Experiments with word alignment, normalization and clause reordering for SMT between English and German

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 393–398, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Experiments with word alignment, normalization and clause reordering for SMT between English and German Maria Holmqvist, Sara Stymne and Lars Ahrenberg Department of Computer and Information Science Link¨oping University, Sweden firstname.lastname@liu.se Abstract This paper presents the LIU system for the WMT 2011 shared task for translation be- tween German and English. For English– German we attempted to improve the trans- lation tables with a combination of standard statistical word alignments and phrase-based word alignments. For German–English trans- lation we tried to make the German text more similar to the English text by normalizing Ger- man morphology and performing rule-based clause reordering of the German text. This re- sulted in small improvements for both transla- tion directions. 1 Introduction In this paper we present the LIU system for the WMT11 shared task, for translation between En- glish and German in both directions. We added a number of features that address problems for trans- lation between German and English such as word or- der differences, incorrect alignment of certain words such as verbs, and the morphological complexity of German compared to English, as well as dealing with previously unseen words. In both translation directions our systems in- clude compound processing, morphological se- quence models, and a hierarchical reordering model. For German–English translation we also added mor- phological normalization, source side reordering, and processing of out-of-vocabulary words (OOVs). For English–German translation, we extracted word alignments with a supervised method and combined these alignments with Giza++ alignments in various ways to improve the phrase table. We experimented with different ways of combining the two alignments such as using heuristic symmetrization and interpo- lating phrase tables. Results are reported on three metrics, BLEU (Pa- pineni et al., 2002), NIST (Doddington, 2002) and Meteor ranking scores (Agarwal and Lavie, 2008) based on truecased output. 2 Baseline System This years improvements were added to the LIU baseline system (Stymne et al., 2010). Our base- line is a factored phrase based SMT system that uses the Moses toolkit (Koehn et al., 2007) for transla- tion model training and decoding, GIZA++ (Och and Ney, 2003) for word alignment, SRILM (Stol- cke, 2002) an KenLM (Heafield, 2011) for language modelling and minimum error rate training (Och, 2003) to tune model feature weights. In addition, the LIU baseline contains: • Compound processing, including compound splitting and for translation into German also compound merging • Part-of-speech and morphological sequence models All models were trained on truecased data. Trans- lation and reordering models were trained using the bilingual Europarl and News Commentary corpora that were concatenated before training. We created two language models. The first model is a 5-gram model that we created by interpolating two language 393 models from bilingual News Commentary and Eu- roparl with more weight on the News Commentary model. The second model is a 4-gram model trained on monolingual News only. All models were cre- ated …

Content truncated to 500 words