Human-centric Question Answering System over Multiple Different Knowledge Graphs

Abstract

The inception of Semantic Web including the Resource Description Framework (RDF) data model that provides a standard framework for publishing and sharing machine-under\-stand\-able data on the Web enables the development of intelligent, semantically-oriented systems. Up to date, thousands of linked RDF datasets, also known as Knowledge Graphs (KGs), have been constructed and are available on the Web. Therefore, there is a need for Question-Answering (QA) systems that provide users with appropriate utilization of existing KGs and provide detailed and summarizing answers to the users' questions. Yet, the diversity of posing questions and the heterogeneity of KGs' schemas make the process of querying KGs quite challenging. Moreover, a single KG often does not provide sufficient information for a variety of questions. In this work, we propose a methodology aiming at developing a QA system that can automatically construct KG queries, use information from multiple different KGs, combine obtained data, and handle conflicting information, summarize obtained data if suitable. To accomplish that, we introduce a set of methods for: aligning properties (determining degrees of equivalence) in different KGs; generating templates based on given question-SPARQL query pairs; and using generated templates for constructing specific SPARQL queries for answering newly asked questions. Besides usual/regular questions, the methods allow for asking questions that contain linguistic terms with imprecise meanings. They also allow for aggregating answers, generating linguistic summaries for some suitable questions, and handling conflicting information retrieved from multiple different KGs.