引入基于主题复述知识的统计机器翻译模型

Improved statistical machine translation model with topic-based paraphrase

针对传统的基于双语平行语料的复述获取方法在复述获取和应用的过程中忽视文档上下文的缺点,引入基于主题模型的上下文信息来改善复述获取—主要致力于如何计算上下文无关的复述生成概率和上下文相关的复述生成概率.研究如何将上述2种概率融入统计机器翻译建模,以提高翻译系统的性能.多个测试集上的实验结果证明了该方法的有效性.

Abstract. To deal with the defect of the conventional parallel corpus based paraphrase extraction method which neglects document-level context, the paraphrase extraction and its application in statistical machine translation were improved by introducing the context based on topic model. The problem that how to better learn two kinds of paraphrase probabilities, topic-insensitive and topic-sensitive ones, was mainly analyzed. Both of the two probabilities can be incorporated into the modeling of statistical machine translation by using different methods. The experimental results on various test sets demonstrated the effectiveness of the approach.