Related Entity Finding: University of Waterloo at TREC 2010 Entity Track

摘要

The University of Waterloo participated in the Related Entity Finding task of the Entity track. Our goal is to investigate whether related entity finding problem can be addressed by unsupervised approaches that rely primarily on statistical methods and common linguistic tools, such as named- entity taggers and syntactic parsers. We approach the related entity finding problem by first retrieving documents in response to the query, and extracting an initial set of candidate entities from the text of the documents. As a separate step, we automatically construct a set of seed entities, which represent hyponyms of the target entity category specified in the narrative, and then rank the candidate entities by their similarity to the seeds. An example of the target entity category name is 'authors', extracted from the narrative 'Authors awarded an Anthony Award at Bouchercon in 2007' (2009 topic No. 14). The system extracts category names from the free-text narrative, finds seed entities belonging to each category, and computes the similarity of candidate entities to the seeds.