Mitigating Metric Bias in Minimum Bayes Risk Decoding

TL;DR

This paper addresses the problem of metric bias in Minimum Bayes Risk decoding for machine translation, demonstrating that using an ensemble of metrics can mitigate bias and improve translation quality according to human evaluations.

Contribution

It introduces the concept of metric bias in MBR decoding and proposes an ensemble approach to reduce bias, validated by human evaluation results.

Findings

Neural metrics overestimate MBR translation quality compared to human ratings.

Using an ensemble of utility metrics during MBR decoding reduces metric bias.

Ensemble-based MBR decoding outperforms single-metric approaches in human evaluations.

Abstract

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.