WIT: Web People Search Disambiguation using Random Walks

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 480–483, Prague, June 2007. c⃝2007 Association for Computational Linguistics WIT: Web People Search Disambiguation using Random Walks Jos´e Iria, Lei Xia, Ziqi Zhang The University of Sheffield 211 Portobello Street Sheffield S1 4DP, United Kingdom {j.iria, l.xia, z.zhang}@sheffield.ac.uk Abstract In this paper, we describe our work on a ran- dom walks-based approach to disambiguat- ing people in web search results, and the im- plementation of a system that supports such approach, which we used to participate at Semeval’07 Web People Search task. 1 Introduction Finding information about people on the Web us- ing a search engine is far from being a quick and easy process. There is very often a many-to-many mapping of person names to the actual persons, that is, several persons may share the same name, and several names may refer to the same person. In fact, person names are highly ambiguous: (Guha and Garg, 2004) reports that only 90.000 thousand dif- ferent names are shared by 100 million people ac- cording to the U.S. Census Bureau. This creates the need to disambiguate the several referents typically found in the web pages returned by a query for a given person name. The Semeval’07 Web People Search challenge (Artiles et al., 2007) formally evaluated systems on this task. In this paper, we describe our work on a random walks-based approach to disambiguating people in web search results, heavily influenced by (Minkov et al., 2006). This particular model was chosen due to its elegance in seamlessly combining lexico-syntactic features local to a given webpage with topological features derived from its place in the network formed by the hyperlinked web pages returned by the query, to arrive at one single mea- sure of similarity between any two pages. 2 Proposed Method In a nutshell, our approach 1) uses a graph to model the web pages returned by the search engine query, 2) discards irrelevant web pages using a few sim- ple hand-crafted heuristics, 3) computes a similarity matrix for web pages using random walks over the graph, and 4) finally clusters the web pages given the similarity matrix. The next subsections detail these steps. 2.1 Web People Search Graph We build a directed weighted typed graph from the corpus. The graph is a 5-tuple G = (V, E, t, l, w), where V is the set of nodes, E : V ×V is the ordered set of edges, t : V → T is the node type function (T = {t1, . . . , t|T|} is a set of types), l : E → L is the edge label function (L = {l1, . . . , l|L|} is a set of labels), and w : L → R is the label weight function. We structure our problem domain with the types and labels presented in Figure 1. In order to transform the text into a graph that conforms to the model shown, we take the output of standard …

Content truncated to 500 words