Projects

HVGuard detects hateful videos, where harm is often carried jointly by speech, visuals, sarcasm, and pacing. It feeds transcripts and video frames to a multimodal LLM pipeline that reasons over evidence across modalities, improving detection of implicit, context-dependent hate in short videos.

HMGuard detects harmful memes, where a small amount of text and imagery can carry hateful, harassing, or propagandistic intent. It uses multimodal LLMs to reason about how the image, the overlaid text, and their cultural references combine, treating moderation as a meme-understanding task.

UGCG-Guard detects illicit promotion that lures users, especially minors, into unsafe user-generated content games. It uses large vision-language models to screen social posts, screenshots, and game imagery for sexualized or exploitative content, despite platform-specific slang and fast-changing promotion styles.

NewWave moderates hate speech that surges around breaking events, when static policies and existing classifiers go stale. It uses chain-of-thought reasoning in LLMs to capture the narratives, slogans, and references tied to each new wave, so detection can adapt without retraining a model for every shift.

LLM4HateSpeech evaluates whether LLMs can detect hate speech in realistic, context-heavy settings. It compares prompting strategies and measures how context, task framing, and external knowledge affect LLM judgments on subtle or ambiguous hateful language, showing where LLMs help and where they remain brittle.