Uncertainty in linked data

Abstract

The semantic web and linked data have become essential in modern data management, providing valuable insights by interconnecting diverse datasets. However, the quality and completeness of linked data remain significant challenges due to uncertainties such as integration errors, ambiguities in semantic relationships, and incomplete data. This thesis addresses these issues through advanced deep learning techniques. We propose four main contributions: (1) LinkED-S2S, a sequence-to-sequence model incorporating an embedding layer and an attention mechanism for detecting missing semantic links in RDF (Resource Description Framework) datasets; (2) An encoder-decoder model based on an embedding layer and an attention mechanism for predicting missing types of RDF entities; (3) SASNN, a Siamese neural network for detecting sameAs links, aimed at improving entity alignment within the dataset and across multiple datasets; and (4) A model relying on three LSTM (Long Short Term Memory) neural networks and embedding layers for detecting erroneous triples in linked data. Each contribution has been tested individually, achieving promising results on several benchmark datasets, demonstrating the effectiveness of our methods. Additionally, our contributions address scalability issues and consider the semantic relationships between entities. A case study on the UniProt dataset further validates the combined application of these contributions in generating and verifying triples, enabling the resolution of incompleteness and errors in linked data.