Improving the BiLSTM model performance for tweet sentiment analysis

Abstract

Twitter is a microblogging website where users can publish brief entries known as tweets. These tweets can occasionally reveal the users' attitudes and feelings. This paper analyses three emoticon processing methods with the BiLSTM model to discover the efficiencies of different methods in helping deep learning models classify tweet sentiments. Firstly, the simply removing method, the replacing with description text method, and the replacing with predefined sentiment method are established. Then the BiLSTM model is used to train and test with different methods on the Sentiment140 dataset. The performances of all models are evaluated by accuracy, F1 score, precision score, and recall score. The experimental results show that the replacing with predefined sentiments method provides the highest accuracy which is 0.84. The simply removing method also produces the testing accuracy as 0.84, but it performs worse in the last epoch, the training and validation accuracy, and the training and validation loss. The replacing with description text method produces the worst accuracy which is 0.83. It indicates that predefining the most possible sentiments of the popular emoticons has a reliable efficiency in optimizing the performance of deep learning models when the tweets with emoticons take a small proportion.