AHA-MEMES: A Fine-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes

As hateful memes become a pervasive form of online harm, their detection requires understanding the complex interplay between text, images, and cultural context. While English-centric resources have advanced, Arabic remains significantly underserved. This paper addresses this gap by creating a dedicated benchmark to study how hate is constructed and targeted in Arabic memes.

The authors developed AHA-MEMES, a dataset curated from major social media platforms (Facebook, Instagram, Pinterest, and X). The core of the dataset consists of 5,000 manually annotated memes, categorized by binary hate labels and fine-grained attack strategies (e.g., mocking, incitement, dehumanization). The annotation process involved native Arabic speakers using bilingual guidelines to capture dialectal nuances. To support broader research, the authors also released 66,000 auxiliary memes with silver labels generated by an LLM.

The study establishes strong baselines across five model families, including unimodal encoders, multimodal fusion models, and both open- and closed-weight vision-language models (VLMs). Key findings include:

• Fine-tuned open-weight VLMs (such as Qwen3-VL-8B) outperform other architectures in binary hate detection.
• Textual information remains the primary signal for detection, though visual information provides essential complementary context.
• Fine-grained classification of hate types remains a significant challenge, highlighting the need for models that can better reason about cultural and rhetorical framing.

By providing a large-scale, fine-grained resource, this work enables researchers to move beyond coarse binary classification. It facilitates the development of more robust, culturally aware AI systems capable of identifying nuanced forms of online harm in Arabic, which is critical for maintaining safer digital spaces in the region.