Treffer: Using enhanced knowledge graph embedding and graph summarisation for question answering over knowledge graphs

Much attention has recently been given to Knowledge Graphs (KGs) that contain entities and relations in the application domains of interest. Knowledge Graph Embedding (KGE), question answering and drug repositioning are three popular KG applications; however, some challenges still exist. In KGE, previous studies focus mostly on independent triples and relational paths but they ignore the semantic hierarchies hidden behind multi-relations. Multi-relations are all relations among an entity pair. In question answering, the common issues of existing works are the uncertain number of answers issue, multi-entities issue, KG incompleteness issue and candidate answer extraction issue. The first two issues mean existing works cannot effectively answer questions with an uncertain number of answers and by using multi-entities. As for drug repositioning, constructing an effective domain-specific KG and predicting missing drug-disease links are the most important tasks. ; In this study, we investigated the above unsolved challenges in KGE (neglect of multirelations), question answering over KGs (uncertain number of answers, multi-entities, KG incompleteness, and candidate answer extraction) and KG-based drug repositioning (KG construction and link prediction). Addressing these challenges is therefore the motivation of this research study. We proposed three deep learning-based models and a KG construction process to address them. The experimental results showed that our methods can successfully enhance performance in comparison to state-of-the-art techniques. ; Keywords: knowledge graph, knowledge graph embedding, question answering, drug repositioning