Improving Machine Reading Comprehension by Integrating External Knowledge through Attention Mechanism and Graph Neural Network

Extractive machine reading comprehension task as a branch of reading comprehension tasks that requires extracting appropriate text fragments from passages as answers to relevant questions. In this paper, we use ConceptNet as an external world knowledge source for the reading comprehension task, and generate knowledge graphs for each question and passage after retrieving knowledge from ConceptNet. The role of Graph neural network is to derive knowledge on a knowledge graph and integrate it into the context of the corresponding question and passage. For better integration with external knowledge and deeper understanding of semantic relationships between questions and passages, attention mechanisms are introduced to accomplish further integration.