Non-Isomorphic Forest Pair Translation

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 440–450, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Non-isomorphic Forest Pair Translation Hui Zhang1, 2, 3 Min Zhang1 Haizhou Li1 Eng Siong Chng2 1Institute for Infocomm Research 2Nanyang Technological University 3USC Information Science Institute huizhang.fuan@gmail.com {mzhang, hli}@i2r.a-star.edu.sg aseschng@ntu.edu.sg Abstract This paper studies two issues, non-isomorphic structure translation and target syntactic structure usage, for statistical machine translation in the context of forest-based tree to tree sequence trans- lation. For the first issue, we propose a novel non-isomorphic translation framework to capture more non-isomorphic structure mappings than tra- ditional tree-based and tree-sequence-based trans- lation methods. For the second issue, we propose a parallel space searching method to generate hypo- thesis using tree-to-string model and evaluate its syntactic goodness using tree-to-tree/tree sequence model. This not only reduces the search complexity by merging spurious-ambiguity translation paths and solves the data sparseness issue in training, but also serves as a syntax-based target language mod- el for better grammatical generation. Experiment results on the benchmark data show our proposed two solutions are very effective, achieving signifi- cant performance improvement over baselines when applying to different translation models. 1 Introduction Recently syntax-based methods have achieved very promising results and attracted increasing interests in statistical machine translation (SMT) research com- munity due to their ability to provide informative context structure information and convenience in carrying out word transformation and sub-span reor- dering. Fundamentally, syntax-based SMT views translation as a structural transformation process. Generally speaking, from modeling viewpoint, a syntax-based model tries to convert the source struc- tures into target structures iteratively and recursively while from decoding viewpoint a syntax-based sys- tem segments an input tree/forest into many sub-fragments, translates each of them separately, combines the translated sub-fragments and then finds out the best combinations. Therefore, from bilingual viewpoint, we face two fundamental problems: the mapping between bilingual structures and the way of carrying out the target structures combination. For the first issue, a number of models have been proposed to model the structure mapping between tree and string (Galley et al., 2004; Liu et al., 2006; Yamada and Knight, 2001; DeNeefe and Knight, 2009) and between tree and tree (Eisner, 2003; Zhang et al., 2007 & 2008; Liu et al., 2009). How- ever, one of the major challenges is that all the cur- rent models only allow one-to-one mapping from one source frontier non-terminal node (Galley et al., 2004) to one target frontier non-terminal node in a bilingual translation rule. Therefore, all those translation equi- valents with one-to-many frontier non-terminal node mapping cannot be covered by the current state-of-the-art models. This may largely compro- mise the modeling ability of translation rules. For the second problem, currently, the combina- tion is driven by only the source side (both tree-to-string model and tree-to-tree model only check the source span compatibility when combining different target structures in decoding) or only the 440 target side (string to tree model). There is no well study in …

Content truncated to 500 words