Verb Alternations and Japanese : How, What and Where

Verb Alternations and Japanese� How, What and Where? Timothy Baldwin and Hozumi Tanaka Tokyo Institute of Technology 2-12-1 Ookayama, Meguro-ku, Tokyo 152-8552 JAPAN ftim,tanakal@cl.cs.titech.ac.jp Abstract We set out to empirically identify the range and frequency of basic verb alternation types in Japanese, through analysis of the Goi-Taikei Japanese pattern-based valency dictionary. This is achieved through comparison of the selectional preference annotation on corresponding case slots, based on the assumption that selectional preferences are preserved under alternation. Three separate extraction methods are considered, founded around: (1) simple match of selectional restrictions; (2) selectional restriction matching, with recourse to penalised backing-off; and (3) semantic density, again with recourse to backing-off. 1 Introduction This is an attempt to document the full range of verb alternations in Japanese from analysis of structural regularities between entries in a valency dictionary, and determine the degree of permeation of each individual alternation type. Various techniques are analysed for extracting alternations, focusing on full or partial preservation of the selectional restrictions describing each case slot, and the degree of restriction of the selectional restrictions. These methods operate around the assumption that selectional restrictions are essentially unchanged under alternation.) Verb alternations have been the target of considerable research within linguistic circles, in terms of describing the range of alternations for a particular language (e.g. (Levin 1993) for English and (Fukui et al. 1985) for Japanese), applying alternations to model verbal semantics (e.g. (Levin 1993; Hale & Keyser 1987)), or analysing alternating potential and interpretational differences between verbs in different clause contexts (e.g. (Goldberg 1995; Wierzbicka 1988)). More recently, alternations have been the object of interest within the natural language processing community, for use in lexical selection in natural language generation (Dorr & Olsen 1996; Jing & McKeown 1998), and as a means of expanding dictionary coverage (Baldwin et al. 1999b). There has also been some work on automatically extracting verbs which undergo particular alternations from corpora (McCarthy Korhonen 1998). Our work represents a variation on this same theme, whereby we compare the different senses 2 of a given verb within the Goi-Taikei pattern-based valency dictionary (Ikehara et al. 1997; Shirai et al. 1997), and exhaustively determine all possible mappings between each valency frame pair. Naturally, not all such mappings are going to constitute true alternations, but by scoring each mapping and combining the scores for each mapping paradigm over all dictionary entries, it is hoped that alternations will find themselves into the top-ranking analyses. One area in which this research attempts to break new ground is, therefore, the fully-unsupervised extraction of alternations. Our purpose in extracting alternations is twofold. Firstly, we are interested in determining the possibilities for collapsing dictionary entries together through dynamic generation of regular alterna- tions. Preliminary research on exactly this subject indicates that the use of prominent alternation types produces a reduction of almost 9% in the number of senses (Baldwin et al. 1999a), suggesting that a fuller set of alternations could produce even greater reduction. Secondly, we are interested in deriving the inventory of …

Content truncated to 500 words