Emotion Detection using Deep Learning
Yasmin Shaaban, Hoda Korashy, Walaa Medhat · 2021
Emotion detection is one of the challenging tasks in Natural Language Processing. The applications of emotion detection can be utilized in different fields like data mining, psychology, and human-computer interaction. The paper aims to train different learning- based models in a supervised framework. In this paper, we have investigated different learning-based methods used in emotion detection. The methods are artificial neural networks and deep learning methods to investigate the best solution for emotion detection problem. The algorithms tested are neural networks approach as Perceptron, Multilayer Perceptron, and deep learning approach as CNN-LSTM, CNN-BiLSTM, CNN-GRU, CNN-BiGRU, BiLSTM and CNN using different features representation methods like TF-IDF, N-Grams, word-based embeddings, and contextualized embeddings. We have tested the algorithms on ISEAR dataset. The results show that the features representation method, applying preprocessing on dataset or not, and the number of emotion classes classified have impact on the models' performance. The results also show that BERT and DNN dense layer as a classifier outperform all other approaches by achieving 0.71 macro-average f1-measure for 7 emotions, 0.76 macro-average f1-measure for 5 emotions and 0.8 macro- average f1-measure for 4 emotions.