Bilingual Knowledge Extraction Using Chunk Alignment

Bilingual Knowledge Extraction Using Chunk Alignment Young-Sook Hwang ATR SLT Research Labs 2-2-2 Hikaridai Seika-cho Soraku-gun Kyoto 619-0288, JAPAN youngsook.hwang@atr.jp Kyonghee Paik ATR SLT Research Labs 2-2-2 Hikaridai Seika-cho Soraku-gun Kyoto 619-0288, JAPAN kyonghee.paik@atr.jp Yutaka Sasaki ATR SLT Research Labs 2-2-2 Hikaridai Seika-cho Soraku-gun Kyoto 619-0288, JAPAN yutaka.sasaki@atr.jp Abstract In this paper, we propose a new method for effectively acquiring bilingual knowledge by ex- ploiting the dependency relations among the aligned chunks and words. We use a monolingual dependency parser to automatically obtain dependency parses of target language using chunk and word alignment. For reducing the computational complexity of structural alignment, we use a bilingual dictionary and adopt a divide-and-conquer strategy. By sharing the depen- dency relations of a given source sentence, we automatically obtain a dependency parse of a target sentence that is structurally consistent with the source sentence. Moreover, we extract bilingual knowledge bases from translation correspondences of singletons to surface verb sub- categorization patterns by exploiting the bilingual dependency relations. To acquire reliable ones, we take a stepwise filtering method based on statistical test. 1 Introduction Corpus-based approaches based on bilingual corpora are promising for automatically acquiring transla- tion knowledge (Alahawi et al,2000) (Brown et al,1993)(Imamura,2002)(Menezes et al,2001) (Utsuro et al,1993)(Watanabe et al,2000)(Yamada et al,2001). Most of statistical translation models based on IBM models(Brown et al,1993) are built from bilingual corpora without considering the structural aspects of the language(Brown et al,1993). They often output ungrammatical or unnatural translations. (Yamada et al,2001) modeled the translation process from a parse tree of source language into a target language sentence. But the computational complexity during alignment is very high because it must handle hierarchical structures. Some methods use parsed sentences in parallel sentence-aligned corpora to extract transfer rules or examples (Aramaki et al,2001) (Watanabe et al,2000) (Menezes et al,2001) (Imamura,2002). However, parse-to-parse matching, which regards parsing and alignment as separate and successive procedures, suffers from grammatical inconsistency between languages. Moreover, it costs a lot to develop parsers of both the source and target language. In this study, we propose a new method for effectively extracting translation knowledge which will reduce the computational complexity of alignment without losing the structural properties of each lan- guage. The main difference of our approach lies in exploiting a pair of a dependency parsed sentence and a POS tagged sentence as an input, and aligning those sentences in a chunk level as well as in a word level. By sharing the dependency relations of a source sentence, it is possible to automatically obtain a dependency parsed target sentence that will be structurally consistent with a given source sentence without using a parser of the target language. Ultimately, we can effectively acquire invaluable bilingual knowledge by exploiting the dependency relations among the aligned chunks and words. The proposed method was implemented into the system that consists of two sub-systems, a bilingual alignment system and a knowledge extraction system. For aligning chunks and words, we utilize bilin- gual dictionaries and take a divide-and-conquer strategy. That is, after …

Content truncated to 500 words