Abstract
In this paper we present a novel approach based on efficient text representation which employs semantic relations between words. We use singular value decomposition of the co-occurrence matrix to overcome its noise and sparseness. Thereby, we obtain a new refined co-occurrence matrix, which allows us to determine relations between words as distances in it. We use these distances as correction factors for the Bag-of-words text representation. In other words, we transform text representation vectors by inclusion relations between words. To validate our representation model, we apply it to binary classification task. We study how our model improves classification of documents, which are relevant to a given domain (topic). For this purpose, we implement Support Vector Machine and classify documents from Reuters-21578 collection. Results of our experiments demonstrate the superiority of our model.