Abstract:
More and more (semi) structured information is becoming available on the Web in the form of documents embedding metadata (e.g., RDF, RDFa, Microformats and others). There are already hundreds of millions of such documents accessible and their number is growing rapidly. This calls for large scale systems providing effective means of searching and retrieving this semi-structured information with the ultimate goal of making it exploitable by humans and machines alike.
This dissertation examines the shift from the traditional web document model to a web data object (entity) model and studies the challenges and issues faced in implementing a scalable and high performance system for searching semi-structured data objects on a large
heterogeneous and decentralised infrastructure. Towards this goal, we define an entity retrieval model, develop novel methodologies for supporting this model, and design a web-scale retrieval system around this model. In particular, this dissertation focuses on the following four main aspects of the system: reasoning, ranking, indexing and querying. We introduce a distributed reasoning framework which is tolerant against low data quality. We present a link analysis approach for computing the
popularity score of data objects among decentralised data sources. We propose an indexing methodology for semi-structured data which o ers a good compromise between query expressiveness, query processing and
index maintenance compared to other approaches. Finally, we develop an index compression technique which increase both the update and query throughput of the system. The resulting system can index billions
of data objects and provides keyword-based as well as more advanced search interfaces for retrieving the most relevant data objects.
This work has been part of the Sindice search engine project at the Digital Enterprise Research Institute (DERI), NUI Galway. The Sindice system currently maintains more than 100 million pages downloaded from the Web and is being used actively by many researchers within and
outside of DERI. The reasoning, ranking, indexing and querying components of the Sindice search engine is a direct result of this dissertation research.