<![CDATA[The widespread use of social media, particularly Platform X, has increased the popularity of memes as a multimodal communication medium that combines textual and visual elements to express emotions, opinions, and social reactions. This study proposes a multimodal emotion classification approach for Indonesian-language memes by integrating IndoBERT for textual analysis and YOLOv11 operating in image classification mode for visual analysis through a weighted late fusion strategy. An initial dataset of 2,810 Indonesian-language memes was collected through web scraping. After removing corrupted or unreadable images, 2,547 valid samples remained. Each meme was manually annotated into one of six emotion categories—Happiness, Disgust, Anger, Sadness, Fear, and Surprise—based on the annotators' judgment of the dominant emotion conveyed by the combination of text and image, following Ekman's Basic Emotion framework. The dataset was divided using a stratified 80:10:10 split into 2,037 training, 255 validation, and 255 testing samples. The validation set was used to determine the optimal fusion weight, while the held-out test set was reserved exclusively for final evaluation. The best-performing weighted late fusion model (α = 0.6) achieved 72.55% accuracy, 73.03% macro precision, 72.76% macro recall, and 72.66% macro F1-score on the test set. Within the constructed dataset, the proposed multimodal approach outperformed the evaluated IndoBERT-only and YOLOv11-only baselines, indicating that combining textual and visual information can improve emotion classification performance for Indonesian-language meme content.]]>
