基于语言理解的机器翻译译文自动评价方法研究

机器翻译译文质量的自动评价方法是推动机器翻译技术快速发展的一条重要途径。它不仅在机器翻译系统性能测量中发挥着重要作用，而且在系统开发中指导着特征参数的调整或规则的优化。针对目前机器翻译译文自动评价过程中不充分利用源语言句子的作用、汉语译文中汉字层面的语言知识难以获取与应用、深层次语义知识难以适当的引入等问题，本项目在以下方面进行深入的研究和探索：（1）研究面向译文自动评价的译文分析与理解方法，结合源语言句子进行译文分析，减少翻译错误给译文分析带来的困难，提高译文中近义词、句法、语义分析的准确率；（2）研究汉语译文汉字层面的语言知识获取与应用方法，在词语层面的译文分析结果上，探索利用马尔可夫逻辑网络迁移学习汉字层面的深层次语言知识；（3）研究融合词语、句法、语义一体化的机器翻译译文自动评价新框架，通过概率图结构降低计算复杂度。本项目最终目标是提高译文自动评价方法与人工评价的相关性。

Automatic evaluation of machine translation is critical to promote the rapid development of machine translation. It not only plays an important role on measuring the performance of machine translation, but also guides feature weights tuning or rules optimization in the development of machine translation system. Due to the source sentences is not being fully utilized, character-level knowledge of Chinese translation is difficult to obtain and exploit, and the deep semantic knowledge is difficult to be appropriately introduced in current automatic evaluation process, this project will investigate and delve deeply in the following areas: (1) study the analyzing and understanding approach that oriented automatic evaluation of machine translation, combine with the source sentence to analyze the translation in order to decrease the difficulty caused by translation error and improve the accuracy of synonym, syntactic, semantic analysis; (2) study the approach to obtain and exploit character-level knowledge of Chinese translation, explore the way used Markov logic network to transfer learn the charcter-level knowledge of Chinese translation based on the word-level analysis results of Chinese translation; (3) investigate the new framework of automatic evaluation of machine translation, that integrated lexical, syntactic, semantic knowledge into the new framework, reduce the computational complexity by the structure of the probabilitic graph. The final objective of this project is to improve effectively the correlation between the new metric and human judgment.