Context Based Entity Graph Convolutional Network for Multi-hop Reading Comprehension
Download as PDF
DOI: 10.23977/mcee2020.040
Author(s)
Lunhua Zhang, Xiaohong Liu
Corresponding Author
Lunhua Zhang
ABSTRACT
Multi-hop reading comprehension poses new challenge on machine reading comprehension tasks which requires reasoning between multiple documents. In this paper, we propose a graph-based model called Context based Entity Graph Convolutional Network (CEG). In order to take full advantage of context information, on the one hand we apply multi-granularity encodings for entity which capture rich context information, on the other hand we extract surrounding context of entity to enrich entity encoding. The surrounding context of entity will further be utilized by Memory Network to support graph reasoning. Experimental evaluation shows CEG achieves competitive performance on the QAngaroo WIKIHOP dataset, and the following ablation test demonstrates our proposed context extraction modules are effective in multi-hop reading comprehension.
KEYWORDS
Multi-hop reading comprehension, Graph convolutional network, Memory network, Multi-granularity encodings