Implicit Emotion Detection from Text with Information Fusion



Computer Engineering Department, Alzahra University, Tehran, Iran


In this paper we have proposed an approach for emotion detection in implicit texts. We have introduced a combinational system based on three subsystems. Each one analyzes input data from a different aspect and produces an emotion label as output. The first subsystem is a machine learning method. The second one is a statistical approach based on vector space model (VSM) and the last one is a keyword-based subsystem with an information fusion component to aggregate the final output of main system. We analyzed the performance of our proposed system on ISEAR dataset with seven emotions: anger, joy, sad, shame, fear, disgust and guilt. The results show that our combinational system outperforms each subsystem with overall f-measure of 0.68 and at least up to 0.71 in terms of F1 in emotion level except for anger. The overall performance of the main system is 9.13% better than machine learning module, 16.6% better than VSM and 23% better than keyword-based.