基于深度学习的机器译文质量估计方法研究

项目名称： 基于深度学习的机器译文质量估计方法研究

项目编号： No.61462044

项目类型： 地区科学基金项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 李茂西

作者单位： 江西师范大学

项目金额： 46万元

中文摘要： 作为一种新的译文自动评价方法，机器译文质量估计不仅能一定程度上替代传统方法自动评价译文质量，而且能提供一种全新的统计翻译系统特征权重优化方式。针对目前机器译文质量估计中特征提取严重依赖语言学分析、算法复杂度高、译文质量估计效果不够理想且这些严重制约着其在统计翻译系统自训练中的应用等不足，本项目将在以下方面进行深入的研究和探索：（1）研究新的机器译文质量估计模型，通过建立基于递归神经网络的机器译文逐层结构化表示和基于soft-max回归的译文质量估计模型，提高估计的效果。（2）不仅研究英语译文的质量估计方法，而且通过标注一定规模的汉语译文质量语料并结合汉语译文分词特点，对汉语译文的质量估计方法展开研究。（3）探索无需开发集的统计翻译系统自训练方法，结合解码时的知识和减少特征权重优化的迭代次数，将新的译文质量估计方法应用于翻译系统自训练中。最终建立机器译文质量估计的新框架，推动机器翻译的发展。

中文关键词： 统计机器翻译；机器译文质量估计；统计翻译系统自训练；递归神经网络；soft-max回归

英文摘要： As a new method for automatic evaluation of machine translation, quality estimation of machine translation not only can replace the traditional methods on automatically evaluating the translation quality to some extent, but also can provide a new weight optimization mode for statistical translation system. However, there exist some shortcomings in nowadays research on quality estimation of machine translation, such as it depend heavily on linguistic analysis and algorithm complexity is high in feature extraction, as well as its performance is not satisfactory, and all of which restricts severely its application in self-training of statistical translation system etc. Thus, this project will investigate and delve deeply in the following areas: (1) study a new model to quality estimation of machine translation, establish a layer by layer structured representation of machine translation based on recursive neural network and a translation quality estimated model based on soft-max regression to improve the effect of quality estimation. (2) not only investigate the approach of quality estimation for English translation, but also investigate the approach for Chinese translation by annotating a certain amount of translation quality of Chinese translation and combining the characteristics of Chinese word segmentation in Chinese translation.(3)investigate the self-training issue of statistical translation system without development set, achieve the optimization algorithm of feature weights by integrating the decoding knowledge of machine translation system and reducing the iteration times in feature weights optimization. Finally, we will establish a new framework for quality estimation of machine translation, which will promote the development of research on machine translation.

英文关键词： Statistical machine translation;Quality estimation of machine translation;Self-training of statistical translation system;Recursive neural network;Soft-max regression