Latent semantics as cognitive components

Abstract

Cognitive component analysis, defined as an unsupervised learning of features resembling human comprehension, suggests that the sensory structures we perceive might often be modeled by reducing dimensionality and treating objects in space and time as linear mixtures incorporating sparsity and independence. In music as well as language the patterns we come across become part of our mental workspace when the bottom-up sensory input raises above the background noise of core affect, and top-down trigger distinct feelings reflecting a shift of our attention. And as both low-level semantics and our emotional responses can be encoded in words, we propose a simplified cognitive approach to model how we perceive media. Representing song lyrics in a vector space of reduced dimensionality using LSA, we combine bottom-up defined term distances with affective adjectives, that top-down constrain the latent semantics according to the psychological dimensions of valence and arousal. Subsequently we apply a Tucker tensor decomposition combined with re-weighted l1 regularization and a Bayesian ARD automatic relevance determination approach to derive a sparse representation of complementary affective mixtures, which we suggest function as cognitive components for perceiving the underlying structure in lyrics.