Multi-channel Embedding LSTM for Sentiment Classification in Twitter Dataset

Twitter, a popular social media platform, used to express opinions and thoughts in the form of short messages. Because of this popularity, the sentiment classification on Twitter is a challenge for understanding the feelings of users. Word embedding is crucial in sentiment classification to capture contextual information. Multi-channel embedding is a solution to overcome the limitations of word embedding. Multi-channel embedding has been utilized in previous studies using multi-channel embedding CNN. However, the use of CNN methods is still not optimal for capturing sequential dependencies of texts. In previous studies, the use of LSTM for sentiment classification has proven to be better than the CNN method. To address this, the multi-channel embedding LSTM method is proposed to improve the classification performance of the previous method. The model combines GloVe and FastText embedding to capture global semantic relationships and subword-level information. LSTM is also utilized to capture sequential dependencies of text. After experiment using hyperparameter variation, multi-channel embedding LSTM achieved the best performance with epoch of 5. The best accuracy achieved at dropout rate of 0.5, while the best loss achieved without applying dropout rate. Compared to previous research, multi-channel embedding LSTM outperformed the previous method, multi-channel embedding CNN, with the best accuracy and loss of 82.96% and 0.382, respectively. These results demonstrate the effectiveness of multi-channel embedding LSTM over previous models in social media text classification.