Advancing Paraphrase Generation through Deep Reinforcement Learning

This study addresses the challenge of generating high-quality paraphrases, a complex task in Natural Language Processing (NLP) that encompasses various essential subproblems. While neural networks have driven advancements, this study investigates the use of Deep Reinforcement Learning (DRL) to enhance paraphrase quality. Initially, a comparison of state-of-the-art supervised text generation approaches is conducted, revealing their limitations. Subsequently, strategies for knowledge transfer between supervised models and RL models are assessed using two intermediary environments. The impact of RL fine-tuning with different reward functions is then analyzed, highlighting the difficulties in designing effective reward functions for paraphrase generation. Findings suggest that employing an adversarial model as a reward function yields the best results. Additionally, the study proposes a general framework for transferring knowledge from supervised models to RL models, facilitating training efficiency. This work not only advances paraphrase generation but also offers a robust pipeline for applying RL in NLP applications.