Improving reordering performance using higher order and structural features

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 features and learn a model to rerank the n-best reorderings produced by the TSP model using 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 translation score assigned to a reordering. Our experiments, involving Urdu-English, show that the proposed approach outperforms a state-of-Theart PBSMT system which uses the TSP model for reordering by 1.3 BLEU points, and a publicly available state-of-The-Art MT system, Hiero, by 3 BLEU points.