AUG: A combined classification and clustering approach for web people disambiguation

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 105–108, Prague, June 2007. c⃝2007 Association for Computational Linguistics AUG: A combined classification and clustering approach for web people disambiguation Els Lefever and V´eronique Hoste LT3 Language and Translation Technology Ghent University Association Groot-Brittanni¨elaan 45, 9000 Gent els.lefever@hogent.be veronique.hoste@hogent.be Timur Fayruzov Computational Web Intelligence Ghent University Association Krijgslaan 281, 9000 Gent Timur.Fayruzov@UGent.be Abstract This paper presents a combined super- vised and unsupervised approach for multi- document person name disambiguation. Based on feature vectors reflecting pairwise comparisons between web pages, a classifi- cation algorithm provides linking informa- tion about document pairs, which leads to initial clusters. In addition, two different clustering algorithms are fed with matrices of weighted keywords. In a final step the “seed” clusters are combined with the results of the clustering algorithms. Results on the validation data show that a combined classi- fication and clustering approach doesn’t al- ways compare favorably to those obtained by the different algorithms separately. 1 Introduction Finding information about people on the World Wide Web is one of the most popular activities of Internet users. Given the high ambiguity of person names and the increasing amount of information on the web, it becomes very important to organize this large amount of information into meaningful clus- ters referring each to one single individual. The problem of resolving name ambiguity on the Internet has been approached from different an- gles. Mann and Yarowsky (2003) have proposed a Web based clustering technique relying on a fea- ture space combining biographic facts and associ- ated names, whereas Bagga and Baldwin (1998) have looked for coreference chains within each doc- ument, take the context of these chains for creating summaries about each entity and convert these sum- maries into a bag of words. Documents get clustered using the standard vector space model. Other re- searchers have taken this search for distinctive key- words one step further and tried to come up with “concepts” describing the documents. Fleischman and Hovy (2004) introduce the “maximum entropy model”: a binary classifier determines whether two concept-instance pairs refer to the same individ- ual. Pedersen (2006) presented an unsupervised ap- proach using bigrams in the contexts to be clustered, thus aiming at a concept level semantic space instead of a word level feature space. For the semeval contest, we approached the task from a double supervised and unsupervised perspec- tive. For the supervised classification, the task was redefined in the form of feature vectors containing disambiguating information on pairs of documents. In addition to this, different clustering approaches were applied on matrices of keywords. These results were then merged by taking the classification output as basic ”seed” clusters, which were then enhanced by the results from the clustering experiments. In the remainder of this paper, Section 2 intro- duces the data sets and describes the construction of the feature vectors and the keyword matrices. The classification and clustering experiments, and the final combination of the different outputs are dis- cussed in Section …

Content truncated to 500 words