Source-primed Multi-turn Conversation Helps LLMs Translate Documents

This is the code to replicate the Source-primed Multi-turn Translation in the paper Source-primed Multi-turn Conversation Helps Large Language Models Translate Documents. [cite]

1. Installation:

bash scripts/start_vllm.sh

2. Prepare data

We get and parse the data from the WMT24 dataset (https://www2.statmt.org/wmt24/translation-task.html), you can also directly use our parsed data in wmt24_processed.

3. Run translation

If you need to use gpt models to translate the data, you should first set your OPENAI_API_KEY:

export OPENAI_API_KEY=your_api_key

Then you can run the following command to translate the data:

bash scripts/run_gpt.sh

If you don't need to use gpt models, you can run the following command to translate the data:

bash scripts/run.sh

'mturn_icl_context' means our source-primed multi-turn translation, 'mturn_icl' and 'mturn' is multi-turn without source-primed.

4. Evaluate the translation

If you need to evaluate through dBLEU, run the following command:

bash eval/eval.sh

If you need to evaluate through Comet, run the following command:

bash eval/eval_comet.sh

Citation

Please consider citing us if you use our materials.

@misc{hu2025sourceprimedmultiturnconversationhelps,
      title={Source-primed Multi-turn Conversation Helps Large Language Models Translate Documents}, 
      author={Hanxu Hu and Jannis Vamvas and Rico Sennrich},
      year={2025},
      eprint={2503.10494},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.10494}, 
}