Research and Implementation of Task-oriented Dialogue Systems Based on Large Language Models and Vector Retrieval
He Rong, Jiamin Gao, Jiawei Guo, Chuan Lin · 2024
This paper focuses on task-oriented dialogue systems, introducing the differences between them and open-domain dialogue systems as well as the existing research methods. In view of the challenges faced by traditional task-oriented dialogue systems, such as the loss of context and poor interpretability. A task-oriented dialogue system based on Language Models and Vector Retrieval is proposed, which includes main modules like dialogue policy, intent recognition, entity extraction and natural language generation. The functions and working processes of each module are elaborated in detail. Through a series of comparative experiments, the advantages and disadvantages of this system compared with traditional dialogue systems are evaluated from multiple aspects such as response time, accuracy and token consumption. The experimental results show that the task-oriented dialogue proposed in this paper has achieved significant improvements in response time, accuracy and resource utilization.