Hierarchical Reinforcement Learning for Adaptive Text Summarization

This study presents a novel approach to adaptive text summarization using hierarchical reinforcement learning. We develop a T5-based hierarchical summarizer with a level selector, implementing and comparing three reinforcement learning algorithms: Proximal Policy Optimization (PPO), Advantage Actor-Critic (A2C), and Soft Actor-Critic (SAC). Our system adapts summary length based on time constraints and is evaluated using ROUGE and BERTScore metrics. Experiments conducted on the CNN/DailyMail dataset illustrate the potential of this approach in balancing summary quality and generation speed. Results show that PPO achieves the highest ROUGE and BERTScores, while A2C demonstrates a better balance between quality and efficiency. The paper emphasizes the potential as well as challenges of employing reinforcement learning for adaptive summarization, paving the way for future research in this critical domain of natural language processing.