Bilingual Lexicon Extraction from Comparable Corpora Using Label Propagation

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 24–36, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Bilingual Lexicon Extraction from Comparable Corpora Using Label Propagation Akihiro Tamura and Taro Watanabe and Eiichiro Sumita Multilingual Translation Laboratory, MASTAR Project National Institute of Information and Communications Technology 3-5 Hikaridai, Keihanna Science City, Kyoto, 619-0289, JAPAN {akihiro.tamura,taro.watanabe,eiichiro.sumita}@nict.go.jp Abstract This paper proposes a novel method for lex- icon extraction that extracts translation pairs from comparable corpora by using graph- based label propagation. In previous work, it was established that performance drasti- cally decreases when the coverage of a seed lexicon is small. We resolve this problem by utilizing indirect relations with the bilin- gual seeds together with direct relations, in which each word is represented by a distri- bution of translated seeds. The seed distri- butions are propagated over a graph repre- senting relations among words, and transla- tion pairs are extracted by identifying word pairs with a high similarity in the seed dis- tributions. We propose two types of the graphs: a co-occurrence graph, representing co-occurrence relations between words, and a similarity graph, representing context sim- ilarities between words. Evaluations using English and Japanese patent comparable cor- pora show that our proposed graph propaga- tion method outperforms conventional meth- ods. Further, the similarity graph achieved im- proved performance by clustering synonyms into the same translation. 1 Introduction Bilingual lexicons are important resources for bilin- gual tasks such as machine translation (MT) and cross-language information retrieval (CLIR). There- fore, the automatic building of bilingual lexicons from corpora is one of the issues that have attracted many researchers. As a solution, a number of pre- vious works proposed extracting bilingual lexicons from comparable corpora, in which documents were not direct translations but shared a topic or domain1. The use of comparable corpora is motivated by the fact that large parallel corpora are only available for a few language pairs and for limited domains. Most of the previous methods are based on as- sumption (I), that a word and its translation tend to appear in similar contexts across languages (Rapp, 1999). Based on this assumption, many methods calculate word similarity using context and then ex- tract word translation pairs with a high-context sim- ilarity. We call these methods context-similarity- based methods. The context similarities are usu- ally computed using a seed bilingual lexicon (e.g. a general bilingual dictionary) by mapping contexts expressed in two different languages into the same space. In the mapping, information not represented by the seed lexicon is discarded. Therefore, the context-similarity-based methods could not find ac- curate translation pairs if using a small seed lexicon. Some of the previous methods tried to alleviate the problem of the limited seed lexicon size (Koehn and Knight, 2002; Morin and Prochasson, 2011; Hazem et al., 2011), while others did not require any seed lexicon (Rapp, 1995; Fung, 1995; Haghighi et al., 2008; Ismail and Manandhar, 2010; Daum´e III and Jagarlamudi, …

Content truncated to 500 words