Mitigating Metric Bias in Minimum Bayes Risk Decoding
This addresses a specific issue in machine translation decoding for researchers and practitioners, but it is incremental as it builds on existing MBR methods.
The paper tackles the problem of metric bias in Minimum Bayes Risk (MBR) decoding for machine translation, where using a single neural metric for decoding leads to overestimated quality scores compared to human ratings. The result shows that using an ensemble of utility metrics in MBR decoding mitigates this bias and outperforms single-metric approaches in human evaluations.
While Minimum Bayes Risk (MBR) decoding using metrics such as COMET or MetricX has outperformed traditional decoding methods such as greedy or beam search, it introduces a challenge we refer to as metric bias. As MBR decoding aims to produce translations that score highly according to a specific utility metric, this very process makes it impossible to use the same metric for both decoding and evaluation, as improvements might simply be due to reward hacking rather than reflecting real quality improvements. In this work we find that compared to human ratings, neural metrics not only overestimate the quality of MBR decoding when the same metric is used as the utility metric, but they also overestimate the quality of MBR/QE decoding with other neural utility metrics as well. We also show that the metric bias issue can be mitigated by using an ensemble of utility metrics during MBR decoding: human evaluations show that MBR decoding using an ensemble of utility metrics outperforms a single utility metric.