Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

CHAI for LLMs: Improving Code-Mixed Translation in Large Language Models through Reinforcement Learning with AI Feedback

Created by
  • Haebom

Author

Wenbo Zhang, Aditya Majumdar, Amulya Yadav

Outline

In this paper, we propose CHAI, a novel framework for improving the code-mixed language understanding capability of multilingual large-scale language models (LLMs). To solve the problem that existing multilingual LLMs are not effective in code-mixed language translation tasks, we perform accurate annotation generation using LLMs, preference data generation through reinforcement learning, and experimental evaluation. CHAI shows 25.66% better performance than the state-of-the-art open source LLMs in code-mixed translation tasks.

Takeaways, Limitations

•
Takeaways:
◦
We present the possibility of annotating mixed-language translation tasks using LLM and improving their performance based on reinforcement learning.
◦
The CHAI framework has been experimentally proven to be effective in improving the code-mixing language processing capability of multilingual LLMs.
◦
An important first step toward the development of an LLM that more comprehensively addresses code-mixing languages.
•
Limitations:
◦
Further research is needed on the generalization ability of the CHAI framework.
◦
Verification of scalability across a variety of code-mixing languages and tasks is required.
◦
There are limitations to performance evaluation methods that rely on human evaluators.
👍
Made with Slashpage