Improving reordering performance using higher order and structural features

Proceedings of NAACL-HLT 2013, pages 315–324, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Improving reordering performance using higher order and structural features Mitesh M. Khapra IBM Research India mikhapra@in.ibm.com Ananthakrishnan Ramanathan IBM Research India anandr42@gmail.com Karthik Visweswariah IBM Research India v-karthik@in.ibm.com Abstract Recent work has shown that word aligned data can be used to learn a model for reordering source sentences to match the target order. This model learns the cost of putting a word immediately before another word and finds the best reordering by solving an instance of the Traveling Salesman Problem (TSP). However, for efficiently solving the TSP, the model is restricted to pairwise features which examine only a pair of words and their neighborhood. In this work, we go beyond these pairwise fea- tures and learn a model to rerank the n-best reorderings produced by the TSP model us- ing higher order and structural features which help in capturing longer range dependencies. In addition to using a more informative set of source side features, we also capture target side features indirectly by using the transla- tion score assigned to a reordering. Our exper- iments, involving Urdu-English, show that the proposed approach outperforms a state-of-the- art PBSMT system which uses the TSP model for reordering by 1.3 BLEU points, and a pub- licly available state-of-the-art MT system, Hi- ero, by 3 BLEU points. 1 Introduction Handling the differences in word orders between pairs of languages is crucial in producing good ma- chine translation. This is especially true for lan- guage pairs such as Urdu-English which have sig- nificantly different sentence structures. For exam- ple, the typical word order in Urdu is Subject Object Verb whereas the typical word order in English is Subject Verb Object. Phrase based systems (Koehn et al., 2003) rely on a lexicalized distortion model (Al-Onaizan and Papineni, 2006; Tillman, 2004) and the target language model to produce output words in the correct order. This is known to be in- adequate when the languages are very different in terms of word order (refer to Table 3 in Section 3). Pre-ordering source sentences while training and testing has become a popular approach in overcom- ing the word ordering challenge. Most techniques for pre-ordering (Collins et al., 2005; Wang et al., 2007; Ramanathan et al., 2009) depend on a high quality source language parser, which means these methods work only if the source language has a parser (this rules out many languages). Recent work (Visweswariah et al., 2011) has shown that it is pos- sible to learn a reordering model from a relatively small number of hand aligned sentences . This elim- inates the need of a source or target parser. In this work, we build upon the work of Visweswariah et al. (2011) which solves the reorder- ing problem by treating it as an instance of the Traveling Salesman Problem (TSP). They learn a model which assigns costs to all pairs of words in a sentence, where the cost represents the penalty of …

Content truncated to 500 words