Improving Machine Translation with Human Feedback: An Exploration of Quality Estimation as a Reward Model

TL;DR

This paper investigates using quality estimation (QE) as a reward model to guide human-preference-aligned feedback training for machine translation. It identifies a pervasive overoptimization problem when QE-based rewards diverge from human judgments and proposes RAFT+—a penalty-based adjustment for QE rewards—to mitigate this issue. Across high- and low-resource multilingual settings, the approach yields consistent BLEURT improvements and is validated by human evaluations, with strong data-efficiency demonstrated in monolingual feedback. The findings highlight that stronger, pretrained base models benefit more from QE-based feedback, while also outlining limitations related to error types, metric vulnerabilities, and resource-language detectors. The work suggests QE-based reward modeling as a promising, data-efficient path for improving MT with human-aligned feedback, while pointing to future work on broader QE granularity, detector reliability, and multilingual balance.

Abstract

Insufficient modeling of human preferences within the reward model is a major obstacle for leveraging human feedback to improve translation quality. Fortunately, quality estimation (QE), which predicts the quality of a given translation without reference, has achieved impressive alignment with human evaluations in the last two years. In this work, we investigate the potential of employing the QE model as the reward model to predict human preferences for feedback training. We first identify the overoptimization problem during QE-based feedback training, manifested as an increase in reward while translation quality declines. We examine the problem and argue that the vulnerability of the QE model might lead to high rewards for incorrect translations, resulting in overoptimization and error propagation. To address the problem, we adopt a simple yet effective method that uses heuristic rules to detect the incorrect translations and assigns a penalty term to the reward scores of them. Experimental results show that the proposed QE-based feedback training achieves consistent and significant improvements across various settings, further verified through human preference studies. Our subsequent analysis demonstrates the high data efficiency of the proposed QE-based feedback training: it outperforms systems using larger parallel corpora by a small amount of monolingual data. Our code is available at: https://github.com/zwhe99/FeedbackMT

Figures (9)

• Figure 1: The relationship between reward score and translation quality. Each point represents the average performance of a checkpoint on the development sets. Comet-qe-da is used as QE-based reward model. BLEURT is a translation quality metric that strongly correlates with human preferences freitag-etal-2022-results.
• Figure 2: Trends in length-ratio and off-target error rates with the training process. Even though these errors do not manifest significantly during the early and middle phases of training, they may still surge in later stages.
• Figure 3: Training curves under various settings. The metrics are average values for all language pairs on the development set. The QE-based reward model is Comet-qe-da.
• Figure 4: Human preference evaluation, comparing RAFT+ to SFT model on En$\Leftrightarrow$Zh test sets.
• Figure 5: Comparison between RAFT+ and continuous training in the low-resource setting.