Computational Aesthetics of the Collective Affective Dynamics of IMDB Movie Reviews

This paper proposes an approach to visualising affective metadata from movie reviews on IMDB. First, a natural language processing (NLP) method is applied for automatic topic modelling and sentiment analysis using contextual valence and subjectivity as measures of emotional expression intensity. A t-Distributed Stochastic Neighbour Embedding (t-SNE) was used to project this meta-data into a compact 2D representation. A cluster analysis was used to extract the spatial dynamics of this representation, which were mapped into a generative visualisation. Each generated visual 'signature' represented the emotional dynamics of a single movie extracted from 150 reviews, for a total of 20 movies. Based on the visualised metadata, a qualitative evaluation of 79 participants demonstrated the capacity to communicate affective metaphors, as well as a robust sense of perceptual consistency. Furthermore, we assert that the generative visualisation of data shows a nuanced expression of an aesthetic approach instead of an abstract articulation of an idea.