Text Classification Aided by Clustering: a Literature Review

Supervised and unsupervised learning have been the focus of critical research in the areas of machine learning and artificial intelligence. In the literature, these two streams flow independently of each other, despite their close conceptual and practical connections. In this work we exclusively deal with the text classification aided by clustering scenario. This chapter provides a review and interpretation of the role of clustering in different fields of text classification with an eye towards identifying the important areas of research. Drawing upon the literature review and analysis, we discuss several important research issues surrounding text classification tasks and the role of clustering in support of these tasks. We define the problem, postulate a number of baseline methods, examine the techniques used, and classify them into meaningful categories. A standard research issue for text classification is the creation of compact representations of the feature space and the discovery of the complex relationships that exist between features, documents and classes. There are several approaches that try to quantify the notion of information for the basic components of a text classification problem. Given the variables of interest, sources of information about these variables can be compressed while preserving their information. Clustering is one of the approaches used in this context. In this vein, an important area of research where clustering is used to aid text classification is the area of dimensionality reduction. Clustering is used as a feature compression and/or extraction method: features are clustered into groups based on selected clustering criteria. Feature clustering methods create new, reduced-size event spaces by joining similar features into groups. They define a similarity measure between features, and collapse similar features into single events that no longer distinguish among their constituent features. Typically, the parameters of the cluster become the weighted average of the parameters of its constituent features. Two types of clustering have been studied: i) one-way clustering, i.e. feature clustering based on the distributions of features in the documents or classes, and ii) coclustering, i.e. clustering both features and documents. A second research area of text classification where clustering has a lot to offer, is the area of semi-supervised learning. Training data contain both labelled and unlabelled examples. Obtaining a fully labelled training set is a difficult task; labelling is usually done using human expertise, which is expensive, time consuming, and error prone. Obtaining unlabelled data is much easier since it involves collecting data that are known to belong to