Towards Collaborative Neural-Symbolic Graph Semantic Parsing via Uncertainty

Abstract

Recent work in task-independent graph semantic parsing has shifted from grammar-based symbolic approaches to data-intensive neural approaches, and has shown strong performance on different types of meaning representations. However, it is still unclear that what are the limitations of these neural parsers, and whether these limitations can be compensated by collaborating with symbolic parsers. In this paper, we attempt to answer these questions by taking English Resource Grammar (ERG) parsing as a case study. Specifically, we first develop a state-of-the-art neural ERG parser, and then conduct detailed analyses on fine-grained linguistic phenomena. The results suggest that the neural parser's performance degrades significantly on long-tail examples, while the symbolic parser performs more robustly. To address this, we further propose a collaborative neural-symbolic semantic parsing framework. Specifically, we improve the beam search strategy by designing a decision criterion that incorporates both the model uncertainty about the testing data distribution and the prior knowledge from a symbolic parser. Experimental results show that this collaborative parsing framework can outperform the single neural parser and concretely improve the model's performance on long-tail examples.

Tasks

Reproductions