アイテム詳細

要旨

This paper pursues the recently emerging paradigm of searching for entities that are embedded in Web pages. We utilize information extraction techniques to identify entity candidates in documents, map them onto entries in a richly structured ontology, and derive a generalized data graph that encompasses {W}eb pages, entities, and ontological concepts and relationships. We exploit this combination of pages and entities for a novel kind of search-result ranking, coined {E}ntity{A}uthority, in order to improve the quality of keyword queries that return either pages or entities. To this end, we utilize the mutual reinforcement between authoritative pages and important entities. This resembles the {HITS} method for Web-graph link analysis and recently proposed {O}bject{R}ank methods, but our approach operates on a much richer, typed graph structure with different kinds of nodes and also differs in the underlying mathematical definitions. Preliminary experiments with topic-specific slices of Wikipedia demonstrate the effectiveness of our approach on certain classes of queries.