Reranking by Voting Mechanism in Neural Machine Translation
Han Yang, Xiang Li, Shike Wang, Jinan Xu · 2022 4th International Conference on Natural Language Processing (ICNLP) · 2022
Reranking technique has demonstrated the effectiveness of further improving the performance in Neural Machine Translation (NMT). Current reranking methods focus on fusing different neural machine translation models. Although even higher quality can be achieved by combining multiple models by techniques. Training multiple models may bring extra time and space consumption. In this paper, we propose a new method that concentrates on a single model's N-Best candidates voting for each other with features produced by BERT (Bidirectional Encoder Representation from Transformers) and n-gram. The experimental results show that our method performs well on WMT'20 Chinese-English, WMT'09 Hungarian-English and CCMT'21 Chinese-Uyghur.