Improving Alignment for SMT by Reordering and Augmenting the Training Corpus

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 120–124, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Improving alignment for SMT by reordering and augmenting the training corpus Maria Holmqvist, Sara Stymne, Jody Foo and Lars Ahrenberg Department of Computer and Information Science Linköping University, Sweden {marho,sarst,jodfo,lah}@ida.liu.se Abstract We describe the LIU systems for English- German and German-English translation in the WMT09 shared task. We focus on two methods to improve the word align- ment: (i) by applying Giza++ in a sec- ond phase to a reordered training cor- pus, where reordering is based on the alignments from the first phase, and (ii) by adding lexical data obtained as high- precision alignments from a different word aligner. These methods were studied in the context of a system that uses com- pound processing, a morphological se- quence model for German, and a part- of-speech sequence model for English. Both methods gave some improvements to translation quality as measured by Bleu and Meteor scores, though not consis- tently. All systems used both out-of- domain and in-domain data as the mixed corpus had better scores in the baseline configuration. 1 Introduction It is an open question whether improved word alignment actually improves statistical MT. Fraser and Marcu (2007) found that improved alignments as measured by AER will not necessarily improve translation quality, whereas Ganchev et al. (2008) did improve translation quality on several lan- guage pairs by extending the alignment algorithm. For this year’s shared task we therefore stud- ied the effects of improving word alignment in the context of our system for the WMT09 shared task. Two methods were tried: (i) applying Giza++ in a second phase to a reordered training corpus, where reordering is based on the alignments from the first phase, and (ii) adding lexical data ob- tained as high-precision alignments from a differ- ent word aligner. The submitted system includes the first method in addition to the processing of compounds and additional sequence models used by Stymne et al. (2008). Heuristics were used to generate true-cased versions of the translations that were submitted, as reported in section 6. In this paper we report case-insensitive Bleu scores (Papineni et al., 2002), unless otherwise stated, calculated with the NIST tool, and case- insensitive Meteor-ranking scores, without Word- Net (Agarwal and Lavie, 2008). 2 Baseline system Our baseline system uses compound split- ting, compound merging and part-of- speech/morphological sequence models (Stymne et al., 2008). Except for these additions it is similar to the baseline system of the workshop1. The translation system is a factored phrase- based translation system that uses the Moses toolkit (Koehn et al., 2007) for decoding and train- ing, GIZA++ for word alignment (Och and Ney, 2003), and SRILM (Stolcke, 2002) for language models. Minimum error rate training was used to tune the model feature weights (Och, 2003). Tuning was performed on the news-dev2009a set with 1025 sentences. All development test- ing was performed on the news-dev2009b set with 1026 sentences. …

Content truncated to 500 words