Automatic Detection Methods of Repressed Anger using Deep Learning

Abstract

Japanese people tend to repress the expression of strong emotions. Conventional emotion detection systems don't detect repressed anger with sufficient accuracy. Documents containing repressed anger are unusual to express anger with all sentences. Therefore the author proposed and evaluated methods to detect emotions for each sentence and detect anger of the document by deep learning based on the emotion type and order. The accuracy rate for the a method which firstly classifies repressed or expressed anger and then classifies document into detail kind of anger improved by 10 percent points compared with the classifier which inputs the whole sentence, but a sufficient method cannot be established because there are types of anger that are never detected. However, since anger could be detected by the type and order of emotions contained in the document, it was suggested that the method of using the sentence as input to the system is effective.