Later-stage Minimum Bayes-Risk Decoding for Neural Machine Translation

Abstract

For extended periods of time, sequence generation models rely on beam search algorithm to generate output sequence. However, the correctness of beam search degrades when the a model is over-confident about a suboptimal prediction. In this paper, we propose to perform minimum Bayes-risk (MBR) decoding for some extra steps at a later stage. In order to speed up MBR decoding, we compute the Bayes risks on GPU in batch mode. In our experiments, we found that MBR reranking works with a large beam size. Later-stage MBR decoding is shown to outperform simple MBR reranking in machine translation tasks.

References in corpus (5)

• Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
• A Simple, Fast Diverse Decoding Algorithm for Neural Generation
• Learning to Decode for Future Success
• Noisy Parallel Approximate Decoding for Conditional Recurrent Language Model
• Trainable Greedy Decoding for Neural Machine Translation

Cited by in corpus (1)

• A Stable and Effective Learning Strategy for Trainable Greedy Decoding