# ParroT: Translating During Chat Using Large Language Models

**Authors:** Wenxiang Jiao, Jen-tse Huang, Wenxuan Wang, Xing Wang, Shuming Shi, Zhaopeng Tu  
**arXiv:** [2304.02426](https://arxiv.org/abs/2304.02426)  
**Date:** 2023-04-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.02426v1/

## Key points

- The paper explores the capabilities of large language models (LLMs) in natural language processing tasks, focusing on machine translation during chat interactions.
- Authors highlight limitations of restricted APIs for further research and progress in the field.
- They propose the ParroT framework utilizing open-sourced LLMs like LLaMA-7b and human-written translation data for evaluation.
- ParroT introduces a novel approach by reformulating translation data into an instruction-following style and incorporating a "Hint" field to guide the translation process effectively.
- Experimental results show that employing translation instructions and error-guided instructions enhance vanilla LLMs' translation performance.
- ParroT models maintain proficiency in general tasks with multi-task datasets during fine-tuning.
- The ParroT framework presents a promising solution for leveraging open-sourced LLMs to enhance real-time translation capabilities during chat interactions while addressing limitations in accessing advanced language models through restricted APIs.

## Summary

The paper "ParroT: Translating During Chat Using Large Language Models" by Wenxiang Jiao, Jen-tse Huang, Wenxuan Wang, Xing Wang, Shuming Shi, and Zhaopeng Tu explores the capabilities of large language models (LLMs) in natural language processing tasks. The authors focus on machine translation during chat interactions and highlight the limitations of restricted APIs for further research and progress in the field. To address this issue, they propose the ParroT framework which utilizes open-sourced LLMs such as LLaMA-7b and human-written translation data for evaluation. This framework introduces a novel approach by reformulating translation data into an instruction-following style and incorporating a "Hint" field to guide the translation process effectively. The authors provide experimental results that demonstrate the effectiveness of employing translation instructions and error-guided instructions in enhancing vanilla LLMs' translation performance. They also show that ParroT models can maintain their proficiency in general tasks with the inclusion of multi-task datasets during fine-tuning. Overall, the ParroT framework presents a promising solution for leveraging open-sourced LLMs to enhance real-time translation capabilities during chat interactions while addressing existing limitations in accessing advanced language models through restricted APIs.