Clustered sub-matrix singular value decomposition

Abstract

This paper presents an alternative algorithm based on the singular value decomposition (SVD) that creates vector representation for linguistic units with reduced dimensionality. The work was motivated by an application aimed to represent text segments for further processing in a multi-document summarization system. The algorithm tries to compensate for SVD's bias towards dominant-topic documents. Our experiments on measuring document similarities have shown that the algorithm achieves higher average precision with lower number of dimensions than the baseline algorithms - the SVD and the vector space model.