Discriminative Reordering Extensions for Hierarchical Phrase-Based Machine Translation

Translate, Predict or Generate: Modeling Rich Morphology in Statistical Machine Translation Ahmed El Kholy and Nizar Habash Center for Computational Learning Systems, Columbia University 475 Riverside Drive New York, NY 10115 {akholy,habash}@ccls.columbia.edu Abstract We compare three methods of modeling morphological features in statistical ma- chine translation (SMT) from English to Arabic, a morphologically rich language. Features can be modeled as part of the core translation process mapping source tokens to target tokens. Alternatively these fea- tures can be generated using target mono- lingual context as part of a separate gen- eration (or post-translation inflection) step. Finally, the features can be predicted us- ing both source and target information in a separate step from translation and genera- tion. We focus on three morphological fea- tures that we demonstrate through a man- ual error analysis to be most problematic for English-Arabic SMT: gender, number and the determiner clitic. Our results show significant improvements over a state-of- the-art baseline (phrase-based SMT) of al- most 1% absolute BLEU on a medium size training set. Our best configuration models the determiner as part of core translation and predicts gender and number separately, and handles the rest of the features through generation. 1 Introduction Translation into English has been the focus of many research efforts in Statistical Machine Trans- lation (SMT). However, recently, translation into other languages has been receiving increasing at- tention, especially translation into morphologi- cally rich languages (Sarikaya and Deng, 2007; Elming and Habash, 2009; Yeniterzi and Oflazer, 2010). One of the main issues in SMT is the spar- sity of parallel data for many language pairs espe- c⃝ 2012 European Association for Machine Translation. cially when the source or target language is mor- phologically rich. Morphological richness comes with many challenges and the severity of these challenges increases when translating from a mor- phologically poor language to a morphologically richer language. In this paper, we address these challenges through different modeling methods.1 In our ap- proach, morphological features can be modeled as part of the core translation process mapping source tokens to target tokens. Alternatively these fea- tures can be generated using target monolingual context as part of a separate generation (or post- translation inflection) step. Finally, the features can be predicted using both source and target in- formation in a separate step before generation. We focus in our experiments on English-Arabic SMT and we work on three morphological features that we found, through a manual error analysis, to be most problematic for English-Arabic SMT: gen- der, number and the determiner clitic. Our results show improvements over a state-of-the-art baseline (phrase-based SMT) of almost 1% absolute BLEU on a medium size training set of 4M words. Our best configuration models the determiner as part of core translation, predicts gender and number fea- tures separately, and handles the rest of the fea- tures through generation. We test our approach on a blind test set and we got the same relative im- provements across the different systems. However, when scaling up the data set, …

Content truncated to 500 words