Non-Acted Text and Keystrokes Database and Learning Methods to Recognize Emotions

The modern computing applications are presently adapting to the convenient availability of huge and diverse data for making their pattern recognition methods smarter. Identification of dominant emotion solely based on the text data generated by humans is essential for the modern human-computer interaction. This work presents a multimodal text-keystrokes dataset and associated learning methods for the identification of human emotions hidden in small text. For this, a text-keystrokes data of 69 participants is collected in multiple scenarios. Stimuli are induced through videos in a controlled environment. After the stimuli induction, participants write their reviews about the given scenario in an unguided manner. Afterward, keystroke and in-text features are extracted from the dataset. These are used with an assortment of learning methods to identify emotion hidden in the short text. An accuracy of 86.95% is achieved by fusing text and keystroke features. Whereas, 100% accuracy is obtained for pleasure-displeasure classes of emotions using the fusion of keystroke/text features, tree-based feature selection method, and support vector machine classifier. The present work is also compared with four state-of-the-art techniques for the same task, where the results suggest that the present proposal performs better in terms of accuracy.