOB]i , GO ( [BUTTER]j , TO ( [BREAD]k )))   Syntax handled by numerous argument-fusing “linking rules”, typically difficult to formalize. Semantic templates mirror alternation patterns, but are ad-hocly constructed. 3. Minimalist/Incorporation Account /put/ =ploc =d vcause (λ(=ploc) (λ(=d) (=ploc =d))) /on/ =d +k ploc (λ(=d) (λ(x) ((go x) (path self =d)))) // >vcause +k =d pred (λ(>vcause) (λ(=d) ((cause >vcause) =d))) /-ed/ >pred ++k t (λ(>pred) (tense >pred past)) Small number of lexical entries handle all syntactic phenomena. Semantics directly encoded in lexical entry. Entries structurally governed by small number of rules, specifying how N/A/P are related. Figure 1: Three Different Accounts of Verb Subcategorization 2 Incorporation Theory At the heart of our new contribution to modeling verb subcategorization is the marriage of Hale and Keyser’s (1993, 1998) argument structure theory with Stabler’s (1997) ‘minimalist’ structure building rules. In the Hale and Keyser’s theory, using the terminology of X-bar syntax, a particular head (labeled X), may or may or may not take a complement (labeled Y) and may or may not project a specifier (labeled S), resulting in 4 possible structural configurations: X © H X Y X ©© H H S X © H X Y α ©© H H S α © H α X X (a) -subj, +comp (V) (b) +subj, +comp (P) (c) +subj, -comp (A) (d) -subj, -comp (N) Figure 2: Four fundamental primitives in Hale and Keyser’s incorporation theory The combinatorial possibilities of incorporation with X=V, A, N, P heads, plus ‘head movement’, is designed to yield the space of possible syntactic argument structure configurations, presumably across all languages. Notions of agent, patient, instrument, theme, goal, etc. are not ‘primitives’, but are derived from positions in structural configurations. In English (but not necessarily in all languages), (a) the category V takes a complement but projects no specifier; (b) the category P takes both a complement and projects a specifier; (c) the category A takes no complement but projects a specifier; (d) the category N takes neither complement nor specifier. A particular verbal entry, being of category V, may incorporate one or more of these structures as its complement, as shown in Figure 3: • Nouns incorporated directly into a verbal entry yield structures such as (a): no subject is projected by the N. The phonetic material of the noun head incorporates (undergoes head movement) into the phonetic material of the verb head, which itself may undergo further movement. Verbs such as these are intransitive by nature, generating, e.g., /The light glow -ed/ but */Bob glow -ed the light/. This argument structure typifies purely internally caused processes.