基于层次短语的统计翻译系统中规则冗余的高效约束方法.pdfVIP

基于层次短语的统计翻译系统中规则冗余的高效约束方法 方李成 宗成庆 中科院自动化研究所模式识别国家重点实验室 北京 100190 {lcfang, cqzong}@nlpr.ia.ac.cn 摘 要： 基于层次短语的统计机器翻译模型是近年来比较流行且翻译质量较好的一种模型。层次短语翻译 系统有效地将同步上下文无关文法的重排序能力构建于成熟的普通短语翻译系统之上，得到了在重排序和捕 捉上下文信息方面都具有优势的模型。然而，层次短语翻译系统在计算复杂度方面远高出普通短语翻译系统， 使用的规则存在大量的冗余。本文分析了基于层次短语的翻译系统的规则冗余问题，提出了一种基于重排序 分割点的约束方法，使得学习重排序规则的训练过程集中在训练语料中重排序真实发生的片段。实验证明这 种方法大幅度减少了规则数量和解码时间，且使训练时间减少了一个量级，而翻译质量仅有微小损失，并保 持了基于层次短语的翻译系统和普通短语翻译系统相比翻译质量的优势。 关键词：统计机器翻译，层次短语，同步上下文无关文法，重排序分割点，重排序 An Efficient Constraint to Reduce the Redundancy of Rules in Hierarchical Phrase-Based Translation Systems FANG Licheng, ZONG Chengqing National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing 100190 {lcfang, cqzong}@nlpr.ia.ac.cn Abstract: Hierarchical phrase-based translation model is a popular statistical translation model which yields high quality translation by combining the reordering power of synchronous context free grammars and the proved wisdom of conventional phrase-based translation models. However, the model suffers from a significant high computational cost and a large redundancy of rules compared with conventional phrase-based systems. This paper analyzes the rule redundancy in hierarchical phrase-based systems, and proposes a rift-based constraint that forces the rules with reordering power to focus on where reordering has actually happens. Experimental results show that our method greatly reduces the number of rules extracted and used in the system, the decoding time, and reduces the training time by an order of magnitude. The sacrifice in translation quality is little and the advantage over conventional phrase-based systems is maintained. key words: Statistical machine translation, hierarchical phrases, synchronous context free grammars, rift, reordering 1