Measuring conceptual incongruity from text-based annotations

Abstract

We propose a method for measuring the conceptual incongruity of a digital object using associated text meta-data. We show that this measure correlates well with empirical creativity ratings elicited from human subjects in laboratory settings. Extending our focus to online resources, we show that the predicted incongruity of a movie plot in the Movielens database is weakly correlated with users’ ratings for the movie, but strongly correlated with variability in ratings. Movies with incongruous plots appear to elicit much more polarized responses. Further, in domains where cognitive theories suggest users are likely to be looking for incongruity, e.g. humor, we show, using the Youtube Comedy Slam Dataset, that user ratings for comedy pieces are considerably well-predicted by their incongruity score. These evaluations provide convergent evidence for the validity of our incongruity measurement, and immediately present several direct application possibilities. We present a case example of including incongruity as a recommender system metric to diversify the set of suggestions made in response to user queries in ways that align with users’ natural curiosity.