Chimera Model of Candidate Soups for Non-Autoregressive Translation

Non-Autoregressive Translation (NAT) models have drawn much attention because of their excellent decoding speed. However, NAT models suffer a significant drop in translation quality compared to Autoregressive Translation (AT) models. Candidate Soups (CandiSoups) is an effective method that can fully use the different candidate translations, significantly improving the translation quality for NAT models. However, it needs to use an additional AT model for re-scoring to achieve the best performance, which slows down its inference speed and takes up more computing resources. In this paper, we propose a Chimera Model framework of CandiSoups (CMCS), which can significantly accelerate inference speed while maintaining superior performance for CandiSoups. Specifically, by modifying the decoder, we fuse the AT and NAT models to construct a Chimera Model that can perform self-rescore. Moreover, we propose a novel adaptive training method to help train Chimera Models better. Experimental results on two major benchmarks demonstrate the effectiveness of our proposed approach, which can significantly improve translation quality while maintaining the excellent inference speed.