Please note: This PhD defence will be given online.
Mina Farid, PhD candidate
David R. Cheriton School of Computer Science
The RDF data model has become a prevalent format to represent heterogeneous data because of its versatility. The capability of dismantling information from its native formats and representing it in triple format offers a simple yet powerful way of modelling data that is obtained from multiple sources. In addition, the triple format and schema constraints of the RDF model make the RDF data easy to process as labeled, directed graphs.
This graph representation of RDF data supports higher-level analytics by enabling querying using different techniques and querying languages, e.g., SPARQL. Anlaytics that require structured data are supported by transforming the graph data on-the-fly to populate the target schema that is needed for downstream analysis. These target schemas are defined by downstream applications according to their information need.
The flexibility of RDF data brings two main challenges. First, the extraction of RDF data is a complex task that may involve domain expertise about the information required to be extracted for different applications. Another significant aspect of analyzing RDF data is its quality, which depends on multiple factors including the reliability of data sources and the accuracy of the extraction systems. The quality of the analysis depends mainly on the quality of the underlying data. Therefore, evaluating and improving the quality of RDF data has a direct effect on the correctness of downstream analytics.
This work presents multiple approaches related to the extraction and quality evaluation of RDF data. To cope with the large amounts of data that needs to be extracted, we present DSTLR, a scalable framework to extract RDF triples from semi-structured and unstructured data sources. For rare entities that fall on the long tail of information, there may not be enough signals to support high-confidence extraction. Towards this problem, we present an approach to estimate property values for long tail entities. We also present multiple algorithms and approaches that focus on the quality of RDF data. These include discovering quality constraints from RDF data, and utilizing machine learning techniques to repair errors in RDF data.
200 University Avenue West
Waterloo, ON N2L 3G1