Parse-realize based paraphrasing and SMT corpus enriching

Abstract:

To resolve the low-coverage problem of the statistic machine translation training corpus, a dependency parsing and sentence realization based paraphrasing method is proposed. The input sentence is first parsed into a dependency tree, and then the tree is realized into multiple natural language sentences. Although the generated sentences have the same lexical words, the expressions of word orders are re-arranged. The experiments shows that the paraphrasing method can be used to enlarge the bilingual corpus for statistic machine translation and the method efficiently relieves the low-coverage problem of training corpora without any extra resources, finally the translation quality is improved. 