Improving Word Alignment with Bridge Languages

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 42–50, Prague, June 2007. c⃝2007 Association for Computational Linguistics Improving Word Alignment with Bridge Languages Shankar Kumar and Franz Och and Wolfgang Macherey Google Inc. 1600 Amphitheatre Parkway Mountain View, CA 94043, U.S.A. {shankarkumar,och,wmach}@google.com Abstract We describe an approach to improve Statistical Machine Translation (SMT) performance using multi-lingual, parallel, sentence-aligned corpora in several bridge languages. Our approach consists of a sim- ple method for utilizing a bridge language to create a word alignment system and a proce- dure for combining word alignment systems from multiple bridge languages. The final translation is obtained by consensus de- coding that combines hypotheses obtained using all bridge language word alignments. We present experiments showing that mul- tilingual, parallel text in Spanish, French, Russian, and Chinese can be utilized in this framework to improve translation performance on an Arabic-to-English task. 1 Introduction Word Alignment of parallel texts forms a cru- cial component of phrase-based statistical machine translation systems. High quality word alignments can yield more accurate phrase-pairs which improve quality of a phrase-based SMT system (Och and Ney, 2003; Fraser and Marcu, 2006b). Much of the recent work in word alignment has focussed on improving the word alignment quality through better modeling (Och and Ney, 2003; Deng and Byrne, 2005; Martin et al., 2005) or alternative approaches to training (Fraser and Marcu, 2006b; Moore, 2005; Ittycheriah and Roukos, 2005). In this paper we explore a complementary approach to improve word alignments using multi-lingual, par- allel (or multi-parallel) corpora. Two works in the literature are very relevant to our approach. Borin (2000) describes a non-statistical approach where a pivot alignment is used to combine direct translation and indirect translation via a third language. Filali and Bilmes (2005) present a multi-lingual extension to the IBM/HMM models. Our current approach dif- fers from this latter work in that we propose a sim- ple framework to combine word alignments from any underlying statistical alignment model without the need for changing the structure of the model. While both of the above papers focus on improv- ing word alignment quality, we demonstrate that our approach can yield improvements in transla- tion performance. In particular, we aim to improve an Arabic-to-English (Ar-En) system using multi- parallel data from Spanish (Es), French (Fr), Rus- sian (Ru) and Chinese (Zh). The parallel data in these languages X ∈ {Es, Fr, Ru, Zh} is used to generate word alignments between Arabic-X and X-English. These alignments are then combined to obtain multiple word alignments for Arabic-English and the final translation systems. The motivation for this approach is two-fold. First, we believe that parallel corpora available in several languages provide a better training ma- terial for SMT systems relative to bilingual cor- pora. Such multi-lingual parallel corpora are be- coming widely available; examples include proceed- ings of the United Nations in six languages (UN, 2006), European Parliament (EU, 2005; Koehn, 2003), JRC Acquis corpus (EU, 2007) and religious …

Content truncated to 500 words