Design of Hybrid Deep Learning Model for Text-based Emotion Recognition

Emotion Detection is part of sentiment analysis which detects emotions at a finer grain level of classifying the exact emotions instead of classifying them in general as positive and negative emotions. Emotion detection research is a recent field where emotions are detected from facial images, video, audio, and text. This paper focuses on emotion detection from textual data. Here this work proposes a hybrid model built using the doc2vec model, CNN model, Bi-GRU, and attention mechanism. In the proposed model initially, a Doc2vec model is used to transform the text into vectors. The doc2vec model captures the semantic information and this information is fed into the CNN model. The CNN model captures the local features. Bi-LSTM, GRU, and Bidirectional Gated Recurrent Unit (Bi-GRU) are used to learn the contextual features. The built model performance is fine-tuned by using an attention mechanism and dropout regularization. The model was tested with the standard benchmark datasets IMDB and Emotions dataset. The proposed hybrid model produced an accuracy of 82.84% for multi-class classification.