# Document-Level Machine Translation with Large Language Models

**Authors:** Longyue Wang, Chenyang Lyu, Tianbo Ji, Zhirui Zhang, Dian Yu, Shuming Shi, Zhaopeng Tu  
**arXiv:** [2304.02210](https://arxiv.org/abs/2304.02210)  
**Date:** 2023-04-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.02210v1/

## Key points

- Wang et al. evaluate Large Language Models (LLMs) like Chat-GPT and GPT-4 in document-level machine translation tasks
- Evaluation focuses on impact of discourse-aware prompts, comparison of translation models, and analysis of discourse modeling abilities
- Leveraging long-text modeling capabilities of Chat-GPT results in outperforming commercial MT systems in human evaluation
- GPT-4 demonstrates strong ability to explain discourse knowledge despite occasional errors in selecting translation candidates during contrastive testing
- Both Chat-GPT and GPT-4 show superior performance and potential as a promising paradigm for document-level translation
- Study highlights challenges and opportunities of discourse modeling for LLMs
- Suggests future exploration of more challenging discourse-aware NLP tasks like the GuoFeng Benchmark by Wang et al. (2023)
- Research contributes valuable insights into capabilities of LLMs in handling complex language tasks at a document level, showcasing potential for further advancements in natural language processing technologies

## Summary

In their study, Wang et al. evaluate the performance of Large Language Models (LLMs) such as Chat-GPT and GPT-4 in document-level machine translation tasks. The evaluation focuses on three main aspects: the impact of discourse-aware prompts, comparison of translation models, and analysis of discourse modeling abilities. They find that leveraging the long-text modeling capabilities of Chat-GPT results in outperforming commercial MT systems in human evaluation. Additionally, GPT-4 demonstrates a strong ability to explain discourse knowledge, despite occasional errors in selecting translation candidates during contrastive testing. Overall, both Chat-GPT and GPT-4 show superior performance and potential to become a promising paradigm for document-level translation. The study highlights the challenges and opportunities of discourse modeling for LLMs and suggests future exploration of more challenging discourse-aware NLP tasks like the GuoFeng Benchmark by Wang et al. (2023). This research contributes valuable insights into the capabilities of LLMs in handling complex language tasks at a document level, showcasing their potential for further advancements in natural language processing technologies.