Discriminative Word Alignment via Alignment Matrix Modeling

Proceedings of the Third Workshop on Statistical Machine Translation, pages 18–25, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Discriminative Word Alignment via Alignment Matrix Modeling Jan Niehues Institut f¨ur Theoretische Informatik Universit¨at Karlsruhe (TH) Karlsruhe, Germany jniehues@ira.uka.de Stephan Vogel Language Technologies Institute Carnegie Mellon University Pittsburgh, PA, 15213, USA stephan.vogel@cs.cmu.edu Abstract In this paper a new discriminative word align- ment method is presented. This approach models directly the alignment matrix by a con- ditional random field (CRF) and so no restric- tions to the alignments have to be made. Fur- thermore, it is easy to add features and so all available information can be used. Since the structure of the CRFs can get complex, the in- ference can only be done approximately and the standard algorithms had to be adapted. In addition, different methods to train the model have been developed. Using this approach the alignment quality could be improved by up to 23 percent for 3 different language pairs compared to a combination of both IBM4- alignments. Furthermore the word alignment was used to generate new phrase tables. These could improve the translation quality signifi- cantly. 1 Introduction In machine translation parallel corpora are one very important knowledge source. These corpora are of- ten aligned at the sentence level, but to use them in the systems in most cases a word alignment is needed. Therefore, for a given source sentence fJ 1 and a given target sentence eI 1 a set of links (j, i) has to be found, which describes which source word fj is translated into which target word ei. Most SMT systems use the freely available GIZA++-Toolkit to generate the word alignment. This toolkit implements the IBM- and HMM- models introduced in (Brown et al., 1993; Vogel et al., 1996). They have the advantage that they are trained unsupervised and are well suited for a noisy- channel approach. But it is difficult to include addi- tional features into these models. In recent years several authors (Moore et al., 2006; Lacoste-Julien et al., 2006; Blunsom and Cohn, 2006) proposed discriminative word align- ment frameworks and showed that this leads to im- proved alignment quality. In contrast to generative models, these models need a small amount of hand- aligned data. But it is easy to add features to these models, so all available knowledge sources can be used to find the best alignment. The discriminative model presented in this pa- per uses a conditional random field (CRF) to model the alignment matrix. By modeling the matrix no restrictions to the alignment are required and even n:m alignments can be generated. Furthermore, this makes the model symmetric, so the model will pro- duce the same alignment no matter which language is selected as source and which as target language. In contrast, in generative models the alignment is a function where a source word aligns to at most one target word. So the alignment is asymmetric. The training of this discriminative model has to be done on …

Content truncated to 500 words