Term weighting using contextual information for categorization of unstructured text documents

During categorization of text documents, term weighting assigns appropriate weights to different terms. All the terms having equal weights have different contribution in deciding context of the document. This paper proposes a novel concept of associating positional context among regions for term weighting. For this, Dynamic Partitioning of text documents with First and Last Partitions (DynaPart-FiLa) is proposed. Experiments show that associating positional context improves F-measure by 11.9% for Reuters-21578, 23.6% for talk.* Newsgroups and 34.82% for Reuters Corpus Volume I (RCV1) in comparison to traditional term weighting scheme. The performance improvement is at the expense of small additional storage cost.