Text-Based Emotion Detection using CNN-BiLSTM

Social media is not only used by the public to convey information but also to express emotions in writing. Text-based emotion detection is needed to find out the emotions contained in the text. This study combines CNN and BiLSTM for text-based emotion detection and compares the use of word embedding such as Word2Vec and GloVe. The data types used in this study are commuter line, transjakarta and commuter line + transjakarta. This study uses two scenarios, namely scenario I which classifies the dataset into emotion and no-emotion classes and scenario II which classifies emotions into five types of emotions, namely happiness, anger, sadness, fear, surprise. This study has two experimental types to emotion detection based on text, namely Word2Vec-CNN-BiLSTM, and GloVe-CNN-BiLSTM. In scenario I, Word2Vec-CNN-BiLSTM outperforms GloVe-CNN-BiLSTM in terms of accuracy. The accuracy values generated by Word2Vec-CNN-BiLSTM on the commuter line, transjakarta and commuter line + transjakarta data are 84.34%,83.%73%, and 83.88%, respectively. Word2Vec-CNN-BiLSTM provides the best overall Precision, Recall and F1-Measure compared with GloVe-CNN-BiLSTM on all data. The same result is also shown in scenario II where accuracy, precision, recall and F1-measure in Word2Vec-CNN-BiLSTM are better than other methods. The results of this study also improve the performance of emotion detection in text compared to the results of previous studies using Word2Vec-BiLSTM.