Assigning semantic labels to data sources

Abstract

There is a huge demand to be able to find and integrate heterogeneous data sources, which requires mapping the attributes of a source to the concepts and relationships defined in a domain ontology. In this paper, we present a new approach to find these mappings, which we call semantic labeling. Previous approaches map each data value individually, typically by learning a model based on features extracted from the data using supervised machine-learning techniques. Our approach differs from existing approaches in that we take a holistic view of the data values corresponding to a semantic label and use techniques that treat this data collectively, which makes it possible to capture characteristic properties of the values associated with a semantic label as a whole. Our approach supports both textual and numeric data and proposes the top k semantic labels along with their associated confidence scores. Our experiments show that the approach has higher label prediction accuracy, has lower time complexity, and is more scalable than existing systems.