A Multi-Level Feature Fusion Framework for Spoken Language Understanding

Spoken language understanding(SLU) constitutes an essential component of task-based dialogue, primarily focusing on intent detection(ID) and slot filling(SF). Generally, researchers treat these two tasks as sentence classification and sequence labeling tasks. Although significant success has been achieved in this area, it seems that the overall accuracy has reached a bottleneck. Through empirical research, we demonstrate that many researchers neglect the information interaction between the two tasks when modeling them jointly, or that most of the carefully designed components do not achieve the expected effect. Based on this, we propose a multi-level feature fusion method for spoken language understanding, which consists of two basic components - the multi-level feature intent feature fusion module and the intentslot bridging unit. The multi-level feature intent module can integrate each token-level corpus information into the intent feature, while the intent-slot bridging unit can combine the sentence-level intent feature with the slot feature vector. With novel and simple design, these two modules can effectively fuse features. We conducted experiments on two classic datasets in the SLU field, and the experiments prove that our proposed method can greatly improve the overall accuracy and achieve state-of-the-art on various metrics.