Selecting training samples from large and noisy corpora for efficient text classification

Abstract

In this thesis, an algorithm is presented that selects samples of documents for training text classifiers. Often the number of documents is very large and the documents are noisy. Both for efficiency purposes and accuracy purposes, one need good samples not just blind samples such as that of simple random sampling. The proposed algorithm is far superior to simple random sampling both for small sampling ratios and in the presence of noise. The proposed algorithm is based on a simple fact that the terms in the set of training sample documents should have approximately equal document frequency as in the whole set (not including the test set).