The use of abstracted knowledge from an automatically sense-tagged corpus for lexical transfer ambiguity resolution

MT Summit VII Sept. 1999 The Use of Abstracted Knowledge from an Automatically Sense-Tagged Corpus for Lexical Transfer Ambiguity Resolution Hui-Feng Li, Namwon Heo, Kyounghi Moon, Jong-Hyeok Lee Dept. of Computer Science & Engineering Pohang University of Science & Technology San 31 Hyojadong, Nam-Gu, Pohang, 790-784, Korea E-mail: {hflee, nwheo, khmoon, jhlee}@kle.postech.ac.kr Abstract This paper proposes a method for lexical transfer ambiguity resolution using corpus and conceptual information. Previous researches have restricted the use of linguistic knowledge to the lexical level. Since the extracted knowledge is stored in words themselves, the- se methods require a large amount of space with a low recall rate. On the contrary, we re- solve word sense ambiguity by using concept co-occurrence information extracted from an automatically sense-tagged corpus. In one ex- periment, it achieved, on average, a precision of 82.4% for nominal words, and 83% for verbal words. Considering that the test corpus is completely irrelevant to the learning cor- pus, this is a promising result. 1 Introduction In Korean-to-Japanese machine translation (MT), which employs a direct MT strategy, a Korean word with multiple senses may be translated into different Japanese equivalents depending on which sense is used in a given context. In general, it is much more difficult for Korean- to-Japanese MT to resolve such word sense ambiguities than its opposite, Japanese-to-Korean MT. This is be- cause unlike Japanese words whose stem is usually writ- ten in Kanji (the Chinese ideographic script), Korean words may appear written in Hangul (the Korean phonetic alphabet) only. Thus, word sense disambiguation (WSD) is essential to lexical transfer in an Korean-to-Japanese MT system. There is a large body of previous work on word sense disambiguation (Kelly 1975; Church 1990; McRoy 1992; Liu 1993; Tanaka 1994; Yarowsky 1995; Hwee 1996). Early work made use of manually coded knowledge, but it required time-consuming and laborious work for knowl- edge acquisition (Kelly and Stone 1975). The recent em- phases on corpus-based WSD usually fall into two cate- gories: supervised and unsupervised. Yarowsky (1995) used an unsupervised learning procedure. However, his experiment was only restricted to “binary” WSD, a kind of coarse sense distinction (Yarowsky 1995; Hwee and Lee 1996). LEXAS (Hwee and Lee 1996) adopted a su- pervised learning method using multiple knowledge sources such as a human sense-tagged corpus, the POS of neighboring words, morphological forms, local colloca- tions, unordered sets of surrounding words, and verb- object syntactic relations. Since the extracted knowledge in these methods is stored in words themselves, they re- quire a large amount of space to store the knowledge, and they have lower applicability. In this paper, a corpus-based WSD method is pro- posed. Unlike previous work (Yarowsky 1995; Hwee 1996) that restricted the use of linguistic knowledge to the lexical level, i.e., surface words, we rely on concept co- occurrence information (CCI) extracted from a sense- tagged corpus. The sense-tagged corpus is automatically constructed from a Japanese raw corpus by an existing Japanese-to-Korean MT system called COBALT-J/K (Collocation-Based Language Translator) (Park 1997). The concept …

Content truncated to 500 words