A Comparative Data-driven Study of Intensity-based Categorical Emotion Representations for MER

Data-driven analysis and modeling of music-perceived emotions have widespread applications in MIR, with representations of perceived musical emotions forming a crucial component. Though some emotion representations are popular in the literature, their relative merits and demerits in terms of expressiveness and broad applicability have been sparsely studied. The application-specific emotion representations used in multiple studies lead to incomparability of algorithms and performance metrics and non-reusability of representation-specific emotion data across studies. In this work, we study an intensity ratings-based, categorical emotion representation called Emotion-Word Intensity-Value (EWIV) representation, with emotion classes adapted from the aesthetic concept of Nava Rasa. We also introduce EmoRaga - a novel clip-set annotated with perceived emotions and emotion motifs for emotion analysis of Hindustani classical music. We explore the applicability of EWIV towards diverse MIR applications, e.g., dominant and secondary emotion identification, and temporal emotion pattern study. Last, we report a data-driven comparison of EWIV, categorical and dimensional representations, using statistical out-of-sample goodness of fit tests to measure and compare their representativeness over both benchmark datasets and collected emotion data. We conclude that EWIV is applicable to a range of MIR tasks, with higher representative and generalization potential compared to popular representations in certain cases.