A Higher-Dimensional Expansion of Affective Norms for English Terms for Music Tagging

The Valence, Arousal and Dominance (VAD) model for emotion representation is widely used in music analysis. The ANEW dataset is composed of more than 2000 emotion related descriptors annotated in the VAD space. However, due to the low number of dimensions of the VAD model, the distribution of terms of the ANEW dataset tends to be compact and cluttered. In this work, we aim at finding a possibly higher-dimensional transformation of the VAD space, where the terms of the ANEW dataset are better organised conceptually and bear more relevance to music tagging. Our approach involves the use of a kernel expansion of the ANEW dataset to exploit a higher number of dimensions, and the application of distance learning techniques to find a distance metric that is consistent with the semantic similarity among terms. In order to train the distance learning algorithms, we collect information on the semantic similarity from human annotation and editorial tags. We evaluate the quality of the method by clustering the terms in the found high-dimensional domain. Our approach exhibits promising results with objective and subjective performance metrics, showing that a higher dimensional space could be useful to model semantic similarity among terms of the ANEW dataset.