Functional Combinatory Categorial Grammar

Abstract

Functional Combinatory Categorial Grammar (FCCG) advances the field of combinatory categorial grammars by enabling semantic dependencies to be determined directly from the syntactic derivation under the action of a small set of extraction rules. Predicates are extracted composably and can be used to apply semantic constraints during parsing. The approach is an alternative to that of classical CCG which requires (i) mapping from categories to lambda expressions, (ii) a set of semantic transformation rules for unary combination, and (iii) an explicit β-reduction stage. GFCCG, a generalised form of the grammar, has previously been applied to situation assessment (McMichael, Jarrad, & Williams, 2006). In FCCG, combinators are largely distinguished by their semantic purpose. Unary combination is only used for preterminal-terminal transitions. Replacing unary type-raising and type-changing by their binary counterparts R and P tends to reduce parse ambiguity. Four other binary combinators are introduced to model various semantic phenomena: functional composition (F), modification (M), apposition (A) and copular modification (Q). Of the combinators of classical CCG, only binary coordination (&) is retained. The category is the natural feature structure for CCG, and we show how it may be extended to host semantic and parsing-related features compactly. The grammar is demonstrated by extracting an FCCG-annotated corpus from the Penn treebank using the com-binator calculator described in (Foreman & McMichael, 2004). Only fifty categories and 140 productions cover 99.7% of the extracted corpus, a substantially more efficient representation than previous conversions. We adopt a factored conditional statistical model, and provide an efficient iterative learning algorithm involving direct feedback of parsing errors that does not require enu-meration of the parse forest. While the best parser performances on data derived from the Penn treebank have come from fully lexicalized parsers, we have been motivated by the need to provide good coverage outside that relatively narrow training domain. We have therefore used a semilexicalised feature model, which does not, for example, contain bilexicalised features. Spurious ambiguity is controlled via probabilistic scoring coupled with agenda-based A* parsing. The parser uses an auxiliary queue prioritized by multi-tag probabilities for terminal node introduction. The parser's syntactic dependency F-score of 79.2% compares well with the best labelled syntactic dependency F-score for a fully lexicalised CCG parser tested on data consistent with the training domain of 84.6% (Clark & Curran, 2004b). The parser obtained a semantic extraction F-score of 83.1%.