A Word Reordering Model for Improved Machine Translation

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 486–496, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics A Word Reordering Model for Improved Machine Translation Karthik Visweswariah IBM Research India Bangalore, India v-karthik@in.ibm.com Rajakrishnan Rajkumar Dept. of Linguistics Ohio State University raja@ling.osu.edu Ankur Gandhe IBM Research India Bangalore, India ankugand@in.ibm.com Ananthakrishnan Ramanathan IBM Research India Bangalore, India aramana2@in.ibm.com Jiri Navratil IBM T.J. Watson Research Center Yorktown Heights, New York jiri@us.ibm.com Abstract Preordering of source side sentences has proved to be useful in improving statistical machine translation. Most work has used a parser in the source language along with rules to map the source language word order into the target language word order. The require- ment to have a source language parser is a ma- jor drawback, which we seek to overcome in this paper. Instead of using a parser and then using rules to order the source side sentence we learn a model that can directly reorder source side sentences to match target word or- der using a small parallel corpus with high- quality word alignments. Our model learns pairwise costs of a word immediately preced- ing another word. We use the Lin-Kernighan heuristic to find the best source reordering ef- ficiently during training and testing and show that it suffices to provide good quality reorder- ing. We show gains in translation performance based on our reordering model for translating from Hindi to English, Urdu to English (with a public dataset), and English to Hindi. For English to Hindi we show that our technique achieves better performance than a method that uses rules applied to the source side En- glish parse. 1 Introduction Languages differ in the way they order words to pro- duce sentences representing the same meaning. Ma- chine translation systems need to reorder words in the source sentence to produce fluent output in the target language that preserves the meaning of the source sentence. Current phrase based machine translation systems can capture short range reorderings via the phrase table. Even the capturing of these local reordering phenomena is constrained by the amount of training data available. For example, if adjectives precede nouns in the source language and follow nouns in the target language we still need to see a particular ad- jective noun pair in the parallel corpus to handle the reordering via the phrase table. Phrase based sys- tems also rely on the target side language model to produce the right target side order. This is known to be inadequate (Al-Onaizan and Papineni, 2006), and this inadequacy has spurred various attempts to overcome the problem of handling differing word order in languages. One approach is through distortion models, that try to model which reorderings are more likely than others. The simplest models just penalize long jumps in the source sentence when producing the target sentence. These models have also been generalized (Al-Onaizan and Papineni, 2006; Till- man, 2004) to allow for lexical dependencies on the source. While these …

Content truncated to 500 words