Extending CCGbank with Quotes and Multi-modal CCG

Extending CCGbank with quotes and multi-modal CCG Daniel Tse and James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia {dtse6695, james}@it.usyd.edu.au Abstract CCGbank is an automatic conversion of the Penn Treebank to Combinatory Categorial Grammar (CCG). We present two exten- sions to CCGbank which involve manipu- lating its derivation and category structure. We discuss approaches for the automatic re-insertion of removed quote symbols and evaluate their impact on the performance of the C&C CCG parser. We also analyse CCGbank to extract a multi-modal CCG lex- icon, which will allow the removal of hard- coded language-specific constraints from the C&C parser, granting benefits to parsing speed and accuracy. 1 Introduction Combinatory Categorial Grammar (Steedman, 2000) is a powerful lexicalised grammar formalism. In CCG, the combination of categories using a small set of combinatory rules allows a parser to simultaneously build up syntax and semantics. CCGbank, a corpus automatically converted from the Wall Street Journal section of the Penn Treebank (Marcus et al., 1994), forms the cornerstone of re- search in wide-coverage CCG parsing. Its coverage enables the construction of practical, efficient and robust CCG parsers (Clark and Curran, 2007). We describe corpus transformations on CCGbank which improve its linguistic fidelity and discrim- inative power. We restore the quote symbols to CCGbank derivations removed by the CCGbank derivation procedure (Hockenmaier, 2003). Quotes yield useful local information for many corpus ap- plications, such as speaker or topic segmentation, and the supertagging phase in a CCG parser (the as- signment of categories to lexical items). Multi-modal Combinatory Categorial Grammar (Baldridge, 2002) is an extension to CCG for finer- grained control over the applicability of combina- tory rules. We develop a method for the semi- automatic extraction of a MMCCG corpus from CCGbank. These corpus transformations increase the fidelity and precision of CCGbank and hence its usefulness in a range of applications. 2 Multi-modal CCG In CCG, lexical items are mapped to categories, which are combined through a set of combinatory rules. Typically, permitting all of the combinators to be considered by the parser leads to the undesired acceptance of ungrammatical examples, known as overgeneration. To prevent this in pure CCG, restric- tions on the set of applicable rules must be specified as hard-coded parser constraints. X/⋆Y Y → X Y X\⋆Y → X X/◦Y Y/◦Z → X/◦Z Y \◦Z X\◦Y → X\◦Z Y/◦Z X\◦Y → X/◦Z X → T/i(T\iX) X → T\i(T/iX) To restore lexicality, Baldridge (2002) devised multi-modal CCG, in which each slash of a category encodes an indication (or mode) defining the rules in which it may participate. Modes simultaneously yield efficiency benefits and a reduction in deriva- tional ambiguity by limiting the set of combinatory rules a MMCCG parser has to consider. 3 Re-quoting CCGbank We restore quotes to CCGbank by consulting the Penn Treebank to determine the original location of Proceedings of the Australasian Language Technology Workshop 2007, pages 149-151 149 the opening and/or closing quote. Having found the corresponding leaf in the CCGbank derivation, we …

Content truncated to 500 words