More Linguistic Annotation for Statistical Machine Translation

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 115–120, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics More Linguistic Annotation for Statistical Machine Translation Philipp Koehn, Barry Haddow, Philip Williams, and Hieu Hoang University of Edinburgh Edinburgh, United Kingdom {pkoehn,bhaddow,p.j.williams-2,h.hoang}@inf.ed.ac.uk Abstract We report on efforts to build large-scale translation systems for eight European language pairs. We achieve most gains from the use of larger training corpora and basic modeling, but also show promising results from integrating more linguistic an- notation. 1 Introduction We participated in the shared translation task of the ACL Workshop for Statistical Machine Trans- lation 2010 in all language pairs. We continued our efforts to integrate linguistic annotation into the translation process, using factored and tree- based translation models. On average we out- performed our submission from last year by 2.16 BLEU points on the same newstest2009 test set. While the submitted system follows the factored phrase-based approach, we also built hierarchical and syntax-based models for the English–German language pair and report on its performance on the development test sets. All our systems are based on the Moses toolkit (Koehn et al., 2007). We achieved gains over the systems from last year by consistently exploiting all available train- ing data, using large-scale domain-interpolated, and consistent use of the factored translation model to integrate n-gram models over speech tags. We also experimented with novel domain adaptation methods, with mixed results. 2 Baseline System The baseline system uses all available training data, except for the large UN and 109 corpora, as well as the optional LDC Gigaword corpus. It uses a straight-forward setup of the Moses decoder. Some relevant parameter settings are: • maximum sentence length 80 words • tokenization with hyphen splitting • truecasing • grow-diag-final-and alignment heuristic • msd-bidirectional-fe lexicalized reordering • interpolated 5-gram language model • tuning on newsdev2009 • testing during development on newstest2009 • MBR decoding • no reordering over punctuation • cube pruning We used most of these setting in our submission last year (Koehn and Haddow, 2009). The main difference to our baseline system from the submission from last year is the use of ad- ditional training data: larger releases of the News Commentary, Europarl, Czeng, and monolingual news corpora. The first two parallel corpora in- creased roughly 10-20% in size, while the Czeng parallel corpus and the monolingual news corpora are five times and twice as big, respectively. We also handled some of the corpus preparation steps with more care to avoid some data incon- sistency problems from last year (affecting mostly the French language pairs). An overview of the results is given in Table 1. The baseline outperforms our submission from last year by an average of +1.25 points. The gains for the individual language pairs track the increase in training data (most significantly for the Czech– English pairs), and the French–English data pro- cessing issue. Note that last year’s submission used special handling of the German–English language pair, which we did not …

Content truncated to 500 words