Towards Sentic-Aware Multimodal Models for Cyberbullying Detection in Thai Memes

Cyberbullying has become an increasingly urgent issue in online communities. Memes, a popular form of online expression, often blend text and imagery in emotionally charged, sarcastic, or offensive ways–posing unique challenges for automatic harmful content detection. This work explores sentic-aware multimodal models for cyberbullying detection in Thai memes, with a focus on integrating affective commonsense knowledge through SenticNet-based features that emphasize conceptual reasoning and structured emotion representation. To enable this, we propose ThaiSenticNet 7, a resource adapted for the Thai language by translating from SenticNet 7, which supports the generation of sentic features. We investigate three representations–sentic vectors, sentic spectrograms, and sentic mel-spectrograms–and their integration with various sequential models to form sentic embeddings. These embeddings are fused with textual and visual information, extracted via a fine-tuned WangchanBERTa and a Swin Transformer, respectively, forming a unified multimodal pipeline. Experiments on a curated Thai meme dataset show that incorporating sentic features significantly enhances classification performance, with the best configuration–combining all three modalities–achieving an F1-score of 0.8044. Notably, the mel-spectrogram transformation proves particularly effective, suggesting that frequency-domain encoding helps capture subtle affective transitions in text-derived emotional signals. Our findings highlight the value of affective knowledge and multimodal modeling in tackling harmful content in memes.