Language independent `bag-of-words' representations are surprisingly effective for text classification. The generic BOW approach is based on a high-dimensional vocabulary which may reduce the generalization performance of subsequent classifiers, e.g., based on ill-posed principal component transformations. In this communication our aim is to study the effect of sensitivity based pruning of the bag-of-words representation. We consider neural network based sensitivity maps for determination of term relevancy, when pruning the vocabularies. With reduced vocabularies documents are classified using a latent semantic indexing representation and a probabilistic neural network classifier. Pruning the vocabularies to approximately 20% of the original size, we find consistent context recognition enhancement for two mid size data-sets for a range of training set sizes. We also study the applicability of the sensitivity measure for automated keyword generation.
|Title of host publication||Proceedings of 17th International Conference on Pattern Recognition (ICPR 2004)|
|Publication status||Published - 2004|