How to Design Translation Prompts for ChatGPT: An Empirical Study

Yuan Gao, Ruili Wang, Feng Hou

Research output: Chapter in Book/Conference proceedingConference contributionpeer-review

Abstract

ChatGPT, a chatbot based on the GPT models, has demonstrated surprising abilities in natural language understanding and generation tasks. Given that machine translation heavily relies on these abilities, there is substantial promise in applying ChatGPT for machine translation. Using naive prompts cannot fully unleash ChatGPT’s translation ability. Thus, in this paper, we propose several translation prompts that contain (i). translation task information (e.g., English-to-German), (ii). context domain information (e.g., News), (iii). Part-of-Speech (POS) tags, respectively. Our experimental results show that our proposed translation prompts can significantly enhance ChatGPT’s translation performance. We evaluate the translation quality using multi-reference test sets which consist of ten different human translations for each source sentence, and ChatGPT achieves superior performance compared to commercial systems. In addition, we also develop few-shot prompts upon our proposed translation prompts, which consistently show improvement across different translation directions.

Original languageEnglish
Title of host publicationProceedings of the 6th ACM International Conference on Multimedia in Asia Workshops, MMAsia 2024 Workshops
PublisherAssociation for Computing Machinery, Inc
ISBN (Electronic)9798400713149
DOIs
Publication statusPublished - 26 Dec 2024
Externally publishedYes
Event6th ACM International Conference on Multimedia in Asia Workshops, MMAsia 2024 Workshops - Auckland, New Zealand
Duration: 3 Dec 20246 Dec 2024

Publication series

NameProceedings of the 6th ACM International Conference on Multimedia in Asia Workshops, MMAsia 2024 Workshops

Conference

Conference6th ACM International Conference on Multimedia in Asia Workshops, MMAsia 2024 Workshops
Country/TerritoryNew Zealand
CityAuckland
Period3/12/246/12/24

Keywords

  • ChatGPT
  • Neural Machine Translation
  • Prompt Design

ASJC Scopus subject areas

  • Human-Computer Interaction
  • Computer Graphics and Computer-Aided Design

Fingerprint

Dive into the research topics of 'How to Design Translation Prompts for ChatGPT: An Empirical Study'. Together they form a unique fingerprint.

Cite this