# Adapting Large Language Models for Document-Level Machine Translation

**Authors:** Minghao Wu, Thuy-Trang Vu, Lizhen Qu, George Foster, Gholamreza Haffari  
**arXiv:** [2401.06468](https://arxiv.org/abs/2401.06468)  
**Date:** 2024-01-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.06468v1/

## Key points

- Moderately-sized LLMs often outperform larger models after task-specific fine-tuning in document-level machine translation (DocMT)
- Prompt strategies have an impact on downstream translation performance
- Extensive experiments were conducted using two fine-tuning methods, three LLM backbones, and 18 translation tasks across nine language pairs
- Specialized models tailored for DocMT can surpass the performance of GPT-4 in certain cases
- Challenges with off-target translations still exist despite exclusive fine-tuning on bilingual parallel documents
- Analysis of aspects such as translation errors, scaling law of parallel documents, out-of-domain generalization, and impact of zero-shot crosslingual transfer was conducted
- The research provides insights into the strengths and limitations of LLM-based DocMT models
- The authors' work is still ongoing and includes 21 pages with 14 tables and 7 figures.

## Summary

In their research titled "Adapting Large Language Models for Document-Level Machine Translation," authors Minghao Wu, Thuy-Trang Vu, Lizhen Qu, George Foster, and Gholamreza Haffari explore the process of adapting large language models (LLMs) for document-level machine translation (DocMT). They find that moderately-sized LLMs often outperform larger models after task-specific fine-tuning. The researchers investigate the impact of prompt strategies on downstream translation performance and conduct extensive experiments using two fine-tuning methods, three LLM backbones, and 18 translation tasks across nine language pairs. The findings reveal that specialized models tailored for DocMT can even surpass the performance of GPT-4 in certain cases. However, these models still face challenges with off-target translations despite being exclusively fine-tuned on bilingual parallel documents. To gain a deeper understanding of these LLMs tailored for DocMT, the authors analyze aspects such as translation errors, the scaling law of parallel documents, out-of-domain generalization, and the impact of zero-shot crosslingual transfer. This research not only sheds light on the strengths and limitations of LLM-based DocMT models but also provides a foundation for future research in this field. The authors' work is still in progress and includes 21 pages with 14 tables and 7 figures.