Ambiguity is the task of building up multiple alter­native linguistic structures for a single input. Most of the approaches focus on word sense disambigua­tion (WSD), where the sense of a word has to be determined depending on the context in which it is used.

The same problem arises for named entities shared by different people or for grandsons named after their grandparents. For instance, querying the name "Michael Hammond" in the World Wide Web where there are huge quantities of massive and un­structured data, a search engine retrieves thousands of documents related to this name. However, there are several individuals sharing the name "Michael Hammon". One is a biology professor at the Univer­sity of Arizona, another is at the University of War­wick, there is a mathematician from Toronto among others. The question is which one of these refer­ents we are actually looking for and interested in. Presently, to be able to answer to this question, we have to skim the content of the documents and re­trieve the correct answers on our own.

To automate this process, the named entities can be disambiguated and the different underlying meanings of the name can be found. On the basis of this information, the web pages can be clustered together and organized in a hierarchical structure which can ease the documents' browsing. This is also the objective of the Web People Search (WePS) task (Artiles et al., 2007). What makes the WePS task even more challenging is the fact that in con­trast to WSD where the number of senses of a word are predefined, in WePS we do not know the exact number of different individuals.

For the resolution of the WePS task, we have de­veloped a web page clustering approach using the title and the body content of the web pages. In ad­dition, we group together the documents that share many location, person and organization names, as well as those that point out to the same sub-links.

The rest of the paper is organized as follows. In Section 2 we describe various approaches for name disambiguation and discrimination. Our approach is shown in Section 3, the obtained results and a dis­cussion are provided in Section 4 and finally we con­clude in Section 5.

Early work in the field of name disambiguation is that of (Bagga and Baldwin, 1998) who pro­posed cross-document coreference resolution algo­rithm which uses vector space model to resolve the ambiguities between people sharing the same name. The approach is evaluated on 35 different mentions of John Smith and reaches 85% f-score.

Mann and Yarowski (2003) developed an unsupervised approach to name discrimination where bi­ographical features (age, date of birth), familiar re­lationships (wife, son, daughter) and associations (country, company, organization) are considered. Therefore, in our approach we use person, organiza­tion and location names in order to construct a social similarity network between two documents.

Another unsupervised clustering technique for name discrimination of web pages is that of Peder-sen and Kulkarni (2007). They used contextual vec­tors derived from bigrams, and measured the impact of several association measures. During the evalu­ation, some names were easily discriminable com­pared to others categories for which was even diffi­cult to find and obtain discriminative feature. We worked with their unigram model (Purandare and Pedersen, 2004) to cluster the web pages using the text content between the title tags.

Our web people clustering approach is presented in Figure 1 and consists of the following steps:

• HTML cleaning: all html tags are stripped

away, the javascript code is eliminated, the non closed WePS tags are repaired, the missing be­gin/end body tags are included and then the content between the title, the body and the an­chor tags is extracted.

• name matching: the location, person and orga­nization names in the body texts are identified with the GATE system (Cunningham, 2005). Each named entity of a document is matched with its corresponding named entity category from the rest of the web pages. This infor­mation is used to calculate the social semantic similarity of the person, the location and the or­ganization names. Our hypothesis is that doc­uments with similar names tend to refer to the same individual. The output of this module is a matrix with binary values, where 1 stands for the documents which share more than the half of their proper names, and 0 otherwise.

• links: for each document, we extract the links situated between the anchor tags. Since the links are too specific, we wrote an url function which transform a given web page d1 with URL http://www.cs.ualberta.ca/~lindek/index.htm into www.cs.ualberta.ca/~lindek, and the web page d2 with URL http://www.cs.ualberta.ca/~ lindek/demos.htm into www.cs.ualberta.ca/~lindek. According to our approach, the two web pages di and d2are linked to each other if their link structures (LS) intersect, that is LS(d1) f LS(d2) = 0. The output of this module is a matrix with binary values, where 1 stands for two web pages having more than 3 links in common and 0 otherwise.

• titles: for each document, we extract the text between the title tags. We create a unigram matrix which is feed into SenseClusters. We use automatic cluster stopping criteria with the gap statistics which groups the web pages into several clusters according to the context of the titles. From the obtained clusters, we generate a new matrix with binary values, where 1 corre­sponds to the documents which were put in the

Table 1 shows the obtained results for the test data set. The average performance of our system is 56% and we ranked on 10-th position from 16 participat­ing teams. Although, we have used different sources of information and various approximations, in the future we have to surmount a number of obstacles.

One of the limitations comes from the usage of the text snippets situated between the body tags. There are a number of web pages which do not contain any text. The semantic space for these documents cannot be built with LSA and their similarity score is zero.

Despite the fact that we have eliminated the stop words from the documents and we have transformed the web pages by encoding the syntactic categories, the classification power of LSA was different for the ambiguous names and for the web pages. To some extend this is due to the varying number of words in the web pages. In the future, we want to con­duct experiments with a fixed context windows for all documents.

In this task, the number of senses (e.g. number of different individuals that share the same name) is unknown, and one of the major drawbacks in our approach is related to the setting up of the number of clusters. The K-Means clustering algorithm we used, did not include an automatic cluster stopping criteria, and we had to set up the number of clus­ters manually. To be able to do that, we have ob­served the average number of clusters per name in the trial data. We have evaluated the performance of our approach with several different numbers of clusters. According to the obtained results, the best clusters are 25 and 50. We used the same number of clusters for the test data, however this is a rough parameter estimation.

Person name disambiguation is a very importanttask whose resolution can improve the performance of the search engine by grouping together web pages which refer to different individuals that share the same name.

For our participation in the WePS task, we pre­sented a name disambiguation approach which uses only the information extracted from the web pages. We conducted an experimental study with the trail data set, according to which the combination of the title, the body, the proper names and sub-links reaches the best performance. Our current approach can be improved with the incorporation ofautomatic cluster stopping criteria.

So far we did not take advantage of the document ranking and the returned snippets, but we want to incorporate this information by measuring the snippet similarity on the basis of relevant domain informa­tion (Kozareva et al., 2007).

Many thanks to Ted Pedersen for useful comments and suggestions. This work was partially funded by the European Union under the project QALLME number FP6 IST-033860 and by the Spanish Min­istry of Science and Technology under the project TEX-MESS number TIN2006-15265-C06-01.