Tuning Syntactically Enhanced Word Alignment for Statistical Machine Translation

Proceedings of the 13th Annual Conference of the EAMT, pages 250–257, Barcelona, May 2009 Tuning Syntactically Enhanced Word Alignment for Statistical Machine Translation Yanjun Ma† Patrik Lambert‡ Andy Way†‡ †National Centre for Language Technology ‡Centre for Next Generation for Localisation School of Computing, Dublin City University Dublin 9, Ireland {yma,plambert,away@computing.dcu.ie} Abstract We introduce a syntactically enhanced word alignment model that is more flex- ible than state-of-the-art generative word alignment models and can be tuned ac- cording to different end tasks. First of all, this model takes the advantages of both unsupervised and supervised word alignment approaches by obtaining anchor alignments from unsupervised generative models and seeding the anchor alignments into a supervised discriminative model. Second, this model offers the flexibility of tuning the alignment according to differ- ent optimisation criteria. Our experiments show that using our word alignment in a Phrase-Based Statistical Machine Trans- lation system yields a 5.38% relative in- crease on IWSLT 2007 task in terms of BLEU score. 1 Introduction Word alignment, which can be defined as a problem of determing word-level correspondences given a parallel corpus of aligned sentences, is a fundamental component in Phrase-Based Statisti- cal Machine Translation (PB-SMT). The dominant approach to word alignment are generative models, including IBM models (Brown et al., 1993) and HMM models (Vogel et al., 1996; Deng and Byrne, 2006). While generative models trained in an un- supervised manner can produce high-quality align- ments given a reasonable amount of training data, it is difficult to incorporate richer features into such models. On the other hand, discriminative models are more flexible to incorporate arbitrary features. c⃝ 2009 European Association for Machine Translation. However, these models need a certain amount of annotated word alignment data, which is often sub- ject to criticism since the annotation of word align- ment is a highly subjective task. Moreover, param- eters optimised on manually annotated data are not necessarily optimal for MT tasks. Recent research attempts to combine the merits of both genera- tive and discriminative models (Fraser and Marcu, 2007), or to tune a discriminative model according to MT metrics (Lambert et al., 2007). In this paper, we introduce a simple yet flexible framework for word alignment. To take the ad- vantage of the strength of generative models, we use these models to obtain a set of anchor align- ments. We then incorporate syntactic features in- duced by the anchor alignments into a discrimi- native word alignment model. The syntactic fea- tures we used are syntactic dependencies. This de- cision is motivated by the fact that if words tend to be dependent on each other, so does the align- ment (Ma et al., 2008). If we can first obtain a set of reliable anchor links, we could take advantage of the syntactic dependencies relating unaligned words to aligned anchor words to expand the align- ment. Figure 1 gives an illustrative example. Note that the link (c2, e4) can be easily identified, but the link involving the fourth Chinese word (a func- tion word denoting …

Content truncated to 500 words