Intent detection model based on dual-channel feature fusion

Abstract

The intelligent customer service system is an important part of power marketing and an important window for the power sector's external services. Spoken language understanding(SLU) is the core module of the dialogue system. Intent detection is a key subtask of the spoken language understanding module, which directly determines whether the spoken language understanding module is correct. Understand the purpose of the user's words. In view of the current use of commonly used deep learning methods such as RNN and CNN to deal with the limited semantic representation and insufficient feature extraction caused by intent detection, this paper proposes an intent detection model based on dual-channel feature fusion (Dual-channel Model), using The capsule network captures rich text features, makes up for the lack of local spatial information in the pooling operation, and further improves the representation limitations of CNN and RNN. At the same time, it uses the high-level features extracted by the pooling operation and the capsule network to form a dual-channel feature extraction mechanism, and The feature fusion of the two channels is used for the final intent detection. Experiments show that on the ATIS and SMP2019-ECDT datasets, the proposed model has an accuracy of 98.1% and 96.25% in intent detection, and an F1 value of 88.5% and 89.76%.