In principle, the number of different people that are represented in the data cannot be known in advance. However, for the clustering process, either the num­ber of classes has to be fixed before clustering, or some kind of termination criterion has to be found that tells the algorithm when to stop clustering.

A good estimation of the number of different en­tities is a necessary prerequisite for successful clus­tering. Clustering into too many classes would mean assigning documents to classes that have actually no own entity they refer to. Clustering into too few classes means merging two entities into one class.

Our initial intuition was to distinguish people by normally unique properties, like phone numbers or email addresses. So we assumed that the number of different email addresses and phone numbers occur­ring in all documents for one name would be a good means to estimate the number of different persons sharing this name, but we could not find any corre­lation between these features and the class number.

Therefore, we decided to estimate the average number of classes from the training data. The aver­age number of different people for one name in the training data was about 18. Based on the observa­tion that an underestimated number of classes leads to better results than assuming too many classes, we decided to guess 12 different persons for each name.

For the extraction of plain text information from the web pages, we used the html2text converter. In case that a web page consisted of more than one html document, we put all the output from the converter into one single file. By omitting any wrapping of the html pages, we obviously lost useful structural information but got the textual information for our linguistic analysis.

Afterward, we applied several linguistic prepro­cessing tools. We used coreference resolution to re­place pronouns referring to a person, and variations of a name (like "Mr. Smith" after a mention of "John Smith" earlier in the text) with the person's name in the form of its first mention in the text.

For NE-tagging, we used the three NE types PER­SON, LOCATION and ORGANIZATION. For both NE tagging and coreference resolution, we used the LingPipe toolkit . We counted the occurrences of every NE in every file and replaced all instances by their specific NE type combined with a uniquely identifying number, e.g. we replaced all occurrences of "Paris" with "LOCATION27", in order to ensure that the predicate argument parser could work cor­rectly and would not split up multi-word NEs into two or more arguments.

We passed all sentences with NEs that con­tained the specified persons family name (e.g. "Mr. Cooper" for the name "Alvin Cooper") to MontyLingua , that returns a semantic represen­tation of the sentence like ("live" "PERSON2" "in LOCATION3"). These representations abstract from the actual surface form of a sentence as they represent every sentence in its underlying semantic form ("predicate" "semantic subject" "semantic ob-ject1"... ) rather than just determining the syntactic subject and objects of a sentence. We called these structures "patterns" and kept only those that actu­ally contained the respective NE.

We decided on building two vectors for every text file, one for the NEs and one for sentence patterns connected to a person's name in order to give to the NEs a weight different from that for the patterns.

After tagging the documents for NEs, we counted the frequency of the different occurring NEs for one name. We built a first feature vector for each docu­ment that contained as entries the counts of the oc­curring NEs in this document. We set a threshold n to use only the n best NEs in the vectors, counted over all documents for one name. We then built for every document a second feature vector containing the counts of the MontyLingua patterns for the doc­ument.

For the actual clustering process, we used hierar­chical clustering. We started with every file, rep­resented by a pair of normalized feature vectors, constituting a single cluster. As distance measure­ment we used the weighted sum of the absolute dis­tances between the centers of two clusters with re­gard to both feature vectors, respectively, i.e. we chose distance = wdistanceN£s+distancepflfferns. In every step, we made a pairwise comparison of all clusters and merged those with the lowest distance. The clustering terminated when the algorithm came down to the pre-set number of 12 clusters. So far we have not made any further use of the binary tree structure within each cluster.

We assigned every file to exactly one cluster. We had neither a "discarded" category nor did we handle the possibility that a page refers to more than one person and would hence belong to different clusters.

We evaluated the system on the provided WePS training data to estimate the following parameters: number of classes, number of best NEs to be consid­ered and weight of the NE vector compared to the pattern vector.

The relevant evaluation score is the F-measure (a = 0.5) as the harmonic mean of purity and in­verse purity as described by Hotho et al. (2003).

As our attempt to use distinctive features for the estimation ofclass numbers failed, we examined the influence of a wrongly estimated number of classes on the clustering results. Table 1 shows exemplarily for 2 person names how the F-measure varies if the correct number of classes is incorrectly assumed as a higher or lower value. We concluded that it is better to estimate the class number too low than too high.

Primarily meant as a means to reduce computa­tion time, we gave our system the possibility not to use all occurring NEs for clustering, but only a cer­tain number of entities with maximal frequencies. Test runs did not confirm our hypothesis that con­sidering a higher number of NEs leads to better re­sults (cf. table 2). For both training of the number of NEs and the NE weight we assumed that we already knew the correct class number.

As the F-measure did not increase for more con­sidered NEs, we believe that the most important NEs are already covered within the best 100 and that adding more NEs rather adds coincidental informa­tion than any new important facts. Usually, the best 100 NEs already cover most of those which occur more than once in a text.

The third parameter to estimate was the weight w given to the NE feature vector compared to the feature vector for sentence patterns. During training, this weight also appeared to have little influence on the clustering results (cf. 2). We have the hypothesis that sentence pattern detection is not very successful for the often unstructured web page texts.

In the WePS evaluation, our system scored with a purity of 0.39, an inverse purity of 0.83 and a result­ing overall F-measure (a = 0.5) of 0.5.

One main reason for our test results to be worse than our training results is the fact that the test data had a much higher average number ofclasses (about 46 classes). Our F-measure was best for those names with the fewest number of referents. We had an av­erage F-Measure (a = 0.5) of 0.66 for those names with less than 30 instances compared to an overall average of 0.50. These numbers show the impor­tance of a correct estimation of the assumed number of referents for a name.

Our purity was much lower than the inverse pu­rity, i.e. there is too much noise in our clustering compared to the real partition, whereas the real clus­ters are well covered by our clustering. This is due to a too low estimation of the number of referents.