Reordering Model for Forest-to-String Machine Translation

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 227–232, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Reordering Model for Forest-to-String Machine Translation Martin ˇCmejrek IBM Watson Group Prague, Czech Republic martin.cmejrek@us.ibm.com Abstract In this paper, we present a novel exten- sion of a forest-to-string machine transla- tion system with a reordering model. We predict reordering probabilities for every pair of source words with a model using features observed from the input parse for- est. Our approach naturally deals with the ambiguity present in the input parse forest, but, at the same time, takes into account only the parts of the input forest used by the current translation hypothesis. The method provides improvement from 0.6 up to 1.0 point measured by (Ter− Bleu)/2 metric. 1 Introduction Various commonly adopted statistical machine translation (SMT) approaches differ in the amount of linguistic knowledge present in the rules they employ. Phrase-based (Koehn et al., 2003) models are strong in lexical coverage in local contexts, and use external models to score reordering op- tions (Tillman, 2004; Koehn et al., 2005). Hierarchical models (Chiang, 2005) use lexi- calized synchronous context-free grammar rules to produce local reorderings. The grammatical- ity of their output can be improved by addi- tional reordering models scoring permutations of the source words. Reordering model can be either used for source pre-ordering (Tromble and Eisner, ), integrated into decoding via translation rules ex- tension (Hayashi et al., 2010), additional lexical features (He et al., ), or using external sources of information, such as source syntactic features ob- served from a parse tree (Huang et al., 2013). Tree-to-string (T2S) models (Liu et al., 2006; Galley et al., 2006) use rules with syntactic struc- tures, aiming at even more grammatically appro- priate reorderings. Forest-to-string (F2S) systems (Mi et al., 2008; Mi and Huang, 2008) use source syntactic forest as the input to overcome parsing errors, and to al- leviate sparseness of translation rules. The parse forest may often represent several meanings for an ambiguous input that may need to be transtated differently using different word or- derings. The following example of an ambiguous Chinese sentence with ambiguous part-of-speech labeling motivates our interest in the reordering model for the F2S translation. S. tˇaol`un (0) SSS. h`ui (1) SSS zˇenmey`ang (2) discussion/NN SS meeting/NN how/VV discuss/VV SSSSSwill/VV There are several possible meanings based on the different POS tagging sequences. We present translations for two of them, together with the in- dices to their original source words: (a) NN NN VV: How2 was2 the0 discussion0 meeting1? (b) VV VV VV: Discuss0 what2 will1 happen1. A T2S system starts from a single parse corre- sponding to one of the possible POS sequences, the same tree can be used to predict word reorder- ings. On the other hand, a F2S system deals with the ambiguity through exploring translation hy- potheses for all competing parses representing the different meanings. As our example suggests, dif- ferent meanings also tend to reorder differently 227 …

Content truncated to 500 words