The University of Maryland Statistical Machine Translation System for the Fourth Workshop on Machine Translation

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 145–149, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics The University of Maryland Statistical Machine Translation System for the Fourth Workshop on Machine Translation Chris Dyer∗†, Hendra Setiawan†, Yuval Marton∗†, and Philip Resnik∗† †UMIACS Laboratory for Computational Linguistics and Information Processing ∗Department of Linguistics University of Maryland, College Park, MD 20742, USA {redpony,hendra,ymarton,resnik} AT umd.edu Abstract This paper describes the techniques we explored to improve the translation of news text in the German-English and Hungarian-English tracks of the WMT09 shared translation task. Beginning with a convention hierarchical phrase-based sys- tem, we found benefits for using word seg- mentation lattices as input, explicit gen- eration of beginning and end of sentence markers, minimum Bayes risk decoding, and incorporation of a feature scoring the alignment of function words in the hy- pothesized translation. We also explored the use of monolingual paraphrases to im- prove coverage, as well as co-training to improve the quality of the segmentation lattices used, but these did not lead to im- provements. 1 Introduction For the shared translation task of the Fourth Work- shop on Machine Translation (WMT09), we fo- cused on two tasks: German to English and Hun- garian to English translation. Despite belonging to different language families, German and Hungar- ian have three features in common that complicate translation into English: 1. productive compounding (especially of nouns), 2. rich inflectional morphology, 3. widespread mid- to long-range word order differences with respect to English. Since these phenomena are poorly addressed with conventional approaches to statistical machine translation, we chose to work primarily toward mitigating their negative effects when construct- ing our systems. This paper is structured as fol- lows. In Section 2 we describe the baseline model, Section 3 describes the various strategies we em- ployed to address the challenges just listed, and Section 4 summarizes the final translation system. 2 Baseline system Our translation system makes use of a hierarchical phrase-based translation model (Chiang, 2007), which we argue is a strong baseline for these language pairs. First, such a system makes use of lexical information when modeling reorder- ing (Lopez, 2008), which has previously been shown to be useful in German-to-English trans- lation (Koehn et al., 2008). Additionally, since the decoder is based on a CKY parser, it can con- sider all licensed reorderings of the input in poly- nomial time, and German and Hungarian may re- quire quite substantial reordering. Although such decoders and models have been common for sev- eral years, there have been no published results for these language pairs. The baseline system translates lowercased and tokenized source sentences into lowercased target sentences. The features used were the rule transla- tion relative frequency P(¯e| ¯f), the “lexical” trans- lation probabilities Plex(¯e| ¯f) and Plex( ¯f|¯e), a rule count, a target language word count, the target (English) language model P(eI 1), and a “pass- through” penalty for passing a source language word to the target side.1 The rule …

Content truncated to 500 words