Improving Neural Machine Translation with Neural Sentence Rewriting

Abstract

A complex expression is more difficult for machine translation than a simplified sentence. One efficient method to handle this problem is to rewrite the source text into a simplified version before translation while keeping its meaning. In this paper, we propose a novel method to automatically rewrite source sentences based on neural machine translation. We propose a round-trip machine translation method to automatically generate a large amount of high quality rewritten pairs from bilingual corpus and then build an end-to-end sentence rewriting system based on neural network. Experimental results on Chinese-English IWSLT translation tasks show that our method leads to significant improvements over a strong baseline system.