Maximum entropy based Chinese-Japanese word alignment

Word alignment can be used for numerous applications in natural language processing, such as lexicography, machine translation and so on. In this paper, we present an efficient context-dependent word alignment model based on maximum entropy (ME) approach. All knowledge sources are treated as feature functions in this model, such as source words, POS information and bilingual dictionary. This approach allows easy integration of more context-dependent information. We perform experiments on a Chinese-Japanese parallel corpus and the results are compared with a manually produced reference alignment. We achieve good alignment accuracy in a very noisy environment using unsupervised train method. And the effects of different features are also evaluated.