cs.CLFeb 6, 2024

Linear-time Minimum Bayes Risk Decoding with Reference Aggregation

Authors: Jannis Vamvas, Rico Sennrich

Organizations: Department of Computational Linguistics, University of Zurich

Abstract

Minimum Bayes Risk (MBR) decoding is a text generation technique that has been shown to improve the quality of machine translations, but is expensive, even if a sampling-based approximation is used. Besides requiring a large number of sampled sequences, it requires the pairwise calculation of a utility metric, which has quadratic complexity. In this paper, we propose to approximate pairwise metric scores with scores calculated against aggregated reference representations. This changes the complexity of utility estimation from O(n2)O(n^2) to O(n)O(n), while empirically preserving most of the quality gains of MBR decoding. We release our source code at https://github.com/ZurichNLP/mbr

Figures & tables

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

CardsList
  1. Noisy-Channel Minimum Bayes Risk Decoding

    Jul 6, 2026Yusuke Sakai, Hidetaka Kamigaito, Taro WatanabeLanguage Model DecodingMBR Decoding

  2. Overfitting Mitigation via Singular Value Decomposition in Minimum Bayes Risk Decoding

    Sep 1, 2026Riza Setiawan Soetedjo, Yusuke Sakai, Hidetaka Kamigaito +2Language Model DecodingMBR Decoding