Noisy-Channel Minimum Bayes Risk Decoding

This paper addresses a directional mismatch in Minimum Bayes Risk decoding: hypotheses are scored against sampled pseudo-references, while common evaluation metrics such as BLEU and COMET are asymmetric. It proposes a noisy-channel decomposition with four interacting terms: hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior. According to the abstract, channel contributions differ across metrics but remain relatively consistent across tasks, offering a unified interpretation of existing MBR variants and metric-specific diagnostics. The authors further suggest, without results included in the supplied summary, that appropriately weighting these channels may improve upon standard MBR decoding.