Advancing Multimodal Hate Speech Detection in Internet Memes with Large Vision-Language Models

Online hate is increasingly spread through memes, where harmful meaning comes from the combination of images and text, making it harder to detect than text-only abuse. Through a 12-week research visit to the University of Hildesheim, I will develop and evaluate efficient AI models that can better identify hateful memes across multiple datasets while remaining practical to run at scale. Building on my prior work with Prof. Thomas Mandl’s group and my graduate training at McGill, I will systematically compare leading open-source multimodal models, test lightweight improvements that help capture subtle and context-dependent hate, and document common failure cases to guide model improvements, and will also help refine evaluation practices so results are consistent and reusable for follow-on research. The project will deliver reproducible experiments and publishable findings, strengthening international collaboration between McGill University and the University of Hildesheim and supporting both institutions’ research in trustworthy AI and online safety.

Faculty Supervisor:

Derek Nowrouzezahrai

Student:

Partner:

University of Hildesheim

Discipline:

Computer science

Sector:

Artificial Intelligence; Information and Communications Technology (ICT); Cyber Security

University:

McGill University

Program: