FICO: Web Person Disambiguation Via Weighted Similarity of Entity Contexts

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 149–152, Prague, June 2007. c⃝2007 Association for Computational Linguistics FICO: Web Person Disambiguation Via Weighted Similarity of Entity Contexts Paul Kalmar Fair Isaac Corporation 3661 Valley Centre Dr. San Diego, CA 92130 USA PaulKalmar@FairIsaac.com Matthias Blume Fair Isaac Corporation 3661 Valley Centre Dr. San Diego, CA 92130 USA MatthiasBlume@FairIsaac.com Abstract Entity disambiguation resolves the many- to-many correspondence between mentions of entities in text and unique real-world en- tities. Fair Isaac’s entity disambiguation uses language-independent entity context to agglomeratively resolve mentions with similar names to unique entities. This pa- per describes Fair Isaac’s automatic entity disambiguation capability and assesses its performance on the SemEval 2007 Web People Search task. 1 Introduction We use the term entity to mean a specific person or object. A mention is a reference to an entity such as a word or phrase in a document. Taken to- gether, all mentions that refer to the same real- world object model that entity (Mitchell et al. 2004). Entity disambiguation inherently involves resolving many-to-many relationships. Multiple distinct strings may refer to the same entity. Si- multaneously, multiple identical mentions refer to distinct entities (Bagga and Baldwin, 1998). Fair Isaac’s entity disambiguation software is based largely on language-independent algorithms that resolve mentions in the context of the entire corpus. The system utilizes multiple types of con- text as evidence for determining whether two men- tions correspond to the same entity and it auto- matically learns the weight of evidence of each context item via corpus statistics. The goal of the Web People Search task (Artiles et al. 2007) is to assign Web pages to groups, where each group contains all (and only those) pages that refer to one unique entity. A page is assigned to multiple groups if it mentions multiple entities, for example “John F. Kennedy” and the “John F. Kennedy Library”. The pages were se- lected via a set of keyword queries, and the disam- biguation is evaluated only on those query entities. This differs from Fair Isaac’s system in a few key ways: our system deals with mentions rather than documents, our system does not require a filter on mentions, and our system is generally used for large collections of documents containing very many names rather than small sets of highly am- biguous documents dealing with one specific name. Nevertheless, it was possible to run the Fair Isaac entity disambiguation system on the Web People Search task data with almost no modifica- tions and achieve accurate results. The remaining sections of this paper describe Fair Isaac’s automatic entity disambiguation meth- odology and report on the performance of the sys- tem on the WePS data. 2 Methodology In unstructured text, each document provides a natural context for entity disambiguation. After cleaning up extraneous markup we carry out within-document co-reference resolution, aggregat- ing information about each entity mentioned in each document. We then use these entity attributes as features in determining which documents deal with the same entity. …

Content truncated to 500 words