<![CDATA[The rapid growth of online gambling promotion on Indonesian social media creates significant challenges for automated moderation systems, particularly because the content often appears in multimodal forms, uses slang expressions, and disguises promotional intent. The purpose of this study is to improve the accuracy and robustness of gambling advertisement detection by proposing a multimodal ensemble deep learning framework that integrates information from text, images, and audio. The method combines three independent feature streams, namely native text, OCR-extracted text from images, and ASR-generated speech transcripts. These inputs are processed using three classifiers, namely CNN, BiLSTM, and IndoBERT, which are then fused using a weighted soft-voting ensemble strategy. A dataset consisting of 12,000 multimodal samples collected from Facebook, Instagram, TikTok, and YouTube was used for evaluation. The results show that the ensemble model achieves an accuracy of 95.42 percent, outperforming each individual classifier, with substantial improvements in handling noisy OCR and ASR outputs as well as implicit gambling slang. Compared with single-model baselines, the proposed approach reduces false positives by 18.6 percent and false negatives by 22.3 percent. The novelty of this study lies in the integration of multimodal feature streams with an optimized ensemble mechanism, enabling more reliable detection of concealed gambling promotional patterns. The findings provide a strong foundation for future research on adaptive moderation systems and real-time harmful content detection in Indonesian social media.]]>
Copyrights © 2025
