Unlocking Fine-Grained Translation Quality Estimation in LRMs through Mutually Boosting Implicit and Explicit Reasoning

EMNLP 2026 (main conference)

This repository contains the data-processing code and training-launch scripts for the paper above. All paths are anonymized as /anonymous/path/... — update them before running.

Overview

The experimental pipeline consists of three parts:

1. Data preprocessing (preprocess_data/): parse raw WMT 2022 / 2023 QE and GeneralMT 2024 evaluation files (sentence-level scores, word-level OK/BAD tags, MQM error marks) into unified JSON records with character-level error spans and scores.
2. Fine-grained subtask construction (subtask_data/): build the "explicit reasoning" training data of the paper — semantic-unit segmentation of translations, unit/sentence-level error-span detection, and MINOR/MAJOR severity classification — then aggregate the subtask samples into SFT-ready instruction–answer datasets.
3. Training (llama-factory-sft-bash.sh, verl-grpo-bash.sh): LoRA SFT of Qwen3-4B-Thinking with LLaMA-Factory, followed by GRPO RL training with verl using char-level subtask rewards, mutually boosting implicit and explicit quality-estimation reasoning.

Data Sources

All data comes from the official WMT QE shared-task and MQM human-evaluation repositories:

• WMT-QE-Task/wmt-qe-2022-data
• WMT-QE-Task/wmt-qe-2023-data
• WMT-QE-Task/wmt-qe-2024-data
• google/wmt-mqm-human-evaluation

Repository Layout

QE/
├── preprocess_data/                 # raw evaluation files → unified JSON
├── subtask_data/                    # fine-grained subtask data builders
│   ├── semanticUnit/                #   semantic-unit segmentation
│   ├── errorDetection/              #   error-span detection (unit / sequential)
│   ├── errorSeverity/               #   MINOR/MAJOR severity classification
│   ├── allTask/                     #   joint detection + severity in one pass
│   ├── generateLFReasoningData.py   #   LLM-filtered (COT) reasoning data
│   ├── aggregateAndShuffle.py       #   per-language-pair aggregation
│   └── mixed_aggregateAndShuffle.py #   cross-language aggregation
├── llama-factory-sft-bash.sh        # LoRA SFT launcher (LLaMA-Factory)
└── verl-grpo-bash.sh                # GRPO RL launcher (verl + vLLM)

Config blocks (language pair, input/output paths, sample sizes) are hardcoded at the top of each script — review them before running.

Usage

1. Download the datasets listed above, then run the scripts in preprocess_data/ to obtain the unified JSON records.
2. Run the builders in subtask_data/, then the aggregation scripts to produce the SFT dataset registered in LLaMA-Factory's dataset_info.json.
3. Launch SFT via llama-factory-sft-bash.sh, then GRPO RL via verl-grpo-bash.sh (expects train.parquet/test.parquet under the verl data directory and the custom char-level subtask reward functions under verl's utils/reward_score/).

Citation

@misc{dang2026unlockingfinegrainedtranslationquality,
      title={Unlocking Fine-Grained Translation Quality Estimation in LRMs through Mutually Boosting Implicit and Explicit Reasoning}, 
      author={Renfei Dang and Xinye Wang and Zhejian Lai and Weilu Xu and Shimin Tao and Daimeng Wei and Min Zhang and Shujian Huang},
      year={2026},
      eprint={2605.31378},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.31378}, 
}