| Abstract | Both low-level semantics of song texts and our emotional responses can be encoded in words. In order to model how we might perceive the emotional context of songs, we propose a simplified cognitive approach to bottom-up define term vector distances between lyrics and affective adjectives, which top-down constrain the latent semantics according to the psychological dimensions of valence and arousal. Projecting the lyrics and adjectives as vectors into a semantic space using LSA latent semantic analysis, their cosine similarities can be mapped as emotions over time. Subsequently we apply a three-way Tucker tensor decomposition to the derived LSA matrices, combined with a hierarchical Bayesian automatic relevance determination to find similarities across a selection of songs, and as a result identify two time series dramatic curvatures and three mixtures of affective components, which might function as emotional building blocks for generating the structure in lyrics. |