An Application of Latent Semantic Analysis for Text Categorization

Abstract

It is a challenge task to discover major topics from text, which provide a better understanding of the whole corpus and can be regarded as a text categorization problem. The goal of this paper is to apply latent semantic analysis (LSA) approach to extract common factors that representing concepts hidden in a large group of text. LSA involves three steps: the first step is to set up a term-document matrix; the second step is to transform the term frequencies into a term-document matrix using various weighting schemes; the third step performs singular value decomposition (SVD) on the matrix to reduce the dimensionality. The reduced-order SVD is the best k-dimensional approximation to the original matrix. The experiment uses more than fifteen hundreds research paper abstracts from a specific field. Because different factor solutions of the LSA suggest different levels of aggregation, this work examines thirteen solutions in the experiment. The results show that LSA is able to identify not only principle categories, but also major themes contained in the text.