# Prompting Large Language Model for Machine Translation: A Case Study

**Authors:** Biao Zhang, Barry Haddow, Alexandra Birch  
**arXiv:** [2301.07069](https://arxiv.org/abs/2301.07069)  
**Date:** 2023-01-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.07069v2/

## Key points

- Authors Biao Zhang, Barry Haddow, and Alexandra Birch focus on prompting in machine translation
- Systematic study on prompting strategies for translation, including prompt template and demonstration example selection
- Investigation of monolingual data use and exploration of cross-lingual, cross-domain, and sentence-to-document transfer learning in prompting
- Importance of number and quality of prompt examples on translation performance
- Certain features of prompt examples (e.g., semantic similarity) correlate with prompting performance
- Leveraging pseudo parallel prompt examples from monolingual data can enhance translation outcomes
- Knowledge transfer from different settings can improve machine translation performance
- Challenges and limitations in current prompting techniques for machine translation discussed
- Promising results in other tasks but obstacles remain for machine translation prompting advancements

## Summary

In their work titled "Prompting Large Language Model for Machine Translation: A Case Study," authors Biao Zhang, Barry Haddow, and Alexandra Birch address the under-explored area of prompting in machine translation. They conduct a systematic study on prompting strategies for translation, focusing on factors such as prompt template and demonstration example selection. Additionally, they investigate the use of monolingual data and explore the potential of cross-lingual, cross-domain, and sentence-to-document transfer learning in prompting. Through extensive experiments using GLM-130B as the testbed, the authors make several key findings. Firstly, they highlight the importance of both the number and quality of prompt examples in influencing translation performance. Suboptimal examples can lead to degraded translation quality. Secondly, they identify that certain features of prompt examples, such as semantic similarity, exhibit significant correlations with prompting performance. However, none of these correlations are deemed strong enough to be conclusive. Moreover, the authors demonstrate that leveraging pseudo parallel prompt examples constructed from monolingual data through zero-shot prompting can enhance translation outcomes. They also show that transferring knowledge from prompt examples selected in different settings can lead to improved performance in machine translation tasks. In their analysis of model outputs, the authors discuss various challenges and limitations that current prompting techniques still face. Despite showing promising results in other tasks, prompting for machine translation presents its own set of obstacles that need to be addressed for further advancements in this area. This study contributes valuable insights into enhancing machine translation through effective prompting strategies and lays a foundation for future research in this domain.