Bidirectional Information Transfer Scheme for Joint Intent Detection and Slot Filling
Rui Feng Sun, Lu Rao, Xingfa Zhou · 2021 17th International Conference on Computational Intelligence and Security (CIS) · 2021
Intent detection and slot filling are two crucial basic tasks in natural language understanding. No matter treat intent detection and slot filling as two separate tasks, or train the two tasks as a joint model using deep learning frameworks, most of the approaches do not completely establish correlation between the intent and slots. Recent joint models build relationship between the two tasks via sharing intermediate representation of network, but we suggest that specific label information of one task is more beneficial to improve the performance of another task. Therefore, a novel bidirectional information transfer model is proposed to build sufficient interaction between intent detection and slot filling, which utilizes more explicit label information extracted from top layer of network. Besides, we introduce a type-aware mechanism to learn discriminative feature from the label information. Experimental results show that our model achieves a better performance on ATIS and SNIPS dataset in most of criteria, and significantly outperforms previous models.