Machine Translation Hallucination Detection for Low and High Resource Languages using Large Language Models

Overview

This repository presents our work for the paper "Machine Translation Hallucination Detection for Low and High Resource Languages using Large Language Models". The paper evaluates hallucination detection approaches using Large Language Models (LLMs) and semantic similarity within massively multilingual embeddings, addressing a critical challenge in machine translation. Code will be added later.

Abstract

Recent advancements in massively multilingual machine translation systems have significantly enhanced translation accuracy; however, even the best performing systems still generate hallucinations, severely impacting user trust. Detecting hallucinations in Machine Translation (MT) remains a critical challenge, particularly since existing methods excel with High-Resource Languages (HRLs) but exhibit substantial limitations when applied to Low-Resource Languages (LRLs). This study spans 16 language directions, covering HRLs and LRLs with diverse scripts, and finds that the choice of model is essential for performance.

Key Findings

• For HRLs, Llama3-70B outperforms the previous state of the art by as much as 0.16 MCC (Matthews Correlation Coefficient).

• For LRLs, Claude Sonnet shows superior performance on average by 0.03 MCC compared to other LLMs.

• LLMs can achieve performance comparable or even better than previously proposed models despite not being explicitly trained for any machine translation task, though their advantage is less significant for LRLs.

Evaluation reproducibility script sample

pip install -r requirements.txt
python main.py --input df_full_embeds.csv --load_embed True --embed_methods cohere mistral

Citation

If you use this repository or refer to our work, please cite our paper as follows:

@article{gongas2024machine,
  title={Machine Translation Hallucination Detection for Low and High Resource Languages using Large Language Models},
  author={Laura Gongas et al.},
  journal={arXiv preprint arXiv:2407.16470},
  year={2024}
}

Contact

For further inquiries or collaboration opportunities, please contact us at [kenza.benkirane.23@ucl.ac.uk] and [laura.gkogka.23@ucl.ac.uk].