Emotion Analysis based on Incremental Online Learning in Social Networks

Social network messages contain reliable information about the emotions that authors have experienced and expressed; based on them, we can conclude about some psycho-emotional characteristics of their authors. In this work, we use the model of emotions proposed by P. Ekman. We present a new iterative learning model that classifies messages from social networks into six basic emotions: joy, sadness, fear, anger, disgust, and surprise. The model considers not only “plain text” as input, but also various emotional indicators: emoticons and emoji. As a result of incremental learning, the averaged performance of the model increased by 8.9% in ROC AUC and by 7.3% in F1-score. The novelty of the approach lies in the implementation of incremental learning on the raw data permanently accepted from the social networks. It helps to maintain the model updated and improve its characteristics.