Acquiring Translation Equivalences of Multiword Expressions by Normalized Correlation Frequencies

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 478–486, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Acquiring Translation Equivalences of Multiword Expressions by Normalized Correlation Frequencies Ming-Hong Bai1,2 Jia-Ming You1 Keh-Jiann Chen1 Jason S. Chang2 1 Institute of Information Science, Academia Sinica, Taiwan 2 Department of Computer Science, National Tsing-Hua University, Taiwan mhbai@sinica.edu.tw, swimming@hp.iis.sinica.edu.tw, kchen@iis.sinica.edu.tw, jschang@cs.nthu.edu.tw Abstract In this paper, we present an algorithm for ex- tracting translations of any given multiword expression from parallel corpora. Given a multiword expression to be translated, the method involves extracting a short list of tar- get candidate words from parallel corpora based on scores of normalized frequency, generating possible translations and filtering out common subsequences, and selecting the top-n possible translations using the Dice coefficient. Experiments show that our ap- proach outperforms the word alignment- based and other naive association-based me- thods. We also demonstrate that adopting the extracted translations can significantly im- prove the performance of the Moses machine translation system. 1 Introduction Translation of multiword expressions (MWEs), such as compound words, phrases, collocations and idioms, is important for many NLP tasks, including the techniques are helpful for dictio- nary compilation, cross language information retrieval, second language learning, and machine translation. (Smadja et al., 1996; Gao et al., 2002; Wu and Zhou, 2003). However, extracting exact translations of MWEs is still an open problem, possibly because the senses of many MWEs are not compositional (Yamamoto and Matsumoto, 2000), i.e., their translations are not composi- tions of the translations of individual words. For example, the Chinese idiom 坐視不理 should be translated as “turn a blind eye,” which has no direct relation with respect to the translation of each constituent (i.e., “to sit”, “to see” and “to ignore”) at the word level. Previous SMT systems (e.g., Brown et al., 1993) used a word-based translation model which assumes that a sentence can be translated into other languages by translating each word into one or more words in the target language. Since many concepts are expressed by idiomatic multiword expressions instead of single words, and different languages may realize the same concept using different numbers of words (Ma et al., 2007; Wu, 1997), word alignment based me- thods, which are highly dependent on the proba- bility information at the lexical level, are not well suited for this type of translation. To address the above problem, some methods have been proposed for extending word align- ments to phrase alignments. For example, Och et al. (1999) proposed the so-called grow-diag- final heuristic method for extending word alignments to phrase alignments. The method is widely used and has achieved good results for phrase-based statistical machine translation. (Och et al., 1999; Koehn et al., 2003; Liang et al., 2006). Instead of using heuristic rules, Ma et al. (2008) showed that syntactic information, e.g., phrase or dependency structures, is useful in ex- tending the word-level alignment. However, the above methods still depend on word-based alignment models, so they are not well suited to extracting the …

Content truncated to 500 words