Emotion Detection in Text Using Convolutional Neural Network

One of the difficult areas of automatic language understanding is emotion detection. Research related to emotion detection in text is important to be developed so that emotion in text can be detected properly. This research develops emotion detection in text by comparing word embedding such as Word2Vec, BERT and GloVe and using Convolutional Neural Network (CNN). Five kinds of emotions are identified in the text, including happy, angry, sad, scared, and surprised. Three kinds of data are used, including commuter line, transjakarta, and commuter line+transjakarta. BERT+CNN generates the best accuracy compared to Word2Vec+CNN and GloVe+CNN. The accuracy values generated by BERT+CNN on the commuter line, transjakarta and commuter line+transjakarta data are 86.47%, 87.23%, and 86.18%, respectively. Word2Vec+CNN occupies the second best accuracy value and the last is GloVe+CNN. The comparison of Precision, Recall, and F1-Measure results also show that BERT+CNN has the highest score when compared to other methods. This demonstrates that the combination of CNN and BERT embedding methods performs well in text emotion detection.