Using Inverse Lexical Rules to Acquire a Wide-coverage Lexicalized Grammar

Automatic grammar extraction from annotated corpora (Xia, 1999; Chen and Vijay-Shanker, 2000; Chiang, 2000; Hockenmaier and Steedman, 2002; Miyao et al., 2004) enabled us to build a widecoverage lexicalized grammar at low cost. They succeeded in extracting a large number of lexical entries with less effort while conventional methods only allow limited lexical entries to be acquired in realworld texts. Lexicalized grammars require many lexical entries to explain various syntactic alternations, and we can hardly expect that all words will appear in all possible syntactic alternations within a limited training corpus. We aimed at improving the coverage of an automatically extracted grammar using lexical rules in this work (Jackendoff, 1975; Pollard and Sag, 1994). The idea behind lexical rules is that the syntactic constraints of a group of words are derived with general rules from their lexemes, which express characteristics common to the group (e.g. “runs” or “running” is derived from the lexeme “run”). We automatically acquired lexemes by applying lexical rules inversely to the lexical entries of the HPSG grammar extracted automatically from the Penn Treebank (Marcus et al., 1993). We could then generate a wide set of lexical entries from the lexemes, and our grammar achieved a higher coverage against real-world texts. Although the lexical rules proposed by Pollard and Sag (1994) treated several parts-ofspeech such as nouns or adjectives, we only formulated rules for verbs. This is because verbs PHON “scold” HEAD verb CAT MODL null MODR null SUBJ < HEAD noun > COMPS < HEAD noun > NONLOCAL SLASH <>