Enhancing machine translation with quality estimation and reinforcement learning

Yang, Zijian Győző and Laki, László János (2023) Enhancing machine translation with quality estimation and reinforcement learning. Annales Mathematicae et Informaticae, 58.. pp. 182-190. ISSN 17876117

Preview

Text AMI_58_from182to190.pdf - Published Version Download (625kB) | Preview

Abstract

In recent times, our research has focused on training large language models and exploring their potential. With the emergence of ChatGPT, it has been demonstrated that it is possible to fine-tune language models in a task-agnostic way. The success of ChatGPT is attributed to the reinforcement learning method, which integrates human feedback into the language model fine-tuning process. As a part of our research, we initially adapted the method of reinforcement learning for a specific task, which is machine translation, respectively. In this paper, we propose a novel approach to enhance machine translation with reinforcement learning and quality estimation methods. Our proposed approach uses reinforcement learning to learn to adjust the machine translation output based on quality estimation feedback, with the goal of improving the overall translation quality. We evaluated our approach on the WMT09 dataset for English-Hungarian language pair. We conducted an analysis to show how our approach improves the quality of machine translation output. Our approach offers a promising avenue for enhancing the quality of machine translation and demonstrates the potential of utilizing reinforcement learning to improve other natural language processing tasks.

Actions (login required)

Edit Item