Syntax-aware Phrase-based Statistical Machine Translation: System Description

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 292–297, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Syntax-aware Phrase-based Statistical Machine Translation: System Description Ulrich Germann University of Toronto Toronto, Ontario, Canada germann@cs.toronto.edu Abstract We present a variant of phrase-based SMT that uses source-side parsing and a constituent re- ordering model based on word alignments in the word-aligned training corpus to predict hi- erarchical block-wise reordering of the input. Multiple possible translation orders are rep- resented compactly in a source order lattice. This source order lattice is then annotated with phrase-level translations to form a lattice of to- kens in the target language. Various feature functions are combined in a log-linear fashion to evaluate paths through that lattice. 1 Introduction Dealing with word order differences is one of the major challenges in automatic translation between human languages. With its moderate context sen- sitivity and reliance on n-gram language models, phrase-based statistical machine translation (PB- SMT) (Koehn et al., 2003) is usually quite good at performing small word order changes — for instance, the inversion of adjective and noun in English-to-French translation and vice versa. How- ever, it regularly fails to execute word order changes over long distances, as they are required, for exam- ple, to accommodate the substantial differences in the word order in subordinate clauses between Ger- man and English, or to cope with the phenomenon of the “sentence bracket” (Satzklammer) in German main clauses, in which the finite part of the verb complex and additional elements (separable pre- fixes, participles, infinitives, etc.) form a bracket that encloses most of the arguments and other adverbial constituents, as shown in Fig. 1. In order to keep de- coding complexity in check, phrase-based decoders such as the Moses system (Koehn et al., 2007) rou- tinely limit the maximum distance for word order changes to six or seven word positions, thus rul- ing out, a priori, word order changes necessary to achieve good and fluent translations. As is generally acknowledged, word order dif- ferences are not entirely arbitrary. By and large they follow syntactic structure. An analysis of word-aligned French-English data by Fox (2002) showed that word alignment links rarely cross syn- tactic boundaries. Wu’s (1997) Inversion Transac- tion Grammar (ITG), assumes that word order dif- ferences can be accounted for by hierarchical inver- sion of adjacent blocks of text. Yamada and Knight (2001) present a stochastic model for transforming English parse trees into Japanese word sequences within a source-channel framework for Japanese- to-English translation. Collins et al. (2005) per- form heuristic word re-ordering from German into English word order based on German parse trees with a particular focus on the aforementioned drastic word order differences between German and English clause structure. Building on Chiang (2007), several systems under active development (e.g., Weese et al., 2011; Dyer et al., 2010) rely on synchronous context-free gram- mars to deal with word order differences. In essence, these systems parse the input while synchronously building a parse tree in the translation …

Content truncated to 500 words