Chinese semantic dependency graph is extended from semantic dependency tree, which uses directed acyclic graphs to capture richer latent semantics of sentences. In this paper, we propose two approaches for Chinese semantic dependency graph parsing. In the first approach, we build a non-projective transition-based dependency parser with the SwAP-based algorithm. Then we use a classifier to add arc candidates generated by rules to the tree, forming a graph. In the second approach, we build a transition-based graph parser directly using a variant of the list-based transition system. For both approaches, neural networks are adopted to represent the parsing states. Both approaches yield significantly better results than the top systems in the SemEval-2016 Task 9: Chinese Semantic Dependency Parsing.
展开▼
