Quality-Informed Segment-Level Error Correction Using Natural Language Explanations from xTower and Large Language Models

AbstractThis paper describes our submission to the WMT25 Automated Translation Quality Evaluation Systems Task 3 - QE-informed Segment-level Error Correction. We propose a two-step approach for Automatic Post-Editing (APE) that leverages natural language explanations of translation errors. Our method first utilises the xTower model to generate a descriptive explanation of the errors present in a machine-translated segment, given the source text, the machine translation, and quality estimation annotations. This explanation is then provided as a prompt to a powerful Large Language Model, Gemini 1.5 Pro, which generates the final, corrected translation. This approach is inspired by recent work in edit-based APE and aims to improve the interpretability and performance of APE systems. We Evaluated across six language pairs (EN→ZH, EN→CS, EN→IS, EN→JA, EN→RU, EN→UK), our approach demonstrates promising results, especially in cases requiring fine-grained edits.