Building Strong Multilingual Aligned Corpora

Proceedings of the 13th Annual Conference of the EAMT, pages 176–181, Barcelona, May 2009 Building Strong Multilingual Aligned Corpora Reza Bosagh Zadeh Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA rezab@cs.cmu.edu Abstract Recent advances have allowed algorithms that learn from aligned natural language texts to exploit aligned sentences in more than two languages. We investigate ways of combining �N 2 � bilingual aligned cor- pora together to create a multilingual aligned corpus across N languages. As a result of the combination of several cor- pora, our algorithms output a multilingual corpus, with each aligned tuple assigned a quality score called ‘strength’ that may be used when learning from the multilin- gual corpus. We show that the addition of bilingual corpora used with alignment strengths can significantly improve Statis- tical Machine Translation quality on an Arabic→English task. 1 Introduction In Machine Translation, it is desirable and intu- itive that bridging through languages other than the source and target should help improve trans- lation quality between the source and target. By using bilingual alignments across all pairs of N languages, (Kumar et al., 2007) was able to gain in translation quality between two languages by way of alignment bridges. They describe an ap- proach to improve Statistical Machine Transla- tion (SMT) performance using multi-lingual, par- allel, sentence-aligned corpora in several bridge languages. Their approach consists of a simple method for utilizing a bridge language to cre- ate a word alignment system and a procedure for combining word alignment systems from multi- ple bridge languages. They present experiments c⃝ 2009 European Association for Machine Translation. showing that multilingual, parallel text in Span- ish, French, Russian, and Chinese can be utilized to improve translation performance on an Arabic- to-English task. Other papers that use multilin- gual aligned corpora include (Borin, 2000; Mann and Yarowsky, 2001; Simard, 1999; Kumar et al., 2007). With the advent of such learning strategies requiring a multilingual aligned corpus, it is desir- able to take existing bilingual aligned corpora and combine them into a single multilingual aligned corpus. We investigate ways of combining �N 2 � bilin- gual aligned corpora together to create a multi- lingual aligned corpus across N languages. As a result of the combination of several corpora, our algorithms not only output a multilingual corpus, but each multilingual aligned tuple is further as- signed a quality score that may be used when learning. We call the alignment quality measure produced by the combination procedure the align- ment’s strength. By using alignment strengths in a setting similar to (Kumar et al., 2007) we show that strengths can be used to improve SMT qual- ity in an Arabic→English task by re-weighing the training corpus to give higher weight to stronger alignments. This allowed us to add more bilingual corpora at training time while consistently improv- ing translation quality. The problem of creating multilingual align- ments from bilingual ones is not trivial. As an example of the problem being solved, consider 4 sentences A, B, C, D in …

Content truncated to 500 words