Reinforcement Learning For Text Classification: An Evaluation Of Policy-Gradient Methods With Various Topologies

Official URL: https://risc01.sabanciuniv.edu/record=b3205818

Usage of reinforcement learning (RL) in natural language processing (NLP) tasks has gained momentum in recent years. In this thesis, we present an improved approach to the task of text classification through the integration of various deep learning topologies such as transformers and large language models (LLMs) into the feature extraction process within a reinforcement learning framework. In this proposed method, the RL policies are trained to observe a portion of the text and determine whether to classify the text or to proceed to the next part of the document. The policies were optimized with the REINFORCE (Williams, 1992) algorithm utilizing a designed reward signal. The effectiveness of the proposed method was evaluated and compared against other state-of-the-art models on standard text classification benchmark datasets, demonstrating the superiority of the proposed approach in terms of efficiency while losing little performance in accuracy. The results indicate that the use of the LLMs in the feature extraction process, coupled with RL policies with designed reward signals, provides a promising avenue for the development of effective and efficient text classification models.