IndicQE-APE: A Benchmark for Quality Estimation and Automatic Post-Editing for Indic Languages
2026-08-17 • Computation and Language
Computation and Language
AI summaryⓘ
The authors combined data from several years of translation quality tests into one big dataset called indicqe, which includes multiple languages and types of quality information. They tested several language models and quality metrics to see how well they estimate translation quality and fix errors automatically. Their results show that confusing quality signals between whole sentences and individual words indicate worse quality. They also found that measures of agreement between human reviewers matter less when compared carefully. The new dataset and tools will be shared for others to use.
quality estimationautomatic post-editingtranslationlanguage pairsdirect assessmentword-level tagsprompted large language modelsCOMET metricsfew-shot promptingannotator disagreement
Authors
Diptesh Kanojia, Archchana Sindhujan, Sourabh Deoghare, Daria Sokova, Shenbin Qian, Girish Koushik, Tharindu Ranasinghe, Constantin Orăsan, Chrysoula Zerva, Ricardo Rei, Frédéric Blain, André F. T. Martins, Marco Turchi, Matteo Negri, Rajen Chatterjee, Anoop Kunchukuttan, Mitesh M. Khapra, Pushpak Bhattacharyya
Abstract
Indic quality estimation (QE) and automatic post-editing (APE) data is spread across separate releases, so no single resource supports training and evaluation across tasks and language pairs on one footing. We consolidate the WMT 2020--2024 shared-task lineage with an extended English--Malayalam resource into \indicqe: $126{,}754$ instances over nine directional pairs, with up to four label types aligned on the same segment, a direct assessment, a human post-edit, word-level OK/BAD tags and an error explanation, and a test set stratified over four difficulty axes. On it, we benchmark six prompted LLMs and three COMET metrics on segment-level QE, and three systems on APE. Two of the axes are defined partly on the direct assessment and select a compressed slice of it, so each axis is compared against a control drawn from the same language pair with the same score distribution. Only one survives that control: segments whose holistic and token-level quality signals conflict are ranked worse than equally-scored segments of the same language, for all nine systems and all seven pairs that carry the axis. Annotator disagreement, which looks second-hardest without the control, has no effect with it. Few-shot prompting costs every model $\leq$ $3.4$B both correlation and output-format compliance. Within-language accuracy does not make scores comparable across pairs: of the three trained metrics, the one with the best within-language correlation loses most when the pairs are pooled. The benchmark and code will be released.