Human-in-the-loop Machine Translation with Large Language Model (MT Summit 2023)

Paper

1. Overview

In this study, we propose a human-in-the-loop pipeline that guides LLMs to produce customized outputs with revision instructions. The pipeline initiates by prompting the LLM to produce a draft translation, followed by the utilization of automatic retrieval or human feedback as supervision signals to enhance the LLM’s translation through in-context learning. The humanmachine interactions generated in this pipeline are also stored in an external database to expand the in-context retrieval database, enabling us to leverage human supervision in an offline setting. We evaluate the proposed pipeline using the GPT-3.5-turbo API on five domain-specific benchmarks for German-English translation. The results demonstrate the effectiveness of the pipeline in tailoring in-domain translations and improving translation performance compared to direct translation instructions. This work was featured in MT Summit 2023.

2. Feedback Collection

• Initial translation results are obtained via translation_base.py.
• Get TER-based generated Feedback via sacrebleu_patch/sacrebleu.py sacrebleu [ref] -m ter --ter-trace-file op.json < [hypo].
• Get In-context demonstrations by running retrieval.py [DATA_STORE_PATH] [TEST_SET_PATH]

3. In-context Refinement Translation Pipeline

• stage1：run translation_hil.py.
• stage2：run compare_hil.py.

*All results about the experiment are stored in data.

Citation

@inproceedings{yang-etal-2023-hilmt,
    title = "Human-in-the-loop Machine Translation with Large Language Model",
    author = "Yang, Xinyi  and
      Zhan, Runzhe  and
      Wong, Derek F.  and
      Wu, Junchao  and
      Chao, Lidia S.",
    booktitle = "Proceedings of Machine Translation Summit XIX Vol. 2: Users Track",
    month = sep,
    year = "2023",
    address = "Macau SAR, China",
    publisher = "Machine Translation Summit",
    url = "https://files.sciconf.cn/upload/file/20230827/20230827195133_32318.pdf",
    pages = "88--98",
}