SMU Data Science Review
Abstract
Like humans, large language models (LLMs) benefit from revision and refinement, especially for complex tasks requiring critical thinking. Inspired by human collaborative problem-solving, this study introduces a novel multi-agent workflow designed to enhance LLM translations from English to low-resource languages. Multi-Agent Translation Team (MATT) involves the collaboration of agents that are assigned specific roles, such as translator, evaluation coordinator, and various levels of editing, to refine the initial translation into the most desired version possible. The agents work collaboratively in an iterative loop until the translation loss meets a satisfactory threshold. It stands out from other multi-agent workflows by combining the strengths of LLMs and Google Translate (GT) to achieve higher translation quality. This approach shows promise in translating short sentences and long chunks from English to languages such as Vietnamese, Hindi, and Malayalam.
Recommended Citation
Peter, Anishka; Dang, Mai; Liu, Michael; Dominguez, Joaquin; and Lohia, Nibhrat
(2024)
"Multi-Agent Translation Team (MATT): Enhancing Low-Resource Language Translation through Multi-Agent Workflow,"
SMU Data Science Review: Vol. 8:
No.
3, Article 3.
Available at:
https://scholar.smu.edu/datasciencereview/vol8/iss3/3
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Included in
Data Science Commons, Language Interpretation and Translation Commons, South and Southeast Asian Languages and Societies Commons
