Detecting Multiclass Emotions from Labeled Movie Scripts

Abstract

Detecting human emotions will likely become a key component in future artificial intelligence (AI) systems, where the challenge lies in the precise discerning of negative emotions that require delicate responses such as anger and sadness. Existing sentiment tools, however, are mostly limited to dichotomous affect scales and are subject to positivity bias. To infer diverse negative emotions, this paper presents a multiclass emotion classifier that focus on negative emotions. By utilizing a rich set of both content and meta information from a labeled movie transcript, we make a novel finding that while negative emotions are hardly distinguishable from each other based on standard approaches, our non-lexical meta features remarkably increase the recall performance by 52% to 113% among the negative emotions. Our model evaluated in cross-validation studies and via human tagging demonstrate an improved performance compared to traditional baselines. This research presents a pilot study, based on small yet rich dataset, which envisions AI systems that can understand the complex negative feelings to better assist human-robot interactions.