DFKI2: An Information Extraction Based Approach to People Disambiguation

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 137–140, Prague, June 2007. c⃝2007 Association for Computational Linguistics DFKI2: An Information Extraction Based Approach to People Disambiguation Andrea Heyl German Research Center for Artificial Intelligence - DFKI, Saarbr¨ucken, Germany andrea.heyl@dfki.de G¨unter Neumann German Research Center for Artificial Intelligence - DFKI, Saarbr¨ucken, Germany guenter.neumann@dfki.de Abstract We propose an IE based approach to people disambiguation. We assume the mentioning of NEs and the relational context of a per- son in the text to be important discriminat- ing features in order to distinguish different people sharing a name. 1 Introduction In this paper, we propose a system with a linguis- tic view on people disambiguation that exploits the relational and NE context of a person name as dis- criminating features. Texts about different people differ from each other by the names of persons, places and organiza- tions connected to these people and by the relations in which a person’s name is connected to other enti- ties. Therefore we had the hypothesis that the NEs in the documents for a person name should be a main distinctive criterion for disambiguating people. Furthermore, the relational context of a person name should also be able to give good clues for dis- ambiguation. Sentence patterns related to a name, i.e. patterns that contain the name as subject or object like “be(Person X, lawyer)” often convey uniquely identifying information about a person. Our system was not built specifically for the web people search task WePS (Artiles et al., 2007), but is an early version of an IE system that has the more general goal to discover relations between NEs. We see the WePS task as a specific instance of the set of tasks our system should be able to handle. There- fore, we only adapted it slightly to work with the WePS data, but did not make any further customiza- tion w.r.t. the special requirements of people disam- biguation. As our system was built to handle pure texts rather than structured web pages, we relied completely on linguistic information and did not ex- ploit the html structure of the documents provided. 2 Related Work Our system was inspired by the preemptive and on- demand IE approaches by Sekine and Shinyama (Sekine, 2006; Shinyama, 2006) that cluster news- paper articles into classes of articles that talk about the same type of event. They proposed a system to discover in advance all possible relations and to re- turn them in form of tables. We took the idea of distinctive personal attributes as a criterion for disambiguation from the work of Bollegala et al. (2006). They propose an unsu- pervised learning approach to extract phrases that uniquely identify a person from the web and use these discriminative features for clustering. 3 System Overview The goal of the WePS task is to cluster the top 100 web pages returned by a web search engine for a certain name as search query and classify them w.r.t. the underlying different people they …

Content truncated to 500 words