Cross document person name disambiguation using entity profiles

Consolidating entity information spread across multiple documents is a critical problem now with the growing use of large open-domain document sources. Associating every entity in a corpus to a unique entry in a growing knowledge base serves a dual purpose of consolidating (disambiguating) entities as well as to build a rich growing knowledge source containing information about each and every entity accumulated from several documents. With the presence of ambiguous names, use of nominals and aliases, the task of hyper-tagging an entity mentioned in a document to a node in a knowledge base requires the use of context in addition to name matching rules. In this paper, we present an approach that computes a similarity between entities identified in a document with those in the knowledge base using a Vector Space Model utilizing document level entity profiles Information accumulated for each entity from the entire document. The technique resulted in a TAC evaluation score of 71.9 at the TAC 2009 KBP track. The same technique was also successfully used in obtaining state of the art F-measures (93.95) in disambiguating person names by clustering the similarity values obtained using hierarchical agglomerative clustering.