Educational Agent Interaction Algorithm Based on Reinforcement Learning: Intent Recognition and Response Generation Optimization of Adaptive Question Answering System

Faced with the problems of difficult identification of learners’ personalized needs in educational scenarios, low response accuracy of the question and answer system, and unstable interactive experience, this paper proposes an educational agent interaction algorithm based on reinforcement learning (RL), which realizes intent recognition and response generation optimization through an adaptive question and answer mechanism. The methods include the following: (1) This article builds an intention recognition model that integrates semantic understanding and attention mechanisms, and performs feature annotation and semantic vector encoding on learner input; (2) This article designs a strategy optimization module based on Deep Q-Network (DQN) to adjust the agent’s response strategy through dynamic reward and punishment functions; (3) This article introduces a multi-round context memory mechanism to enable the system to achieve long-term intention tracking and knowledge transfer in continuous interactions; (4) This article uses a Generative Adversarial Network (GAN) structure to optimize the quality of response generation, and improves the semantic consistency of the answer through discriminator feedback. The experiment was verified on the EduQA and STU-Learn public data sets, and the average intent recognition accuracy reached 93.5%, and the average response generation consistency score was 92.7%. The results show that the system has significant adaptive and stable performance in complex educational question and answer tasks.