Inferring highly-dense representations for clustering broadcast media content

Abstract

We propose to employ a low-resolution representation for accurately categorizing spoken documents. Our proposed approach guarantees document clusters using a highly dense rep resentation. Performed experiments, using a dataset from a German TV channel, demonstrate that using low-resolution concepts for representing the broadcast media content allows obtain ing a relative improvement of 70.4% in terms of the Silhouette coefficient compared to deep neural architectures.