Conference,
A Chip Off the Old Block – Extracting Typical Attributes for Entities based on Family Resemblance
(2013)
Abstract
Google’s Knowledge Graph offers structured summaries for entity searches. This provides a better user experience by focusing on the main aspects of the query entity only. But to do this Google relies on strictly curated knowledge bases. In consequence, only well-known entities included in such knowledge bases can benefit from such a feature. In this paper, we propose ARES, a system that automatically discovers a manageable number of attributes well-suited for high precision entity summarization. Starting from any entity-centric query and exploiting diverse facts automatically extracted from Web documents, for each entity type ARES derives a common structure (or schema) comprising attributes typical for entities of the same or similar entity type. To do this, we extend the concept of typicality from cognitive psychology and define a practical measure for attribute typicality together with a novel algorithm for its efficient calculation. We evaluate the quality of derived structures for various entities and entity types in terms of precision and recall against Wikipedia article structures, as well as human assessments. ARES achieves results superior to the structure presented by Google’s Knowledge Graph or to frequency-based statistical approaches for structure extraction.
