Semantic Query Processing: Estimating Relational Purity

The use of semantic information found in structured knowledge bases has become an integral part of the processing pipeline of modern intelligent information systems. However, such semantic information is frequently insufficient to capture the rich semantics demanded by the applications, and thus corpus-based methods employing natural language processing techniques are often used conjointly to provide additional information. However, the semantic expressiveness and interaction of these data sources with respect to query processing result quality is often not clear. Therefore, in this paper, we introduce the notion of relational purity which represents how well the explicitly modelled relationships between two entities in a structured knowledge base capture the implicit (and usually more diverse) semantics found in corpus-based word embeddings. The purity score gives valuable insights into the completeness of a knowledge base, but also into the expected quality of complex semantic queries relying on reasoning over relationships, as for example analogy queries.