Social media platforms like X generate massive volumes of user-generated text containing diverse communication styles, including humorous and serious expressions that are difficult to distinguish due to their reliance on contextual and semantic nuance. This study proposes a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model to classify humorous and serious communication styles in Indonesian social media text, comparing its performance against a single CNN model. The dataset was drawn from the SmSA subset of the IndoNLU benchmark on Hugging Face, comprising 12,679 Indonesian texts. Since the corpus was originally designed for sentiment analysis, the texts were re-labeled into humorous and serious styles using a reproducible rule-based approach. After preprocessing and balancing, the final dataset contained 10,000 samples (5,000 humorous and 5,000 serious) split 80:20 for training and testing. Experimental results show that while the standalone CNN achieved 91.60% accuracy and F1-score, the hybrid CNN-LSTM model demonstrated superior performance, reaching an accuracy and F1-score of 94.10%. These findings confirm that combining CNN for local feature extraction with Bidirectional LSTM for sequential context modeling produces more discriminative text representations. Within the scope of this relabeled dataset, the proposed hybrid model provides an effective framework for automated communication style classification and serves as a strong foundation for future Indonesian social media text analysis.
Copyrights © 2026