Maximizing LLMs Potential: Enhancing Mongolian Chinese Machine Translation with RL Agents and Adversarial Multi Knowledge Distillation

Xu Liuxu, YiLa Su, Nier Wu, Yatu Ji, Huinuan Zhang

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Large Language Models, Reinforcement Learning, Adversarial Knowledge Distillation, Mongolian Chinese Machine Translation

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose an innovative approach that leverages Reinforcement Learning agents to efficiently transfer knowledge from Large Language Models to low-resource translation models, significantly improving translation quality.

Abstract:

Despite the impressive performance of Large Language Models (LLMs) in Natural Language Processing (NLP), they still face challenges in low-resource translation tasks, particularly in Mongolian to Chinese machine translation, often yielding suboptimal results. To address this issue, we propose an innovative approach that combines multi-source knowledge distillation and incorporates Reinforcement Learning (RL) to help models acquire and transfer knowledge from LLMs more effectively. RL plays a crucial role in this, making dynamic decisions to determine useful information for low-resource translation models and how to extract it efficiently. We introduce a new reward function to comprehensively guide knowledge distillation, and experiments show that this approach harnesses the potential of LLMs, significantly improving translation quality in low-resource settings.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1959