A new dependency analysis method based on semantically embedded sentence structures and its performance on Japanese subordinate clauses

Satoshi SHIRAI+, Satoru IKEHARA+, Akio YOKOO+ and Junko KIMURA+


+ NTT Communication Science Laboratories + NTT Advanced Technology Corporation
Take 1-2356, Yokosuka, 238-03, Japan Kawakami-cho 90-6, Totsuka-ku, Yokohama, 244, Japan
{shirai,ikehara,ayokoo}@nttkb.ntt.jp kimura@totsuka.natc.co.jp


Abstract

The ambiguity in dependency analysis of Japanese predicates has been one of the major causes of the failure of long sentence analyses. To overcome this problem, a new dependency analysis method for Japanese subordinate clauses based on semantically embedded structure has been proposed and its accuracy has been evaluated.

In this paper, the tree levels of hierarchical subordinate clauses proposed by the Japanese philologist F. Minami, are reviewed for their syntax and semantics, and broken down into 13 basic categories and 4 detailed categories. A hierarchical ordering of these categories yields the dependency relations between predicates.

Experiments are conducted with some 972 newspaper articles sentences (totaling 2,327 predicates of which 661 are classified as having ambiguous dependency). Where as the conventional method leaves 356 predicates of ambiguous dependency, this number is reduced to 54 predicates by the new method. Accordingly, the accuracy of the first candidate of the dependency relations within a sentence is improved from 92% to 98%.



Keywords

dependency analysis, syntactic analysis, Japanese analysis, subordinate clause, semantically embedded structure



[ Transaction of Information Processing Society of Japan, pp.2353-2361 (October, 1995). ]