# Unleashing the Power of ChatGPT for Translation: An Empirical Study

**Authors:** Yuan Gao, Ruili Wang, Feng Hou  
**arXiv:** [2304.02182](https://arxiv.org/abs/2304.02182)  
**Date:** 2023-04-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.02182v1/

## Key points

- ChatGPT has impressive capabilities in natural language understanding and generation.
- Machine translation relies heavily on language understanding and generation abilities.
- Translation prompts are used to assist machine translation with ChatGPT across a wide range of translations.
- With designed translation prompts, ChatGPT achieves comparable or superior performance to professional translation systems for high-resource language translations.
- However, it lags significantly behind for low-resource translations.
- ChatGPT outperforms professional systems when evaluated using multiple references.
- ChatGPT comprehends domain keywords and adjusts its output accordingly for accurate domain-specific translations.
- Few-shot prompts consistently improve performance across different base prompts.
- This empirical study provides evidence of ChatGPT's potential in machine translation tasks due to its effective understanding and generation abilities with appropriate prompts.

## Summary

In their paper titled "Unleashing the Power of ChatGPT for Translation: An Empirical Study," authors Yuan Gao, Ruili Wang, and Feng Hou explore how to enhance machine translation using ChatGPT. They highlight that ChatGPT has demonstrated impressive capabilities in natural language understanding and generation. Machine translation is a crucial task in natural language processing, relying heavily on language understanding and generation abilities. The researchers adopt various translation prompts across a wide range of translations to assist machine translation with ChatGPT. The experimental results reveal that when equipped with designed translation prompts, ChatGPT achieves comparable or even superior performance compared to professional translation systems for high-resource language translations. However, it lags significantly behind for low-resource translations. To evaluate the quality of translations, the authors use multiple references and find that ChatGPT outperforms professional systems. Furthermore, the researchers conduct experiments on domain-specific translations and observe that ChatGPT comprehends provided domain keywords and adjusts its output accordingly to produce accurate translations. Additionally, they perform few-shot prompts which consistently improve performance across different base prompts. Overall, this empirical study provides evidence that ChatGPT has great potential in machine translation tasks due to its ability to understand and generate translations effectively when supported by appropriate prompts. The findings demonstrate its effectiveness in producing accurate translations for both high-resource languages as well as domain-specific tasks with few-shot prompts.