Multi-class Textual Emotion Categorization using Ensemble of Convolutional and Recurrent Neural Network

Categorizing emotion refers to extracting the individuals’ behaviour from texts and assigning textual units into an emotion from predefined emotional connotations. Identification and categorization of emotion content have mostly been made for English, French, Chinese, Arabic, and other high-resource languages. However, very few studies have investigated emotion from the under-resourced language like Bengali. This work proposes an ensemble-based technique for classifying textual emotions into six classes: anger, disgust, fear, joy, sadness and surprise. An emotion corpus containing 9000 Bengali texts is developed to perform the emotion classification. This work investigates 22 standard classifier models developing based on three deep learning techniques (Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), Bidirectional Long Short Term Memory (BiLSTM) with different ensemble strategies and embedding models (i.e., Word2Vec, FastText). All the models are tuned, trained and tested on the developed dataset (EBEmoD-Extended Bengali Emotion Dataset) and a publicly available emotion dataset (BYCD-Bengali Youtube Comment dataset). The experimental result demonstrates that the ensemble of CNN and BiLSTM (i.e., CNN+BiLSTM) outdoes all other models by acquiring the highest weighted $$f_1$$ -score of 62.46% (for EBEmoD) and 67.57% (for BYCD), respectively.