Searching RDF Graphs with SPARQL and Keywords

The proliferation of knowledge-sharing communities like Wikipedia and the advances in automated information extraction from Web pages enable the construction of large knowledge bases with facts about entities and their relationships. The facts can be represented in the RDF data model, as so-called subject-property-object triples, and can thus be queried by structured query languages like SPARQL. In principle, this allows precise querying in the database spirit. However, RDF data may be highly diverse and queries may return way too many results, so that ranking by informativeness measures is crucial to avoid overwhelming users. Moreover, as facts are extracted from textual contexts or have community-provided annotations, it can be beneficial to consider also keywords for formulating search requests. This paper gives an overview of recent and ongoing work on ranked retrieval of RDF data with keyword-augmented structured queries. The ranking method is based on statistical language models, the state-of-the-art paradigm in information retrieval. The paper develops a novel form of language models for the structured, but schema-less setting of RDF triples and extended SPARQL queries. 1 Motivation and Background Entity-Relationship graphs are receiving great attention for information management outside of mainstream database engines. In particular, the Semantic-Web data model RDF (Resource Description Format) is gaining popularity for applications on scientific data such as biological networks [14], social Web2.0 applications [4], large-scale knowledge bases such as DBpedia [2] or YAGO [13], and more generally, as a light-weight representation for the “Web of data” [5]. An RDF data collection consists of a set of subject-property-object triples, SPO triples for short. In ER terminology, an SPO triple corresponds to a pair of entities connected by a named relationship or to an entity connected to the value of a named attribute. As the object of a triple can in turn be the subject of other triples, we can also view the RDF data as a graph of typed nodes and typed edges where nodes correspond to entities and edges to relationships (viewing attributes as relations as well). Some of the existing RDF collections contain more than a billion triples. As a simple example that we will use throughout the paper, consider a Web portal on movies. Table 1 shows a few sample triples. The example illustrates a number of specific requirements that RDF data poses for querying: Copyright 2010 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering