Using Weights with a Text Proximity Matrix
Angel R. Martinez, Edward J. Wegman, Wendy L. Martinez · 2004
In previous work, we introduced a way of encoding free-form documents called the bigram proximity matrix (BPM). When this encoding was used on a corpus of documents, where each document is tagged with a topic label, results showed that the documents could be classified based on their tagged meaning. In this paper, we investigate methods of weighting the elements of the BPM, analogous to the weighting schemes found in natural language processing. These include logarithmic weights, augmented normalized frequency, inverse document frequency and pointwise mutual information. Results presented in this paper show that some of the weights increased the proportion of correctly classified documents.