Guha and Garg (2004) report that approximately 4% of internet searches are to locate named indi­viduals. Yet, many people share the same name with for example 157630 individuals in the UK sharing the most common name 'David Jones' (UK statistics cited by Ash 2006). Consequently identifying web pages on specific individuals is a significant problem that will grow as everyone ac­quires a web presence.

There are several proposed approaches to identi­fying which individuals correspond to which web pages. For example, Bollegala et al. (2007) pro­pose augmenting queries in the style of relevance feedback (Salton and Buckley 1990), Kalashnikov (2007) treat Web Person Search (WePS) as a dis­ambiguation problem whose objective is to distin­guish individuals, whilst Wan et al. (2005) see WePS as a clustering problem.

WePS has both similarities and differences to word sense disambiguation (WSD). Both seek to classify instances of usage, but in WSD the sense inventory is fixed. WSD then is more amenable to a classification solution where a system can be ef­fectively trained using learning algorithms. In WePS we do not know from the outset how many individuals our pages correspond to. Consequently we took the view that WePS is better seen as a clustering rather than a classification problem.

UNN-WePS was made up of three components, comprising modules to:

1. Create seed clusters that associated files with person names other than those being searched for.

2. Match similarity of unallocated documents to micro clusters using lexical chains de­rived from Roget's thesaurus.

3. Identify entities using single link agglom-erative clustering algorithm.

In detail, a part of speech tagger (Coburn et al. 2007) was used to identify sequences of proper nouns. Person names were identified from these sequences using the following simple names 'grammar' coupled with data from the US Census (1990).

We also used a gazetteer to forms seed clusters using data from the World Gazetteer (2007). This did not form part of the submitted system.

UNN-WePS achieved an average purity of 0.6, and inverse purity of 0.73 in Semeval Task 13, achiev­ing seventh position out of sixteen competing sys­tems (Artiles et al. 2007). However there was con­siderable variance in UNN-WePS results as shown in graph 1 below.

Graph 1: UNN-WePS purity performance

Graph 1 shows the purity scores for UNN-WePS on the Semeval 13 test data on three conditions: (1) as submitted (solid line), (2) using the gazetteer (dashed line), and (3) without the lexical chain based similarity matching (dotted line).

Note although the purity is lower when similar­ity matching is included the number of discarded documents is approximately halved.

An examination of the data suggests that where performance was especially poor it was due to ge­nealogical data. Firstly this contains multiple indi­viduals sharing the same name violating the 'one sense per discourse' heuristic. Secondly genealogi­cal data includes birth and death information which was outside the scope of UNN-WePS. Further­more, the large number of names confounds the statistical utility of co-present names.

We have described a system, UNN-WePS that dis-ambiguates individuals in web pages as required for Semeval task 13 (Artiles et al. 2007).

UNN-WePS was composed of three modules. The first formed seed clusters based on names pre­sent in web pages other than the individual for whom we are searching. The second used a lexical chain based similarity measure to associates re­maining files with clusters, whilst the third joined the clusters to identify identities using a single link hierarchical algorithm.

UNN-WePS performed surprisingly well con­sidering the simplicity of its basic seeding algo­rithm. The use however of the 'one sense per dis­course' heuristic was flawed. Names do re-occur across generations in families.

Genealogy is a popular Internet pastime, and web pages containing genealogy data frequently refer to multiple individuals that share a name at different time periods. As UNN-WePS did not ac­count for time, this could not be detected. Further­more, the large number of names in on-line genea­logical data does lead to spurious associations.

As WePS was time limited, several extensions and refinements were envisaged, but not executed. Firstly, as described, the world gazetteer (2007) did not lead to performance improvements. We speculate therefore the disambiguation effect from using place names was exceeded by the ambiguity introduced by using them blindly. We note espe­cially the inference between unidentified names (or street names, or building names) being interpreted as place data.

A further system deficiency was the lack of rec­ognition of date data. This is essential to differenti­ate between identically named individuals in ge­nealogical data.

Finally, we are currently experimenting with dif­ferent clustering algorithms using the CLUTO toolkit (Karypis 2002) to improve on UNN-WePS baseline performance.