Exploration of a collection of documents in neuroscience and extraction of topics by clustering

This paper presents a preliminary analysis of the neuroscience knowledge domain, and an application of cluster analysis to identify topics in neuroscience. A collection of posters presented at the Society for Neuroscience (SfN) Annual Meeting in 2006 is first explored by viewing existing topics and poster sessions using multidimensional scaling. Based on the Vector Space Model, several Term Spaces were built on the basis of a set of terms extracted from the posters' abstracts and titles, and a set of free keywords assigned to the posters by their authors. The ensuing Term Spaces were compared from the point of view of retrieving the genuine category titles. Topics were extracted from the abstracts of posters by clustering the documents using a bisecting k-means algorithm and selecting the most salient terms for each cluster by ranking. The terms extracted as topic descriptors were evaluated by comparing them to existing titles assigned to thematic categories defined by human experts in neuroscience. A comparison of two approaches for terms ranking (Document Frequency and Log-Entropy) resulted in better performance of the Log-Entropy scores, allowing to retrieve 31.0% of original title terms in clustered documents (and 37.1% in original thematic categories).