Using Background Contextual Knowledge for Document Representation

We describe our approach to document representation that captures contextual dependencies between terms in a corpus and makes use of these dependencies to represent documents. We have tried our representation scheme for automatic document categorisation on the Reuters’ test set of documents. We achieve a precision recall break even point of 84% which is comparable to the best known published results. Our approach acts as a feature selection technique that is an alternative to applying the techniques from machine learning and numerical taxonomy.