A Parallel Implementation of the SOM Algorithm for Visualizing Textual Documents in a 2D Plane

With the increase of computer usage, the number of digital documents is reaching values that make unviable conducting the tasks of text organization by humans. There is a demand for text organization tools that can operate with little human intervention and that can display the results of the organization in the most commonly used visual interface: the two-dimensional (2D) plane. One of the techniques used for automatic text organization is the Self Organized Map (SOM), a class of artificial neural network introduced by kohonen in 1982. With SOM, documents can be automatically arranged into groups associated with areas of a 2D plane. Most of SOM implementations are secuential. That is, implementations in which a processor carries out one instruction at once. SOM secuential implementations work fine in domains where the number of objects to organize is low. In large-scale domains like text organization, where is not rare to deal with thousands of documents and millions of words, the application of secuential SOM can become unfeasible in terms of computational time. In an attempt for improving the performance of SOM in large-scale domains, parallel implementation of SOM have been devised. Such implementations take profit of the inherent parallel architecture of SOM. Provided that in the parallel implementation of SOM there are several processors, each one carrying out one instruction at once, there is a reduction of the time required for organizing the objects of the domain. In this paper, a parallel implementation of SOM is achieved by using a Beowulf Cluster of personal computers each with low processing capacity. The speed-up gained by the cluster parallel implementation was measured. This was achieved by measuring the execution time of both a secuential implementation and a parallel implementation. The evaluation was conducted on a constant text corpus. The cluster implementation obtained will help make viable ongoing research on text organization at University of Antioquia where there is no funding for investing in costly parallel processors.