Towards a model of computational beauty: extracting aesthetic meaning from music reviews

Recent research on cognitive neuroscience and artificial intelligence has shown how aesthetic experiences can be bound to concrete qualities of the object which causes them. When trying to deal with the challenge of whether it is possible to automatically extract such features that make a piece of music beautiful, we find ourselves restricted by a semantic problem: the one of providing a universally accepted definition of beauty. We propose to extend existing research in philosophy, neuroaesthetics, biology and computer science with a data driven approach rooted in Natural Language Processing. In particular, we try to study whether it is possible to build a model able to retrieve the main concepts addressed by music critics when they write about musical beauty. In order to do so, we first built a word embedding by training a word2vec neural network architecture on music reviews, and then tried to identify meaningful clusters in such embedding close to a list of aesthetic terms. Results, although with some limitations, show that our approach shows potential. The model appears to have succesfully learned some of the semantic relationships we were after, while other semantic relationships learned were still unclear.