HASHIMOTO Chikara: Computational Parsimony in the Case of V-V Compounds in Japanese

by HASHIMOTO Chikara

As a type of Multiword Expression (Sag et atl., 2002; Baldwin & Bond, 2002), Japanese verbal compounds (V₁-V₂ compounds, hereafter) pose serious problems for Japanese Natural Language Processing (NLP) and require a sophisticated linguistic treatment. Hashimoto (2004) presented such a treatment of V₁-V₂ compounds based on the JACY framework (Siegel & Bender, 2002). However, the treatment suffered from overgeneration involving what Matsumoto (1996) calls `V₁-V₂ with semantically deverbalized V₁', which is peculiar in that, even though it shows (partial) compositionality, its productivity is very restricted. In this paper, I propose an alternative analysis for V₁-V₂ with semantically deverbalized V₁, where all V₁-V₂s of that kind are regarded as single words to account for their restricted productivity but are given (partially) compositional semantic representations. Also, I report on an evaluation experiment that shows an advantage of the alternative treatment. Finally, I argue that grammar developers should take into account computational parsimony; we should not try too hard to generalize phenomena that we can easily enumerate exhaustively.