<![CDATA[Text on social media platforms such as Twitter generates large volumes of data that can be utilized to automatically understand user emotions. However, the characteristics of Twitter text, which are short, unstructured, and dominated by informal language and the use of emojis, pose significant challenges for emotion detection, particularly in Indonesian-language texts. This study aims to develop and evaluate a specific preprocessing pipeline to improve the performance of deep learning–based emotion classification models. The proposed preprocessing pipeline includes social media text cleaning, informal language and slang normalization, removal of irrelevant characters, and emoji-to-text conversion, while word representation is performed using FastText word embedding. An experimental method is employed to compare the performance of three deep learning architectures, namely Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and a hybrid CNN–GRU model, in predicting five emotion categories: anger, joy, sadness, fear, and love. The dataset consists of 5,079 tweets, which are divided using a stratified split with an 80:20 ratio between training and testing data. Model performance is evaluated using accuracy, precision, recall, and F1-score metrics. The results show that the CNN with FastText model outperforms the other models, achieving an accuracy of 81.2%, precision of 81.1%, recall of 81.9%, and an F1-score of 81.2%, confirming that a specifically designed preprocessing pipeline plays a crucial role in improving emotion detection accuracy for Indonesian Twitter text.]]>
Copyrights © 2026
