# Iterative Translation Refinement with Large Language Models

**Authors:** Pinzhen Chen, Zhicheng Guo, Barry Haddow, Kenneth Heafield  
**arXiv:** [2306.03856](https://arxiv.org/abs/2306.03856)  
**Date:** 2023-06-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.03856v1/

## Key points

- Large language models have impressive capabilities in understanding instructions and performing natural language tasks.
- The authors propose a method called iterative translation refinement that uses large language models to enhance translation and post-editing processes.
- Incorporating a large language model into an iterative process improves the quality of translations beyond mere translation.
- Extensive testing using GPT-3.5 shows that iterations may decrease string-based metric scores, but neural metrics indicate comparable or improved translation quality.
- Human evaluations confirm that this approach effectively reduces "translationese" compared to initial GPT translations and even human references, especially for into-English directions.
- Ablation studies emphasize the importance of anchoring the refinement process to the source input and ensuring a reasonable initial translation.
- Leveraging large language models can lead to more natural translations and better post-editing outcomes.

## Summary

, , , , 

Large language models have shown impressive capabilities in understanding instructions and performing natural language tasks. In this paper, the authors propose a method called iterative translation refinement that utilizes large language models to enhance the quality of translation and post-editing processes. By incorporating a large language model into an iterative process, the resulting translations surpass mere translation. Extensive testing using GPT-3.5 reveals that while iterations may decrease string-based metric scores, neural metrics indicate comparable or even improved translation quality. Human evaluations further confirm that this approach effectively reduces "translationese" compared to initial GPT translations and even human references, particularly for into-English directions. Ablation studies emphasize the importance of anchoring the refinement process to the source input and ensuring a reasonable initial translation. These findings highlight how leveraging large language models can lead to more natural translations and better post-editing outcomes.