Tunable Distortion Limits and Corpus Cleaning for SMT

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 225–231, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Tunable Distortion Limits and Corpus Cleaning for SMT Sara Stymne Christian Hardmeier J¨org Tiedemann Joakim Nivre Uppsala University Department of Linguistics and Philology firstname.lastname@lingfil.uu.se Abstract We describe the Uppsala University sys- tem for WMT13, for English-to-German translation. We use the Docent decoder, a local search decoder that translates at the document level. We add tunable dis- tortion limits, that is, soft constraints on the maximum distortion allowed, to Do- cent. We also investigate cleaning of the noisy Common Crawl corpus. We show that we can use alignment-based filtering for cleaning with good results. Finally we investigate effects of corpus selection for recasing. 1 Introduction In this paper we present the Uppsala University submission to WMT 2013. We have submitted one system, for translation from English to German. In our submission we use the document-level de- coder Docent (Hardmeier et al., 2012; Hardmeier et al., 2013). In the current setup, we take advan- tage of Docent in that we introduce tunable dis- tortion limits, that is, modeling distortion limits as soft constraints instead of as hard constraints. In addition we perform experiments on corpus clean- ing. We investigate how the noisy Common Crawl corpus can be cleaned, and suggest an alignment- based cleaning method, which works well. We also investigate corpus selection for recasing. In Section 2 we introduce our decoder, Docent, followed by a general system description in Sec- tion 3. In Section 4 we describe our experiments with corpus cleaning, and in Section 5 we describe experiments with tunable distortion limits. In Sec- tion 6 we investigate corpus selection for recasing. In Section 7 we compare our results with Docent to results using Moses (Koehn et al., 2007). We conclude in Section 8. 2 The Docent Decoder Docent (Hardmeier et al., 2013) is a decoder for phrase-based SMT (Koehn et al., 2003). It differs from other publicly available decoders by its use of a different search algorithm that imposes fewer restrictions on the feature models that can be im- plemented. The most popular decoding algorithm for phrase-based SMT is the one described by Koehn et al. (2003), which has become known as stack decoding. It constructs output sentences bit by bit by appending phrase translations to an initially empty hypothesis. Complexity is kept in check, on the one hand, by a beam search approach that only expands the most promising hypotheses. On the other hand, a dynamic programming technique called hypothesis recombination exploits the lo- cality of the standard feature models, in particu- lar the n-gram language model, to achieve a loss- free reduction of the search space. While this de- coding approach delivers excellent search perfor- mance at a very reasonable speed, it limits the information available to the feature models to an n-gram window similar to a language model his- tory. In stack decoding, it is difficult to implement models with sentence-internal long-range depen- dencies and …

Content truncated to 500 words