Enhancing Machine Comprehension Using Multi-Knowledge Bases and Offline Answer Span Improving System

Enhancing Machine Comprehension Using Multi-Knowledge Bases and Offline Answer Span Improving System

Machine Reading Comprehension (MRC) is a challenging but meaningful task in natural language processing (NLP) that requires us to teach a machine to read and understand a given passage and answer questions related to that passage. In this paper, we present a rich knowledge-enhanced reader (RKE-Reader), a hierarchical MRC model that employs double knowledge bases with an NER system as its knowledge enhancement unit. Besides, we are the first to propose an offline answer-imporving method to help model to determine the uncertain answer without extra online training process. Our experimental results indicate that on most datasets, the RKE-Reader significantly outperforms most of the published models that do not have knowledge base, especially on datasets that need commonsense reasoning. And the ablation study also reflects that external knowledge bases and answer-selecting unit do make a positive contribution in the entire model.