Improving Word Alignment with Language Model Based Confidence Scores

Proceedings of the Third Workshop on Statistical Machine Translation, pages 151–154, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Improving Word Alignment with Language Model Based Confidence Scores Nguyen Bach, Qin Gao, Stephan Vogel InterACT, Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {nbach, qing, vogel+}@cs.cmu.edu Abstract This paper describes the statistical machine trans- lation systems submitted to the ACL-WMT 2008 shared translation task. Systems were submitted for two translation directions: English→Spanish and Spanish→English. Using sentence pair confidence scores estimated with source and target language models, improvements are observed on the News- Commentary test sets. Genre-dependent sentence pair confidence score and integration of sentence pair confidence score into phrase table are also in- vestigated. 1 Introduction Word alignment models are a crucial component in sta- tistical machine translation systems. When estimating the parameters of the word alignment models, the sen- tence pair probability is an important factor in the objec- tive function and is approximated by the empirical prob- ability. The empirical probability for each sentence pair is estimated by maximum likelihood estimation over the training data (Brown et al., 1993). Due to the limitation of training data, most sentence pairs occur only once, which makes the empirical probability almost uniform. This is a rather weak approximation of the true distribution. In this paper, we investigate the methods of weighting sentence pairs using language models, and extended the general weighting method to genre-dependent weight. A method of integrating the weight directly into the phrase table is also explored. 2 The Baseline Phrase-Based MT System The ACL-WMT08 organizers provided Europarl and News-Commentary parallel corpora for English ↔ Span- ish. Detailed corpus statistics is given in Table 1. Follow- ing the guidelines of the workshop we built baseline sys- tems, using the lower-cased Europarl parallel corpus (re- stricting sentence length to 40 words), GIZA++ (Och and Ney, 2003), Moses (Koehn et al., 2007), and the SRI LM toolkit (Stolcke, 2002) to build 5-gram LMs. Since no News development sets were available we chose News- Commentary sets as replacements. We used test-2006 (E06) and nc-devtest2007 (NCd) as development sets for Europarl and News-Commentary; test-2007 (E07) and nc-test2007 (NCt) as held-out evaluation sets. English Spanish Europarl (E) sentence pairs 1,258,778 unique sent. pairs 1,235,134 avg. sentence length 27.9 29.0 # words 35.14 M 36.54 M vocabulary 108.7 K 164.8 K News-Commentary (NC) sentence pairs 64,308 unique sent. pairs 64,205 avg. sentence length 24.0 27.4 # words 1.54 M 1.76 M vocabulary 44.2 K 56.9 K Table 1: Statistics of English↔Spanish Europarl and News- Commentary corpora To improve the baseline performance we trained sys- tems on all true-cased training data with sentence length up to 100. We used two language models, a 5-gram LM build from the Europarl corpus and a 3-gram LM build from the News-Commentary data. Instead of interpolat- ing the two language models, we explicitly used them in the decoder and optimized their weights via minimum- error-rate (MER) training (Och, 2003). To shorten the training time, a multi-threaded …

Content truncated to 500 words