基于双语语料库的翻译等价对自动抽取

Automatic Extraction of Translational Equivalence Based on Bilingual Corpora

提出了一种利用双语语料库自动抽取多词翻译等价对的方法。首先利用N-gram模型获得候选翻译单元,然后根据统计同现计算候选等价对的翻译概率,并用贪心策略实现翻译等价对的自动抽取。在翻译概率的计算中对3种常用的统计同现测度进行了比较。实验表明,当语料规模较小时,对数似然比(Log Likelihood Ratio)测度对于翻译等价对的抽取具有较好的效果。与现有方法相比,该方法较好地解决了翻译等价对抽取中多词单元对应及间接相关问题。

This paper describes a method to acquire multi-word translational equivalences from English-Chinese parallel corpora. Translation candidates are firstly obtained using N-gram model. Then, an iterative algorithm is used to extract translation equivalences according to statistical translation measures. Three statistical translation measures: Dice coefficient, Phi-Square Coefficient and Log Likelihood Ratio are compared in experiments and it is proved that Log Likelihood Ratio works better when training corpus is small. Compared with previous works, the proposed method solves the difficulty of multi-word unit correspondences and the problem of indirect association. Experiments on real corpus produced very promising results.