Multi-intent detection for task-oriented dialogue systems based on joint learning
Qian Zhang, Zesan Liu, Ziheng Wang, Xiaozhen Li · Fifth International Conference on Mechatronics and Computer Technology Engineering (MCTE 2022) · 2022
Dialogue system uses natural language as a medium to achieve friendly communication between humans and machines. The performance of intent detection is crucial to the effectiveness of a task-oriented dialogue system. When developing a task-oriented dialogue system in a specific domain, users typically express multiple intents in the same sentence. In this study, we compare two approaches for multi-intent detection at the sentence level, which aims to investigate multi-intent detection for task-oriented dialogue systems based on joint learning. Experiment results on the ATIS and MixATIS datasets show that the multi-classification approach improves slot prediction by combining relevant intent information, whereas the multi-label approach based on joint learning improves intent detection by making predictions at each possible intent.