Enriching Parallel Corpora for Statistical Machine Translation with Semantic Negation Rephrasing

This paper presents an approach to improving performance of statistical machine translation by automatically creating new training data for difficult to translate phenomena. In particular this contribution is targeted towards tackling the poor performance of a state-of-the-art system on negated sentences. The corpus expansion is achieved by high quality rephrasing of existing sentences to their negated counterparts making use of semantic transfer. The method is designed to work on both sides of the parallel corpus while preserving the alignment. Our results show an overall improvement of 0.16 BLEU points, with a statistically significant increase of 1.63 BLEU points when tested on only negated test data.