Centroid-Based Efficient Minimum Bayes Risk Decoding

Minimum Bayes risk (MBR) decoding has achieved state-of-the-art translation performance using COMET, which is a neural metric that has a high correlation with human evaluation. However, MBR decoding requires quadratic time because it computes the expected score between a translation hypothesis and all reference translations. We propose centroid-based MBR (CBMBR) decoding to improve the speed of MBR decoding. Our method clusters reference translations in the feature space and then calculates the score using the centroids of each cluster. The experimental results demonstrate that our CBMBR not only improved the decoding speed of the expected score calculation by 5.7 times but also outperformed vanilla MBR decoding in terms of translation quality by up to 0.5 COMET% in the WMT'22 En↔Ja, En↔De, En↔Zh, and WMT'23 En↔Ja translation tasks. 1