KSRL: Knowledge Selection Based Reinforcement Learning for Knowledge-Grounded Dialogue

In the domain of multi-turn knowledge-grounded dialogues, the sequential coherence among knowledge elements chosen across various conversational turns presents potential cues for knowledge selection. However, this aspect has been largely overlooked in preceding studies. To tackle this issue, the present study introduces an innovative methodology that employs reinforcement learning to enhance knowledge selection in open-domain dialogue systems. By recasting the knowledge selection challenge as a sequential decision-making task and implementing reinforcement learning, the dialogue system is capable of discerning which knowledge to choose based on the conversational context and preceding dialogue turns, thereby generating high-quality responses. The system acquires a reward signal contingent upon the quality of the generated responses and subsequently updates its policy to maximize the expected reward over time. Harnessing the capabilities of reinforcement learning, our proposed method effectively learns to identify the most pertinent knowledge, thereby generating superior-quality responses. The study assesses the proposed approach using multiple open-domain dialogue datasets, demonstrating that it surpasses the performance of prior methodologies.