Deriving and validating emotional dimensions from textual data

This paper proposes and analyzes a methodology for extracting the underlying emotional dimensions connected to different textual data, including social-media posts and online reviews. Our experiments result in a coherent conclusion across all 16 studied datasets. In particular, the found orthogonal emotional dimensions are a combination of valence (positive–negative sentiment), activation arousal (arousal–dominance), and expectancy tension (the intensity of the expectations concerning the future). We confirm the existence of both valence and arousal as core dimensions. On the other hand, dominance appears as an attribute connected to the variability of both valence and activation arousal dimensions. We also find some evidence for the existence of an “unpredictability/novelty” dimension discussed in recent academic work. Our key empirical contribution is that an additional orthogonal emotional dimension should be defined and named “expectancy tension” in that it captures the variability linked to the intensity of expectations regarding the future. Finally, our work contributes to the social computing literature by suggesting a novel methodology to derive emotional spaces from multiple textual data through eigenvector analyses.