Multi-source named entity typing for social media

Abstract

Typed lexicons that encode knowledge about the semantic types of an entity name, e.g., that ‘Paris’ denotes a geolocation, product, or person, have proven useful for many text processing tasks. While lexicons may be derived from large-scale knowledge bases (KBs), KBs are inherently imperfect, in particular they lack coverage with respect to long tail entity names. We infer the types of a given entity name using multi-source learning, considering information obtained by alignment to the Freebase knowledge base, Web-scale distributional patterns, and global semi-structured contexts retrieved by means of Web search. Evaluation in the challenging domain of social media shows that multi-source learning improves performance compared with rule-based KB lookups, boosting typing results for some semantic categories.