High Quality Rather than High Model Probability: Minimum Bayes Risk Decoding with Neural Metrics
This work addresses the problem of improving translation quality for users of neural machine translation systems by proposing an alternative decoding strategy, though it is incremental as it builds on existing MBR methods with new metrics.
The paper challenges the assumption that high model probability correlates with high translation quality in neural machine translation, showing that they only vaguely correlate. It demonstrates that using Minimum Bayes Risk decoding with neural metrics like BLEURT significantly improves human evaluations, even though these translations have lower model likelihood and perform worse on surface metrics like BLEU.
In Neural Machine Translation, it is typically assumed that the sentence with the highest estimated probability should also be the translation with the highest quality as measured by humans. In this work, we question this assumption and show that model estimates and translation quality only vaguely correlate. We apply Minimum Bayes Risk (MBR) decoding on unbiased samples to optimize diverse automated metrics of translation quality as an alternative inference strategy to beam search. Instead of targeting the hypotheses with the highest model probability, MBR decoding extracts the hypotheses with the highest estimated quality. Our experiments show that the combination of a neural translation model with a neural reference-based metric, BLEURT, results in significant improvement in human evaluations. This improvement is obtained with translations different from classical beam-search output: these translations have much lower model likelihood and are less favored by surface metrics like BLEU.