Detection of multiple emotions in texts using a new deep convolutional neural network

Identifying emotions from the texts can be used in almost every aspect of our daily lives, such as improving computer-human interactions, monitoring people’s mental health, or modifying/improving business strategies based on customers’ emotions. Deep learning techniques have performed well compared to other machine learning methods in all learning problems. All existing machine learning methods for recognizing emotions are taught on datasets that includes single-emotional and multi-emotional samples. Our observations of working with these techniques show that these systems tend to learn more single-emotion samples than multi-emotion samples in a dataset. We also looked at a large number of texts and found that the number of texts from which only one emotion can be deduced is very small compared to texts from which more than one emotion can be deduced. Therefore, in general, the accuracy of existing methods is low. To deal with these two limitations, in this paper, we first created a dataset using all available data so that all texts have at least two emotions. To improve accuracy, because the convolutional neural network (CNN) has performed so well in image processing, we have then proposed a new CNN architecture for extracting emotions from texts. To find semantic, syntactic, and word similarity, we have used fast text and GloVe to embed text data into numeric representation. We have also improved the accuracy of the proposed architecture by using the attention property. The results demonstrated that the proposed model is more accurate compared to the existing methods.