The University of Washington Machine Translation System for ACL WMT 2008

Proceedings of the Third Workshop on Statistical Machine Translation, pages 123–126, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics The University of Washington Machine Translation System for ACL WMT 2008 Amittai Axelrod, Mei Yang, Kevin Duh, Katrin Kirchhoff Department of Electrical Engineering University of Washington Seattle, WA 98195 {amittai,yangmei,kevinduh,katrin} @ee.washington.edu Abstract This paper present the University of Washing- ton’s submission to the 2008 ACL SMT shared ma- chine translation task. Two systems, for English-to- Spanish and German-to-Spanish translation are de- scribed. Our main focus was on testing a novel boosting framework for N-best list reranking and on handling German morphology in the German-to- Spanish system. While boosted N-best list reranking did not yield any improvements for this task, simpli- fying German morphology as part of the preprocess- ing step did result in significant gains. 1 Introduction The University of Washington submitted systems to two data tracks in the WMT 2008 shared task competition, English-to-Spanish and German-to- Spanish. In both cases, we focused on the in-domain test set only. Our main interest this year was on in- vestigating an improved weight training scheme for N-best list reranking that had previously shown im- provements on a smaller machine translation task. For German-to-Spanish translation we additionally investigated simplifications of German morphology, which is known to be fairly complex due to a large number of compounds and inflections. In the fol- lowing sections we first describe the data, baseline system and postprocessing steps before describing boosted N-best list reranking and morphology-based preprocessing for German. 2 Data and Basic Preprocessing We used the Europarl data as provided (version 3b, 1.25 million sentence pairs) for training the transla- tion model for use in the shared task. The data was lowercased and tokenized with the auxiliary scripts provided, and filtered according to the ratio of the sentence lengths in order to eliminate mismatched sentence pairs. This resulted in about 965k paral- lel sentences for English-Spanish and 950k sentence pairs for German-Spanish. Additional preprocess- ing was applied to the German corpus, as described in Section 5. For language modeling, we addition- ally used about 82M words of Spanish newswire text from the Linguistic Data Consortium (LDC), dating from 1995 to 1998. 3 System Overview 3.1 Translation model The system developed for this year’s shared task is a state-of-the-art, two-pass phrase-based statisti- cal machine translation system based on a log-linear translation model (Koehn et al, 2003). The trans- lation models and training method follow the stan- dard Moses (Koehn et al, 2007) setup distributed as part of the shared task. We used the training method suggested in the Moses documentation, with lexical- ized reordering (the msd-bidirectional-fe option) enabled. The system was tuned via Mini- mum Error Rate Training (MERT) on the first 500 sentences of the devtest2006 dataset. 123 3.2 Decoding Our system used the Moses decoder to generate 2000 output hypotheses per input sentence during the first translation pass. For the second pass, the N-best lists were rescored with the additional lan- guage models described below. …

Content truncated to 500 words