The authors introduce a noisy-channel decomposition of Minimum Bayes Risk (MBR) decoding that incorporates bidirectional effects to address asymmetries in evaluation metrics like BLEU and COMET. This approach decomposes MBR into four interacting components: hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior.
- The decomposition provides a unified interpretation of existing MBR variants by isolating the contribution of each channel.
- Analysis reveals that channel-wise contributions exhibit distinct characteristics across metrics while remaining consistent across tasks.
- The study suggests that appropriate channel weighting may lead to improvements over original MBR decoding.
This framework enables metric- and task-specific interpretability, offering a path toward more robust text generation by accounting for directional effects in hypothesis selection.