Introduction

The clustering of narrow-domain short texts is an emergent area that has been not attended into detail by the computational linguistic community and only few works can be found in literature [1] [11] [15] [19]. This behaviour may be derived from the high challenge that this problem implies, since the obtained results are very unstable or imprecise when clustering abstracts of scientific papers, technical reports, patents, etc. Therefore, it is difficult to deal with this kind of data: if a term selection method is applied, this has to be done very carefully because term frequencies in the texts are very low. Generally only 10% or 20% of the keywords from the complete keyword list occur in every document and their absolute frequency usually is one or two, and only sometimes three or four [1]. In this situation, changing a keyword frequency by one can significantly change the clustering results.

However, most current digital libraries and other web-based repositories of scientific and technical information provide free access only to abstracts and not to the full texts of the documents. Evenmore, some repositories such as the well known MEDLINE, and the Conseil Européen pour la Recherche Nucléaire (CERN), receive hundreds of publications every day that must be categorized on some specific domain, sometimes with an unknown number of categories a priori. This led to construct novel methods for dealing with this real problem. Although sometimes, keywords are provided by authors for each scientific document, it has been seen that this information is insufficient for conforming a good clustering [21]; evenmore, some of these keywords can lead to more confusion on the clustering process.

We have carried out a set of experiments and our results have been compared with those published earlier in this field. We have used the two corpora presented in [19] and the one suggested in [21], which we consider the most appropiate for our investigation because of their intrinsic characteristics: narrow-domain, short texts and number of documents. The two best hierarchical clustering methods reported in [19] were also implemented. Finally, we have used, as refered by [11], three different feature selection techniques in order to improve the clustering task.

The comparison between documents is performed introducing a symmetric Kullback-Leibler (KL) divergence. As the texts may differ in the terms, the frequency of many compared terms in the document will be zero. This causes problems in the KL distance computation when probabilities are estimated by frequencies of occurrence. In order to avoid this issue, a special type of back-off scheme is introduced. The next section explains into detail the use of the Kullback and Leibler distance as a similarity measure in the clustering task. In Section 3 we present the characteristics of every corpus used in our experiments, describing the use of feature selection techniques for selecting only the most valuable terms from each corpus. The description and the results obtained in our executions are presented in Section 4 and, finally the conclusions of our experiments are given.