基于树到串模型强化的层次短语机器翻译解码方法  

On Decoding with Augmented Hierarchical Phrase-Based Translation Models Using Tree-to-String Models

作　　者：肖桐[1] 朱靖波[1] 

机构地区：[1]东北大学信息科学与工程学院,沈阳110819

出　　处：《计算机学报》2016年第4期808-821,共14页Chinese Journal of Computers

基　　金：国家自然科学基金(61300097;61272376;61432013);中国博士后科学基金(2013M530131)资助

摘　　要：文中研究了利用树到串模型对层次短语模型进行强化的统计机器翻译解码方法.其基本框架是把层次短语模型作为基础模型,而把树到串模型作为层次短语模型的补充,增加翻译推导空间大小.文中重点研究了在该框架下的统计机器翻译解码技术,并提出了多种解码策略,包括基于树的精确解码策略、基于树的模糊解码策略和基于串的解码策略.通过NIST汉英翻译任务上的实验结果显示,文中所研究的方法可以十分有效地提升基线层次短语系统的翻译性能,比如在newswire和web数据上分别提高了1.3和1.2个BLEU点.此外,文中分析了若干影响翻译性能的因素,并给出了对比实验结果.We study decoding methods to augment a hierarchical phrase-based Machine Translation（MT）system with a tree-to-string model in this paper.In this framework the hierarchical phrasebased model is regarded as the base model,and the tree-to-string model is employed to enlarge the derivation space.In particular,we present several decoding strategies,including tree-based exact decoding,tree-based fuzzy decoding and string-based decoding.We experiment with our approach in a state-of-the-art MT system on the NIST MT evaluation data.Experimental results show that it outperforms a strong baseline over 1.3 and 1.2BLEU points on the newswire and web data respectively.Moreover,we show a systematic comparison of several factors that affect the translation quality.

关 键 词：统计机器翻译 层次短语模型 树到串模型 规则抽取 解码 社会媒体 社交网络 自然语言处理 机器翻译 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]