Improving the precision of RDF question/answering systems- A why not approach

Abstract

Given a natural language question qNL over an RDF dataset D, an RDF Question/Answering (Q/A) system first translates qNL into a SPARQL query graph Q and then evaluates Q over the underlying knowledge graph to figure out the answers Q(D). However, due to the challenge of understanding natural language questions and the complexity of linking phrases with specific RDF items (e.g., entities and predicates), the translated query graph Q may be incorrect, leading to some wrong or missing answers. In order to improve the system's precision, we propose a self-learning solution based on the users' feedback over Q(D). Specifically, our method automatically refines the SPARQL query Q into a new query graph Q0 with minimum modifications (over the original query Q). The new query will fix the errors and omissions of the query results. Furthermore, each amendment will also be used to improve the precision in answering subsequent natural language questions.