UC3M_13: Disambiguation of Person Names Based on the Composition of Simple Bags of Typed Terms

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 362–365, Prague, June 2007. c⃝2007 Association for Computational Linguistics UC3M_13: Disambiguation of Person Names Based on the Composition of Simple Bags of Typed Terms David del Valle-Agudo César de Pablo-Sánchez María Teresa Vicente-Díez Universidad Carlos III de Madrid Escuela Politécnica Superior Av. de la Universidad, 30 – 28911 Leganés (Madrid) Spain {dvalle, cdepablo, tvicente}@inf.uc3m.es Abstract This paper describes a system designed to disambiguate person names in a set of Web pages. In our approach Web documents are represented as different sets of features or terms of different types (bag of words, URLs, names and numbers). We apply Ag- glomerative Vector Space clustering that uses the similarity between pairs of analo- gous feature sets. This system achieved a value of 66% for Fα=0.2 and a value of 48% for Fα=0.5 in the Web People Search Task at SemEval-2007 (Artiles et al., 2007). 1 Introduction Name queries account for a substantial part of Web queries in commercial search engines. Name que- ries often aim at retrieving information about par- ticular persons. Nevertheless, the same query or mention name usually recalls several people and the user is unaware of the potential ambiguity and expects to find the related person after skimming some results. Similar problems are also common for products, organizations and almost any other named object in real world. A related problem appears for differ- ent kinds of objects receiving the same name. For example, Firebird can refer to a car, a guitar, a fic- tion superhero or a database product among more than twenty different senses collected in Wikipe- dia. In all these cases, the user could benefit from a structured representation that facilitates browsing results. Other applications like Question Answer- ing would also benefit from name disambiguation and person names disambiguation, in particular. In this work we focus on the task of disambiguating Web pages retrieved for a person name query as proposed in the Web People Search Task at SemE- val-2007. 2 Background and Related Research In recent work in named entity disambiguation, Malin (2005) identifies two different dimensions to classify approaches to the task depending on the information type that is used and whether the method to train the system is supervised or unsu- pervised. Regarding the information type, Malin (2006) identifies personal information like bio- graphical facts (Bagga and Baldwin, 1998; Mann and Yarowsky, 2003) or relational information (Bekkerman and McCallum, 05), collocations with other entities. Personal name disambiguation has been studied in relation with citation analysis and record linkage and their use to improve Web search results have attracted more interest recently (Guha and Garg 2004; Bollegala, 2006), but it is evaluated only at a small scale. In contrast Bekkerman and McCallum (2005) have focused on disambiguating complete social networks and not only results for one name. 3 System description Web People Search proposes a task to find differ- ent people sharing the same name referred in a set of Web pages and associate each …

Content truncated to 500 words