Two stage machine translation fixes guided by detailed error spotting
Diagnose, Then Repair: A Two-Stage MQM-Guided Post-Editing Framework for Domain-Specific Machine Translation
Computation and LanguageArtificial Intelligence
Summary
Machine translation systems often make mistakes that are hard to fix automatically. This paper presents a method where one AI model first finds specific errors in translated text, and then a second AI model makes small, targeted fixes only where needed. This approach helps keep translations accurate and prevents unnecessary changes, improving overall quality. The authors tested their method on several languages and showed it works better than simply asking one model to fix everything at once.
What this means in practice
- •For language service providers: Improves post-editing accuracy by targeting specific translation errors detected automatically, enhancing domain-specific translation workflows.$Commercial implications: Enables delivery of higher-quality translations with less human rework, making it a sellable upgrade for translation service platforms.
- •For software localization teams: Supports efficient correction of translation outputs by highlighting precise error spans, reducing manual effort in software UI translation maintenance.
Authors
Ji Hun Wang, Siyu Wu
Abstract
LLM-based machine translation evaluation can closely match human judgments, but in practice it remains largely diagnostic, with the signals rarely translating into direct quality improvements under real production constraints. We propose a two-stage, evaluator-guided automatic post-editing framework that turns MQM-style evaluation into targeted repairs: a retrieval-augmented LLM evaluator outputs structured, span-level MQM diagnoses under an explicit edit contract, and a separate LLM post-editor applies minimal edits restricted to those diagnoses. This separation improves controllability and reduces paraphrastic drift compared to one-stage "judge-and-refine" baselines. In a systematic study involving seven LLMs spanning three model providers and seven languages, our best configuration consistently improves both COMET-22 and COMETKiwi scores over one-stage post-edit methods, while the evaluator's error spans and severities show strong agreement with human MQM annotations and human editor preferences.