Construction and optimization of personalized human-computer interaction system based on pre-training large model

Traditional human-computer interaction (HCI) system relies on preset rules or general models, which is difficult to meet the diversity and dynamics of users' needs. Pre-training large model has become one of the core technologies of personalized HCI system with its powerful semantic representation ability. In this paper, a personalized human-computer interaction system based on pre-training large model is proposed, which realizes efficient and personalized interaction experience through key technologies such as dynamic user modeling, lightweight fine-tuning and multimodal interaction optimization. The system adopts hierarchical architecture design, including application layer, strategy layer, model layer and data layer, and can handle various interactive forms such as text, voice and vision. In the aspect of dynamic user modeling, by combining explicit feedback and implicit behavior data, using dual-channel coding network and sliding window mechanism, the user feature vector is updated in real time to capture the user portrait accurately. The lightweight fine-tuning technology combines the design of fixed universal layer and flexible personality layer with adaptive loss function to realize personalized adjustment while maintaining the universal ability of the model. Multi-modal interaction optimization adopts hierarchical fusion strategy, which improves the accuracy of modal fusion and reduces the response delay through cross-modal attention mechanism and dynamic weighting method. The experimental results show that the system has obvious advantages in customer service scenario, with personalized matching reaching 0.82, response delay reduced to 185 milliseconds, and user satisfaction increased to 89.3%. The comparison of multi-modal fusion results shows that the early fusion method achieves a F1 score of 75.0% in fusion accuracy and reduces the response time by 41%. Ablation experiments further verify the importance of dynamic user modeling, lightweight fine-tuning and reinforcement learning (RL) strategy to system performance. In addition, the system shows good delay control ability in high concurrency scenarios, which verifies the effectiveness of the optimization method.