Learning SPARQL Queries from Expected Results

Abstract

We present LSQ, an algorithm learning SPARQL queries from a subset of expected results. The algorithm leverages grouping, aggregates and inline values of SPARQL 1.1 in order to move most of the complex computations to a SPARQL endpoint. It operates by building and testing hypotheses expressed as SPARQL queries and uses active learning to collect a small number of learning examples from the user. We provide an open-source implementation and an on-line interface to test the algorithm. In the experimental evaluation, we use real queries posed in the past to the official DBpedia SPARQL endpoint, and we show that the algorithm is able to learn them, 82 % of them in less than a minute and asking the user just once.