Fully Lexicalising CCGbank with Hat Categories

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1212–1221, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Fully Lexicalising CCGbank with Hat Categories Matthew Honnibal and James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia {mhonn,james}@it.usyd.edu.au Abstract We introduce an extension to CCG that al- lows form and function to be represented simultaneously, reducing the proliferation of modifier categories seen in standard CCG analyses. We can then remove the non-combinatory rules CCGbank uses to address this prob- lem, producing a grammar that is fully lex- icalised and far less ambiguous. There are intrinsic benefits to full lexi- calisation, such as semantic transparency and simpler domain adaptation. The clear- est advantage is a 52-88% improvement in parse speeds, which comes with only a small reduction in accuracy. 1 Introduction Deep grammars return parses that map transpar- ently to semantic analyses, allowing information extraction systems to deal directly with content representations. Usually, this mapping is lexically specified, by linking lexical entries to semantic analyses. This property, lexicalisation, is central to some of the linguistic theories behind deep gram- mars, particularly Combinatory Categorial Gram- mar (Steedman, 2000) and Lexicalised Tree Ad- joining Grammar (Joshi, 1999). Lexicalisation can also help deep grammars achieve satisfactory parse times. Lexicalised grammars use few rules, which simply manipu- late the lexical categories. The categories can be quickly assigned in a supertagging pre-process, dramatically reducing the search space the parser must explore (Bangalore and Joshi, 1999). Combinatory Categorial Grammar (CCG) is well suited to this strategy, and Clark and Curran (2007) have highlighted the division of labour be- tween the parser and the supertagger as one of the critical aspects of their approach to statistical CCG parsing. In their system, the division is managed with parameters that control how many categories the parser’s chart is seeded with. But even if the parser is only given one category per word, it still has a lot of freedom — because the grammar it uses is not fully lexicalised. In a fully lexicalised CCG grammar, modifier categories refer to the category of their head. This category does not necessarily represent the head’s constituent type. For instance, the category of an adverb like still depends on whether it is modify- ing a predicate verb (1), or a clausal adjunct (2): (1) The lion NP was VP/VP lying VP still VP\VP (2) The lion NP waited, VP lying VP\VP still (VP\VP)\(VP\VP) Analyses like these are problematic because the training data is unlikely to include examples of each word in every syntactic environment that re- quires a new category. Hockenmaier and Steed- man’s (2007) solution was to add category spe- cific phrase-structure rules to the grammar, which disrupts the linguistic principles of the formalism, and introduces over-generation and ambiguity as shown in Figure 1. This paper proposes a new way to balance lex- ical and grammatical ambiguity in CCG. We in- troduce an extension to the formalism that allows type-changing rules to be lexically specified. The …

Content truncated to 500 words