Maylinna Rahayu Ningsih, Jumanto Unjung
Socialmedia has become an integral and inseparable part of daily life, generating vast textual traces of public emotion yet posing challenges for large-scale analysis. Realizing that social media plays an important role in both doing business or just understanding other people's sentiments on a topic. Analyzing sentiment and emotion on social media can help identify the factors that drive people to participate in discussions in social media environments. In this paper, we propose an optimized emotion-based sentiment analysis framework using a Long Short-Term Memory (LSTM) network combined with TextVectorization for social media text. Unlike conventional sentiment analysis approaches that focus on polarity classification, this study performs fine-grained emotion classification into 15 distinct emotional categories. The optimization focuses on text preprocessing, vocabulary size, and sequence length configuration to improve classification stability on informal and unstructured social media text. Experimental results on a multi-platform social media dataset demonstrate an accuracy of 99.62% within the evaluated setting, indicating the effectiveness of the proposed configuration for emotion recognition tasks. © 2026 IEEE.
Universitas Negeri Semarang, Faculty of Mathematics and Natural Sciences, Computer Science Department, Semarang, Indonesia
