Maximum Entropy Based Phrase Reordering for Hierarchical Phrase-Based Translation

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 555–563, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Maximum Entropy Based Phrase Reordering for Hierarchical Phrase-based Translation Zhongjun He Yao Meng Hao Yu Fujitsu R&D Center CO., LTD. 15/F, Tower A, Ocean International Center, 56 Dongsihuan Zhong Rd. Chaoyang District, Beijing, 100025, China {hezhongjun, mengyao, yu}@cn.fujitsu.com Abstract Hierarchical phrase-based (HPB) translation provides a powerful mechanism to capture both short and long distance phrase reorder- ings. However, the phrase reorderings lack of contextual information in conventional HPB systems. This paper proposes a context- dependent phrase reordering approach that uses the maximum entropy (MaxEnt) model to help the HPB decoder select appropriate re- ordering patterns. We classify translation rules into several reordering patterns, and build a MaxEnt model for each pattern based on var- ious contextual features. We integrate the MaxEnt models into the HPB model. Ex- perimental results show that our approach achieves significant improvements over a stan- dard HPB system on large-scale translation tasks. On Chinese-to-English translation, the absolute improvements in BLEU (case- insensitive) range from 1.2 to 2.1. 1 Introduction The hierarchical phrase-based (HPB) model (Chi- ang, 2005; Chiang, 2007) has been widely adopted in statistical machine translation (SMT). It utilizes synchronous context free grammar (SCFG) rules to perform translation. Typically, there are three types of rules (see Table 1): phrasal rule, a phrase pair consisting of consecutive words; hierarchical rule, a hierarchical phrase pair consisting of both words and variables; and glue rule, which is used to merge phrases serially. Phrasal rule captures short distance reorderings within phrases, while hierar- chical rule captures long distance reorderings be- Type Constituent Examples Word Variable PR √ - X → ⟨��, one of⟩ HR √ √ X → ⟨X�, ofX⟩ GR - √ S → ⟨SX, SX⟩ Table 1: A classification of grammar rules for the HPB model. PR = phrasal rule, HR = hierarchical rule, GR = glue rule. tween phrases. Therefore, the HPB model outper- forms conventional phrase-based models on phrase reorderings. However, HPB translation suffers from a limita- tion, in that the phrase reorderings lack of contex- tual information, such as the surrounding words of a phrase and the content of sub-phrases that rep- resented by variables. Consider the following two hierarchical rules in translating a Chinese sentence into English: X → ⟨X1 � X2, X1 ’s X2⟩ (1) X → ⟨X1 � X2, X2X1⟩ (2) � ��� � �� with Russia ’s talks talks with Russia Both pattern-match the source sentence, but pro- duce quite different phrase reorderings. The first rule generates a monotone translation, while the sec- ond rule swaps the source phrases covered by X1 and X2 on the target side. During decoding, the first 555 rule is more likely to be used, as it occurs more fre- quently in a training corpus. However, the exam- ple is not a noun possessive case because the sub- phrase covered by X1 is not a noun but a preposi- tional …

Content truncated to 500 words