Bootstrapping Phrase-based Statistical Machine Translation via WSD Integration

International Joint Conference on Natural Language Processing, pages 1042–1046, Nagoya, Japan, 14-18 October 2013. Bootstrapping Phrase-based Statistical Machine Translation via WSD Integration Hien Vu Huy†,♯ , Phuong-Thai Nguyen†,♯ , Tung-Lam Nguyen†,♯ and M.L Nguyen†,‡ † University of Engineering and Technology, VNU Hanoi {hienvuhuy, thainp, lamnt 52}@vnu.edu.vn ‡ Japan Advanced Institute of Science and Technology (JAIST) nguyenml@jaist.ac.jp Abstract Beside the word order problem, word choice is another major obstacle for ma- chine translation. Though phrase-based statistical machine translation (SMT) has an advantage of word choice based on lo- cal context, exploiting larger context is an interesting research topic. Recently, there have been a number of studies on inte- grating word sense disambiguation (WSD) into phrase-based SMT. The WSD score has been used as a feature of translation. In this paper, we will show that by bootstrap- ping WSD models using unlabeled data, we can bootstrap an SMT system. Our ex- periments on English-Vietnamese transla- tion showed that BLEU scores have been improved significantly. 1 Introduction Conventional phrase-based systems use local context information from phrase table and lan- guage model. Though phrase based SMT achieves a jump in translation quality in comparison with word based SMT, there are still cases in which local context cannot capture correctly the mean- ings of source words. WSD can use features from much larger contexts and those features can over- lap each other. The idea of integrating WSD into SMT rises naturally from this perspective. Previ- ously, Varea et al. (2001) directly used context sen- sitive lexical models, applying these models for re-ranking n-best for their word-based maximum entropy model (MEM) SMT and achieving slight improvements in translation quality. Chan et al. (2007) made use of WSD for hi- erarchical phrase-based translation for Chinese- English by utilizing two new WSD features for SMT and proposing an algorithm for scoring syn- chronous rules. Phrases which do not exceed a length of two were computed WSD models. Their experiments showed that WSD can improve SMT significantly. Simultaneously with Chan et al. (2007), Carpuat and Wu (2007) used a similar approach to the problem. The main difference was that they focused on conventional phrase-based SMT in Koehn et al. (2003) and used only one WSD fea- ture for SMT. The limit of phrase length was the same as the value used by their SMT system. Their experiments led to the same conclusion: WSD can improve SMT. However, approaches based on statistic fre- quently against deficiencies of parallel and spe- cific domain corpora. Only a few popular lan- guages are derived continuous financial support and interest of researchers. Therefore, it becomes an immense obstacle to apply these approaches for the remaining languages. Recently, there are several approaches to ad- dress this impediment. Ambati et al. (2011) ap- plied multi-strategy methods in active learning for machine translation by combining several tech- niques in sentence selection process. They at- tained significantly results while parallel training data was scarce. In this paper, we present our study on this topic. First, by integrating WSD …

Content truncated to 500 words