Unsupervised Word Alignment by Agreement Under ITG Constraint

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1998–2004, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Unsupervised Word Alignment by Agreement Under ITG Constraint Hidetaka Kamigaito1 kamigaito@lr.pi.titech.ac.jp Akihiro Tamura2 akihiro.tamura@nict.go.jp Hiroya Takamura1 takamura@pi.titech.ac.jp Manabu Okumura1 oku@pi.titech.ac.jp Eiichiro Sumita2 eiichiro.sumita@nict.go.jp 1Tokyo Institute of Technology 2National Institute of Information and Communication Technology Abstract We propose a novel unsupervised word align- ment method that uses a constraint based on Inversion Transduction Grammar (ITG) parse trees to jointly unify two directional mod- els. Previous agreement methods are not helpful for locating alignments with long dis- tances because they do not use any syntactic structures. In contrast, the proposed method symmetrizes alignments in consideration of their structural coherence by using the ITG constraint softly in the posterior regulariza- tion framework (Ganchev et al., 2010). The ITG constraint is also compatible with word alignments that are not covered by ITG parse trees. Hence, the proposed method is ro- bust to ITG parse errors compared to other alignment methods that directly use an ITG model. Compared to the HMM (Vogel et al., 1996), IBM Model 4 (Brown et al., 1993), and the baseline agreement method (Ganchev et al., 2010), the experimental results show that the proposed method significantly im- proves alignment performance regarding the Japanese-English KFTT and BTEC corpus, and in translation evaluation, the proposed method shows comparable or statistical sig- nificantly better performance on the Japanese- English KFTT and IWSLT 2007 corpus. 1 Introduction Word alignment is an important component of sta- tistical machine translation (SMT) systems such as phrase-based SMT (Koehn et al., 2003) and hier- archical phrase-based SMT (Chiang, 2007). In ad- dition, word alignment is utilized for multi-lingual tasks other than SMT, such as bilingual lexicon ex- traction (Liu et al., 2013). The most conventional approaches to word alignment are the IBM models (Brown et al., 1993) and the HMM model (Vogel et al., 1996), which align each source word to a sin- gle target word (i.e., directional models). In these models, bidirectional word alignments are tradition- ally induced by combining the Viterbi alignments in each direction using heuristics (Och and Ney, 2003). Matusov et al. (2004) exploited a symmetrized pos- terior probability for bidirectional word alignments. In these methods, each directional model is indepen- dently trained. Previous researches have improved bidirectional word alignments by jointly training two directional models to agree with each other (Liang et al., 2006; Grac¸a et al., 2008; Ganchev et al., 2010). Such a constraint on the agreement in a training phase is one of the most effective approaches to word alignment. However, none of the previous agree- ment constraints have taken into account syntactic structures. Therefore, they have difficulty recover- ing the alignments with long distances, which fre- quently occur, especially in grammatically different language pairs. Some unsupervised word alignment models such as DeNero and Klein (2007) and Kondo et al. (2013), have been based on syntactic structures. In particu- lar, it has been proven that Inversion Transduction Grammar …

Content truncated to 500 words