Personal names are often submitted to search en­gines as query keywords, as described in a reportindicating that about 10% of the English queries from the search engine ALLTheWeb contain per­sonal names. However, in response to a personal name query, search engines return a long list of search results containing that contains Web pages about several namesakes. For example, when a user submits a personal name like "William Cohen" as a query to the search engine Google, the re­turned results represent more than one person named "William Cohen." In the results, a computer sci­ence professor, an American politician, a surgeon, and others are not classified into separate clusters but mixed together.

Most of the previous works on disambiguating personal names in Web search results employ sev­eral kinds of agglomerative clustering approach as described in Section 2. However, in these ap­proaches, a lot of clusters that contain only one el­ement tend to be generated, depending on a certain criterion for merging similar clusters. In addition, in person search results from the World Wide Web (WWW), we can often observe that a small num­ber of entities have a lot of search-result Web pages, while others have only one or two. In light of these facts, if a labeled Web page that describes a person is introduced, clustering for personal name disam­biguation would be much more accurate. In the fol­lowing, we refer to such a labeled Web page as the "seed page'" Then, in order to disambiguate per­sonal names in Web search results, we introduce semi-supervised clustering that uses the seed page to aid the clustering of unlabeled search-result Web pages. Our semi-supervised clustering approach is characterized by controlling the fluctuation of the centroid of a cluster.

(Mann and Yarowsky, 2003) first extract biographi­cal information, such as birthdates, birthplaces, oc­cupations, and so on. Then, for each document, they generate a feature vector composed of the ex­tracted biographical information, proper nouns, and the TF-IDF score computed from the documents in the search results. Finally, using this feature vec­tor, they disambiguate personal names by generating clusters based on a bottom-up centroid agglomerative clustering algorithm. (Wan et al., 2005) employ an approach similar to that of (Mann and Yarowsky, 2003), and have developed a system called Web-Hawk.

(Pedersen et al., 2005) recently proposed a method for discriminating names by clustering the instances of a given name into groups. They ex­tract the context of each instance of an ambiguous name and generate second-order context vectors us­ing significant bigrams. The vectors are then clus­tered such that instances that are similar to each other are grouped into the same cluster.

(Bekkerman and McCallum, 2005) propose the following three unsupervised approaches: (1) an approach based on the hyperlink structures of Web pages; (2) an approach based on agglomera-tive/conglomerative double clustering (Bekkerman et al., 2005); and (3) a hybrid approach combining the first two.

(Bollegala et al., 2006) first agglomeratively clus­ter a set of documents and then select key phrases from the resulting clusters to distinguish different namesakes. They extract key phrases from the doc­uments and merge the clusters according to the sim­ilarity between the extracted phrases.

In this section, we first review the pure agglomera-tive clustering approach that most of the previous re­lated works employ and then describe our proposed semi-supervised clustering approach.

In the following discussion, we denote the feature vector of a search-result Web page in a set of search results as follows:

where is the number of distinct terms in the Web page , and denotes each term.

Stop words were eliminated from all Web pages in the search results based on the stopword list, and stemming was performed using Porter stemmer. In our preliminary experiments, we found that gain (Papineni, 2001) is the most effective term weight­ing scheme for generating feature vectors for clus­tering in this kind of task. Using the gain scheme, we also define each element tuf of wp as follows:

Algorithm: Agglomerative clustering Input: Set of search-result Web page pi (t = 1,2, P = {pi,p2) ■ ■ *Pn}.

Output: Clusters that contain the Web pages that refer to the same person. Method:

1. Set the each element in P as initial clusters.

until all of the similarities between two clusters are less than the predefined threshold.

Tables 2 and 3 show evaluation results in each doc­ument set obtained using pure agglomerative clus­tering and our proposed semi-supervised clustering, respectively. "Set 1," "Set 2," and "Set 3" con­tain the names from participants in the ACL con­ference, from biographical articles in the English Wikipedia, and from the US Census, respectively. According to these tables, we found that, although agglomerative clustering outperforms our proposed semi-supervised clustering by 0.21 in the value of purity, our proposed semi-supervised clustering out­performs agglomerative clustering by 0.4 and 0.06 in the values of inverse purity and F-measure, re­spectively. This indicates that our proposed method tends to integrate search-result Web pages into a seed page and a small number of clusters are gen­erated compared with agglomerative clustering. In terms of these facts, it is easier for a user to browse Web pages clustered based on each personal entity. On the other hand, the small values of purity indi­cate that irrelevant search-result Web pages are of­ten contained in the generated clusters. Therefore, we can guess that irrelevant search-result Web pages are integrated into a seed page. In fact, we observed that more than 50 search-result Web pages could be grouped together with a seed page.

In this paper, we described our participating system in the SemEval-2007 Web People Search Task (Ar-tiles et al., 2007). Our system used a semi-supervised clustering which controls the fluctuation of the centroid of a cluster. The evaluation results showed that our proposed method achieves high scores in inverse purity, with the lower scores in pu­rity. This fact indicates that our proposed method tends to integrate search-result Web pages into a seed page. This clustering result makes it easier for a user to browse the results of a person Web search. However, in the generated cluster with a seed page, irrelevant search-result Web pages are also contained. This problem can be solved by in-

Table 4: Evaluation results based on each seed page obtained using our proposed semi-supervised clus­tering._ troducing multiple seed pages. In our experiment, we used the full contents of search-result Web pages and a seed page. We consider that this can cause lower scores in purity. Therefore, in future work, in order to improve the accuracy of clustering, we plan to conduct further experiments by introducing multiple seed pages and using parts of search-result Web pages and seed pages such as words around an ambiguous name.