Disambiguating Music Emotion Using Software Agents.
Yang Dan, Won‐Sook Lee · 2004
Annotating music poses a cognitive load on listeners and this potentially interferes with the emotions being reported. One solution is to let software agents learn to make the annotator’s task easier and more efficient. Emo is a music annotation prototype that combines inputs from both human and software agents to better study human listening. A compositional theory of musical meaning provides the overall heuristics for the annotation process, with the listener drawing upon different influences such as acoustics, lyrics and cultural metadata to focus on a specific musical mood. Software agents track the way these choices are made from the influences available. A functional theory of human emotion provides the basis for introducing necessary bias into the machine learning agents. Conflicting positive and negative emotions can be separated on the basis of their different function (reward-approach and threat-avoidance) or dysfunction (psychotic). Negative emotions have strong ambiguity and these are the focus of the experiment. The results of mining psychological features of lyrics are promising, recognisable in terms of common sense ideas of emotion and in terms of accuracy. Further ideas for deploying agents in this model of music annotation are presented.