A Two Step Data Mining Approach for Amharic Text Classification

Abstract

Traditionally, text classifiers are built from labeled training examples (supervised). Labeling is usually done manually by human experts (or the users), which is a labor intensive and time consuming process. In the past few years, researchers have investigated various forms of semi-supervised learning to reduce the burden of manual labeling. In this paper is aimed to show as the accuracy of learned text classifiers can be improved by augmenting a small number of labeled training documents with a large pool of unlabeled documents. This is important because in many text classification problems obtaining training labels is expensive, while large quantities of unlabeled documents are readily available. In this paper, intended to implement an algorithm for learning from labeled and unlabeled documents based on the combination of Expectation- Maximization (EM) and two classifiers: Naive Bayes (NB) and locally weighted learning (LWL). NB first trains a classifier using the available labeled documents, and probabilistically labels the unlabeled documents while LWL uses a class of function approximation to build a model around the current point of interest. An experiment conducted on a mixture of labeled and unlabeled Amharic text documents showed that the new method achieved a significant performance in comparison with that of a supervised LWL and NB. The result also pointed out that the use of unlabeled data with EM reduces the classification absolute error by 27.6%. In general, since unlabeled documents are much less expensive and easier to collect than labeled documents, this method will be useful for text categorization tasks including online data sources such as web pages, e-mails and news group postings. If one uses this method, building text categorization systems will be significantly faster and less expensive than the supervised learning approach.