Incorporating Complete Syntactical Knowledge for Spoken Language Understanding

Spoken Language Processing (SLU) is important in taskoriented dialog systems. Intent detection and slot filling are two significant tasks of SLU. State-of-the-art methods for SLU jointly solve these two tasks in an end-to-end fashion using pre-trained language models like BERT. However, existing methods ignore the syntax knowledge and longrange word dependencies, which are essential supplements for semantic models. In this paper, we utilize the Graph Convolutional Networks (GCNs) and dependency tree to incorporate the syntactical knowledge. Meanwhile, we propose a novel gate mechanism to model the label of the dependency arcs. Therefore, the labels and geometric connection of dependency tree are both encoded. The proposed method can adaptively attach a weight on each dependency arc based on dependency types and word contexts, which avoids encoding redundant features. Extensive experimental results show that our model outperforms strong baselines.