Open Domain Question Answering over Knowledge Graphs Using Keyword Search, Answer Type Prediction, SPARQL and Pre-trained Neural Models

Question Answering (QA) in vague or complex open domain information needs is hard to be adequate, satisfying and pleasing for end users. In this paper we investigate an approach where QA complements a general purpose interactive keyword search system over RDF. We describe the role of QA in that context, and we detail and evaluate a pipeline for QA that involves a general purpose entity search service over RDF, answer type prediction, entity enrichment through SPARQL, and pre-trained neural models. The fact that we start from a general purpose keyword search over RDF, makes the proposed pipeline widely applicable and realistic, in the sense that it does not pre-suppose the availability of knowledge graph-specific training dataset. We evaluate various aspects of the pipeline, including the effect of answer type prediction, as well as the performance of QA over existing benchmarks. The results show that, even by using different data sources for training, the proposed pipeline achieves a satisfactory performance. Moreover we show that the ranking of entities for QA can improve the entity ranking.