Dependency-based Reordering Model for Constituent Pairs in Hierarchical SMT

Dependency-based Reordering Model for Constituent Pairs in Hierarchical SMT Arefeh Kazemi†, Antonio Toral⋆, Andy Way⋆, Amirhassan Monadjemi†, Mohammadali Nematbakhsh† † Department of Computer Engineering, University of Isfahan, Isfahan, Iran {akazemi,monadjemi,nematbakhsh}@eng.ui.ac.ir ⋆ ADAPT Centre, School of Computing, Dublin City University, Ireland {atoral,away}@computing.dcu.ie Abstract We propose a novel dependency-based re- ordering model for hierarchical SMT that predicts the translation order of two types of pairs of constituents of the source tree: head-dependent and dependent-dependent. Our model uses the dependency struc- ture of the source sentence to capture the medium- and long-distance reorder- ings between these pairs of constituents. We describe our reordering model in de- tail and then apply it to a language pair in which the languages involved follow dif- ferent word order patterns, English (SVO) and Farsi (free word order being SOV the most frequent pattern). Our model out- performs a baseline (standard hierarchical SMT) by 0.78 BLEU points absolute, sta- tistically significant at p = 0.01. 1 Introduction Reordering is a fundamental problem in machine translation (MT) that significantly affects trans- lation quality, especially between languages with major differences in word order. While a great deal of work has been carried out to address this prob- lem, none of the existing approaches can perform all the required types of reordering operations in a principled manner. In general, there are four main approaches to address the reordering problem in statistical ma- chine translation (SMT): distortion models, lexical phrase-based models, hierarchical phrase-based models and syntax-based models. Despite the rela- tive success of each of these approaches in improv- ing the overall performance of the SMT systems, they suffer from a number of shortcomings: c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. • Inability to capturing long-distance reorder- ing. Distortion and lexical phrase-based mod- els assign probability only to the adjacent word or phrase pairs, so they can only per- form local reordering between adjacent units and fail to capture long distance reorder- ing. This weakness has motivated research on tree-based models, such as the hierar- chical phrase-based model (HPB). Although HPB models outperform phrase-based mod- els (PB-SMT) on medium-range reordering, they still perform weakly on handling long distance reordering due to complexity con- straints. • Sparsity. Most of the approaches can perform the reordering of common words or phrases, but they usually cannot be generalized to un- seen patterns which have the same linguistic structure. For example, if the object follows the verb in the source language and preceeds the verb in the target language, we still need to see a particular instance of a verb and an object in the training data to be able to per- form reordering between them. • Context insensitivity. Lexical and hierarchi- cal phrase-based models determine the order- ing of the phrases based solely on the lexi- cal items in those phrases. However, a phrase might have different orderings in different contexts, so it is essential to include more context in order to …

Content truncated to 500 words