tailieunhanh - Báo cáo khoa học: "Filtering Syntactic Constraints for Statistical Machine Translation"

Source language parse trees offer very useful but imperfect reordering constraints for statistical machine translation. A lot of effort has been made for soft applications of syntactic constraints. We alternatively propose the selective use of syntactic constraints. A classifier is built automatically to decide whether a node in the parse trees should be used as a reordering constraint or not. Using this information yields a BLEU point improvement over a full constraint-based system. | Filtering Syntactic Constraints for Statistical Machine Translation Hailong Cao and Eiichiro Sumita Language Translation Group MASTAR Project National Institute of Information and Communications Technology 3-5 Hikari-dai Seika-cho Soraku-gun Kyoto Japan 619-0289 hlcao @ Abstract Source language parse trees offer very useful but imperfect reordering constraints for statistical machine translation. A lot of effort has been made for soft applications of syntactic constraints. We alternatively propose the selective use of syntactic constraints. A classifier is built automatically to decide whether a node in the parse trees should be used as a reordering constraint or not. Using this information yields a BLEU point improvement over a full constraint-based system. 1 Introduction In statistical machine translation SMT the search problem is NP-hard if arbitrary reordering is allowed Knight 1999 . Therefore we need to restrict the possible reordering in an appropriate way for both efficiency and translation quality. The most widely used reordering constraints are IBM constraints Berger et al. 1996 ITG constraints Wu 1995 and syntactic constraints Yamada et al. 2000 Galley et al. 2004 Liu et al. 2006 Marcu et al. 2006 Zollmann and Venugopal 2006 and numerous others . Syntactic constraints can be imposed from the source side or target side. This work will focus on syntactic constraints from source parse trees. Linguistic parse trees can provide very useful reordering constraints for SMT. However they are far from perfect because of both parsing errors and the crossing of the constituents and formal phrases extracted from parallel training data. The key challenge is how to take advantage of the prior knowledge in the linguistic parse trees without affecting the strengths of formal phrases. Recent efforts attack this problem by using the constraints softly Cherry 2008 Marton and Resnik 2008 . In their methods a candidate translation gets an extra .