A pattern-learning based, hybrid model for the syntactic analysis of structural relationshi.pdf
文本预览下载声明
A Pattern-Learning Based, Hybrid Model for the Syntactic
Analysis of Structural Relationships among Japanese Clauses
Akitoshi OKUMURA Kazunori MURAKI Kiyoshi YAMABANA
NEC Corp. C C Information Technology Research Laboratories
4-1-1 Miyazaki, Miyamae-ku, Kawasaki 216, JAPAN
okumura%mtl.cl.nec.co.jp@
Abstract
This paper presents a model for analyzing Japanese compound sentences by taking advantage
of users examples. Syntactic analysis of the Japanese compound sentences is one of the most
difficult problems for machine translation (MT) systems. One particularly difficult problem is
selecting, from all possible candidates, the correct global structure of an individual sentence, i.e.
recognizing the set of structural relationships actually linking the various clauses in a compound
sentence. MT systems are generally equipped with correction tools for users to modify the
incorrect results of the system analyses. Users must use such tools over and over again to modify
the same kind of sentences, however, because there is no effective method to memorize the
previously corrected examples. The authors here propose a pattern-learning based, hybrid model
for analyzing structural relationships among Japanese clauses. The model consists of rule-based
modules and a learning module which memorizes correct global structures as well as incorrect
ones. All structures are memorized in the form of five-region patterns which are represented by
salient features. The model is investigated in comparison with the properties of connectionist
approaches, and then the validity of the model is supported by the the results of preliminary
experiments.
1 Introduction
This paper presents a model for analyzing Japanese compound sentences by taking advantage of
users examples.
Syntactic analysis of Japanese compound sentences is one of the most difficult problems for
machine translation (MT) systems. It requires two kinds of analyses: one for local structures,
that is, to
显示全部