Utilizing Deep Learning Heuristics in Knowledge Graph Applications

Abstract

Knowledge graphs (KGs) are databases containing relational triples of structured, machine-readable data. Large-scale KGs, such as DBpedia, Freebase, and YAGO have played a crucial role in numerous applications, such as question answering systems, recommendation systems, and expert systems. Deep learning (DL) is a subset of machine learning techniques comprised of artificial neural network layers. It is a relatively new technology that has attained cutting-edge performance in many domains of research, business, and government in a very short period of time. In this work, we illuminate the progress that DL heuristics have brought about in five KG tasks: KG completion (KGC), entity summarization (ES), entity linking (EL), KG-based recommendation systems (RSs), and KG embedding. In addition, we contributed by developing two ES models and one RS model. Our first ES model, ESDL [1], is a simple yet effective supervised DL model that encodes resource description framework (RDF) triples using textual semantics. As a supervised model, it employs labels in the form of frequency scores collected programmatically from human-generated summaries. Performance and effectiveness experiments conducted on a publicly available benchmark dataset demonstrate that ESDL achieved cutting-edge performance. Our second ES model, ESDL2, is an improvement to the first model mentioned above. We upgraded the model by modifying its design to include additional input features retrieved from the entity’s descriptions and utilize gated recurrent unit (GRU) layers. Experimental findings indicate that the new model outperforms its predecessor by an average of 11.8%. Experiments also demonstrate that ESDL2 achieves state-of-the-art performance by significantly outperforming all competing models. We also propose WikiRS, a hybrid RS for Wikidata editors that takes each editor's editing frequency and profile similarity to other editors into consideration. Additionally, it incorporates items’ content exemplified by their labels and ontological classes. Our experiments reveal that WikiRS outperforms rival models in a statistically significant manner.