Optimizing Machine Translation through Paraphrasing Ranking

ACL ARR 2024 April Submission179 Authors

Abstract: This paper proposes a novel approach for optimizing the performance of a machine translation system. By paraphrasing an input into multiple different phrases, that maintain the semantic meaning, and ranking them using only source-side information, we show that performance can be significantly improved. Experiments on the IWSLT En-De and En-Nl datasets show that the family of Flan-T5 models can be improved by several COMET points, a notable gain in performance. Furthermore, this can be combined with traditional output-side rankers on n-best list outputs to obtain further gains.

Paper Type: Short

Research Area: Machine Translation

Research Area Keywords: machine translation, ranking, reranking, paraphrasing

Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: english, german, dutch

Submission Number: 179