Syntax Based Reordering with Automatically Derived Rules for Improved Statistical Machine Translation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1119–1127, Beijing, August 2010 Syntax Based Reordering with Automatically Derived Rules for Improved Statistical Machine Translation Karthik Visweswariah IBM Research v-karthik@in.ibm.com Jiri Navratil IBM Research jiri@us.ibm.com Jeffrey Sorensen Google, Inc. sorenj@google.com Vijil Chenthamarakshan IBM Research vijil.e.c@in.ibm.com Nanda Kambhatla IBM Research kambhatla@in.ibm.com Abstract Syntax based reordering has been shown to be an effective way of handling word order differences between source and target languages in Statistical Machine Translation (SMT) systems. We present a simple, automatic method to learn rules that reorder source sentences to more closely match the target language word or- der using only a source side parse tree and automatically generated alignments. The resulting rules are applied to source lan- guage inputs as a pre-processing step and demonstrate significant improvements in SMT systems across a variety of lan- guages pairs including English to Hindi, English to Spanish and English to French as measured on a variety of internal test sets as well as a public test set. 1 Introduction Different languages arrange words in different or- ders, whether due to grammatical constraints or other conventions. Dealing with these word order permutations is one of the fundamental challenges of machine translation. Given an exceptionally large training corpus, a phrase-based system can learn these reordering on a case by case basis. But, if our systems are to generalize to phrases not seen in the training data, they must explicitly cap- ture and model these reorderings. However, per- mutations are difficult to model and impractical to search. Presently, approaches that handle reorderings typically model word and phrase movements via a distortion model and rely on the target language model to produce words in the right order. Early distortion models simply penalized longer jumps more than shorter jumps (Koehn et al., 2003) independent of the source or target phrases in question. Other models (Tillman, 2004), (Al-Onaizan and Papineni, 2006) generalize this to include lexical dependencies on the source. Another approach is to incorporate features, based on the target syntax, during modeling and decoding, and this is shown to be effective for var- ious language pairs (Yamada and Knight, 2001), (Zollmann and Venugopal, 2006). Hierarchical phrase-based decoding (Chiang, 2005) also al- lows for long range reordering without explic- itly modeling syntax. While these approaches have been shown to improve machine translation performance (Zollmann et al., 2008) they usually combine chart parsing with the decoding process, and are significantly more computationally inten- sive than phrase-based systems. A third approach, one that has proved to be useful for phrase-based SMT systems, is to re- order each source-side sentence using a set of rules applied to a parse tree of the source sen- tence. The goal of these rules is to make the word order of the source sentence more sim- ilar to the expected target sentence word or- der. With this approach, the reordering rules are applied before training and testing with an SMT system. The efficacy of these methods has been shown on …

Content truncated to 500 words