Subspaces of text discrimination with application to biological literature

This paper is about the application of statistical pattern recognition techniques to the classification of text with the objective of retrieving documents relevant for the construction of gene networks. We start from the usual practice of representing a document, electronically available abstracts of scientific papers in this case, as a high dimensional vector of term of occurrences. We consider the problem of retrieving documents corresponding to the metabolic pathway of the organism yeast, Saccharomyces Cerevisiae, using a trained classifier as filter. We use support vector machines (SVMs) as classifiers and compare techniques for reducing the dimensionality of the problem: latent semantic kernels (LSK) and sequential forward selection (SFS). In order to deal with the issue of having only a small set of accurately labelled documents, we used the approach of transductive inference. In this case, LSK leads to a subspace formed as a linear combination of features (terms in the lexicon) while SFS selects a subset of the dimension. We find, for this problem, that the discriminant information appears to lie in a subspace, which is very small in dimensionality compared to that of the original formulation. By matching against the gene ontology (GO) database, we further find that the selection process (SFS) picks out the discriminant terms that are of biological significance for this problem.