Dependency-Gated Cascade Biaffine Network for Chinese Semantic Dependency Graph Parsing

The Chinese Semantic Dependency Graph (CSDG) parsing breaks the limitation of the syntactic or semantic tree structure dependency system with a richer representation ability to express more complex language phenomena and semantic relationships. Most of the existing CSDG parsing systems used transition-based approach. It needs to define a complex transition system and its performance depends heavily on whether the model can properly represent the transition state. In this paper, we adopt neural graph-based approach which using Biaffine network to solve the CSDG parsing task. Furthermore, considering that dependency edge and label have the strong relationship, we design an effective dependency-gated cascade mechanism to improve the accuracy of dependency label prediction. We test our system on the SemEval-2016 Task 9 dataset. Experiment result shows that our model achieves state-of-the-art performance with 7.48% and 6.36% labeled F1-score improvement compared to the previous best model in TEXTBOOKS and NEWS domain respectively.