Information Search, Integration, and Personalization: 13th International Workshop, ISIP 2019, Heraklion, Greece, May 9–10, 2019, Revised Selected Papers

In this paper we introduce an approach, called LODQA, for open domain Question Answering over Linked Open Data. We confine ourselves to three kinds of questions: factoid, confirmation, and definition questions. By using LODQA it is feasible to answer questions over 400 millions of entities of any domain without using any training data, since we exploit simultaneously 400 Linked datasets. In particular, we exploit the services of LODsyndesis, a suite of services (based on semantics-aware indexes) which supports cross-dataset reasoning over hundreds of Linked datasets and 2 billion triples. The proposed Question Answering process follows an information extraction approach and comprises several steps including question cleaning, heuristic based question type identification, entity recognition, linking and disambiguation using Linked Data-based methods and pure NLP methods (specifically DBpedia Spotlight and Stanford CoreNLP), WordNet-based question expansion for tackling the lexical gap (between the input question and the underlying sources), and triple scoring for producing the final answer. We discuss the benefits of this approach in terms of answerable questions and answer verification, and we investigate, through experimental results, how the aforementioned steps of the process affect the effectiveness and the efficiency of question answering.