Knowledge Graph Post-Processing for QA Systems

Xiaofei Lin,Yuzhi Zhang

DOI: https://doi.org/10.1145/3467707.3467758

2021-04-23

Abstract:Constructed large-scale Chinese knowledge graph may have many problems when used. In the process of implementing a knowledge-based QA system, the paper finds that the knowledge graph has many defects reducing the accuracy of the QA system. These defects are 1) synonymous attributes and relationships, 2) ambiguous attributes, 3) lacking data needed for application, 4) numeric attributes are not uniform. Wrong results caused by these defects may be hard to avoid by existing algorithms. Paper solves the troubles from the perspective of the knowledge graph, not the system. Research first puts forward the viewpoint of application-driven knowledge graph post-processing and proposes a framework to post-process KGs. The framework is based on the feedback mechanism of the QA system. Our framework has four modules, including 1) merge synonymous attributes based on iteration, 2) divide one ambiguous attribute into multiple attribute s based on Dublin Core™ Metadata Element Set, 3) use knowledge graph completion technology to add critical data, 4) standardize non-standard numerical attributes based on the template. The post-processing of the knowledge graph is to modify the accuracy of the QA system and reduce the difficulties of implementation

Computer Science