Multimodal Framework for Detecting Implicit Hate Speech on Social Media

Designed a hybrid framework combining text encoders (BERT, LLaMA) with vision backbones (Swin Transformer, ViT) to detect subtle hate speech.

Sept 2024 – Apr 2025 Koç University (Repository Private)
  • Designed a hybrid framework combining text encoders (BERT, LLaMA) with vision backbones (Swin Transformer, ViT) to detect subtle hate speech.
  • Improved robustness via focal loss, progressive unfreezing, and adversarial training.
  • Demonstrated superior performance on multimodal hate speech datasets, addressing explainability and ethical implications.