Improved word alignment in patent domain

摘要

This paper presents a new method for word alignment in patent domain which incorporates both generative and discriminative models. In this framework, the advantage of generative model that can learn large numbers of parameters from a sentence-aligned parallel corpus automatically in a unsupervised way can be kept, as well as get an improvement through discriminative models which can deploy various features in a supervised way. Even with only 300 word-aligned Chinese-English sentence pairs, incorporates with a 1M parallel Chinese-English patent sentences released by NTCIR9, experiments show that our method can get a promising performance.