ICTNET at Entity Track TREC 2010

This paper gives an overview of our work for related entity finding which is proposed in TREC 2010 Entity Track. The goal of the Entity Track is to find the entities relevant to a given query from the web corpus. In this paper, we propose a bipartite graph reinforcement model for entity ranking. As is well known, the entities on the web are embedded not only in the natural language text, but also in the tables and lists. Given a query, both the candidate entities and relevant tables/lists are extracted from web documents. Then the candidate entities extracted from unstructured text are ranked based on a probabilistic model. But the result contains a lot of noise. If some candidate entities are in a relevant table/list, they are more relevant to the given query. And Vice versa, if a table/list contains several candidate entities, it is also more relevant to the query. Based on the above intuition, we construct a bipartite graph and then perform a reinforcement algorithm to re-rank the candidate entities.