Using NLP to efficiently visualize text collections with SOMs

Self-Organizing Maps (SOMs) are a good method to cluster and visualize large collections of text documents, but they are computationally expensive. In this paper, we investigate ways to use natural language parsing of the texts to remove unimportant terms from the usual bag-of-words representation, to improve efficiency. We find that reducing the document representation to just the heads of noun and verb phrases does indeed reduce the heavy computational cost without degrading the quality of the map, while more severe reductions which focus on subject and object noun phrases degrade map quality.