Unsupervised translation disambiguation based on Equivalent Pseudo Translation model

This paper describes an unsupervised translation disambiguation method based on the Equivalent Pseudo Translation (EPT). EPT is constructed by using non-ambiguous words of target language, which is semantically equivalent to the source ambiguous words. Sense-tagged examples are automatically extracted from a large scale Chinese corpus, by which a semantic classifier of EPT is formed. In order to apply the EPT classifier, English examples are mapped into a set of Chinese words by Hownet. This method is evaluated on the Senseval-2 framework of English lexical sample task, achieving the top performance against all other previous works which extract sense-tagged examples automatically. The results are very close to the state-of-the-art comparable unsupervised systems.