Emotion classification in social media posts related to telecommunication services using bidirectional LSTM

Abstract

Social media has become a vital platform for sharing opinions, with 139 million users in Indonesia as of January 2024. Telecommunications companies can leverage feedback from platforms like Twitter and Instagram to understand customer sentiment and improve service quality. This study applies Bidirectional Long Short-Term Memory (BiLSTM) and FastText word embeddings to classify emotions in social media posts related to major Indonesian telecom providers, including Telkomsel, Indosat, XL, and Axis. Using Borderline Synthetic Minority Oversampling Technique (B-SMOTE) to address class imbalance, the model categorizes six basic emotions: happiness, sadness, fear, anger, disgust, and surprise. The optimal model, trained over 18 epochs, includes 64 BiLSTM units, 128 dense layer neurons, a 0.3 dropout rate, a batch size of 32, and a learning rate of 0.001. It achieved 93.51% accuracy and a 93.48% F1 score on unseen data, demonstrating strong performance in predicting customer emotions. This approach provides valuable insights for improving customer engagement and service in the telecommunications industry.