Pengembangan Model Klasifikasi Citra Meme Menggunakan Fitur Bi-Modal

The growth of social media has made memes a popular form of digital communication, combining text and image to convey various emotions. Sentiment analysis on memes is challenging due to their multimodal nature, where meaning can only be fully understood through the combination of both modalities. This study develops an emotion classification system for memes using multimodal approach integrating RoBERTa-large for text analysis and ReNet-50 for visual analysis. Text is extracted using OCR then cleaned and tokenized. Text features of 1024 dimensions and visual features of 2-48 dimensions are combined via a late fusion mechanism, then classified by a FCNN into seven emotion categories. The MET-Meme Dataset consist of 3.937 valid samples with oversampling applied to address class imbalance. The multimodal model achieves accuracy of 0.91, outperforming the text-only (0.89) and image-only (0.90) unimodal models.