Unsupervised Text Classification as Topic Annotation

Abstract

Multi-class Text Classification is the task of classifying a given text into one or more than one classes taken form a set of predefined classes. A class can be a topic of a text, for example, a class of any text about a movie can be ``entertainment’’. In this research I investigate unsupervised learning to accurately identify the topic of a given text. The cost involved in labeling a large amount of data and availability of huge amount of unlabeled data makes unsupervised learning an ideal choice. The probabilistic algorithm used for text classification can be termed as topic modeling and is capable to extract multiple topics within a single text of a document. LDA model used in this report exploits co-occurrence patterns of words in documents to extract semantically meaningful probabilistic clusters of words called topics .Each of those clusters is labeled using the significant terms selected in each cluster. Semantic distance between the significant terms from the clusters and Wikipedia documents is measured to identify labels for each cluster.