Section: Application Domains
Mining Referring Expressions in Knowledge Bases. A referring expression (RE) is a description that identifies a concept unambiguously in a domain of knowledge. For example, the expression “X is the capital of France” is an RE for Paris, because no other city holds this title. Mining REs from data is a central task in natural language generation, and is also applicable to automatic journalism and query generation (e.g., for benchmarking purposes). A common requirement for REs is to be “intuitive”, that is, to resort to concepts that are easily understandable by users. For this reason, existing methods required users to provide a lexical ranking of concepts that conveys their preferences for certain predicates and entities in descriptions. In addition, state-of-the-art methods are not tailored for large current knowledge bases and, due to data incompleteness, are often unable to provide an answer. The internship of Julien Delaunay was conceived to tackle these issues by designing a parallel method to mine intuitive REs on large knowledge bases. The system extends the state-of-the-art language bias for REs to deal with incompleteness and proposes a notion of intuitiveness based on information theory that does not require a lexical ranking from the user. The description of the system, named REMI, is under review at the Extended Semantic Web Conference (ESWC) 2019.