Black-Box Emotion Detection: On the Variability and Predictive Accuracy of Automated Emotion Detection Algorithms

The current research demonstrates considerable variability in predictive accuracy across major emotion detection systems (such as Google ML or Microsoft Cognitive Services) with lower (higher) classification accuracy for negative (positive) discrete emotions. We provide two modeling strategies to improve prediction accuracy by either combining feature sets or using ensemble methods.