--- title: "AHA-Memes: A Fine-Grained Multimodal — Kmainasi 2026" description: "Hateful memes express hostility through the combination of picture, caption and shared cultural background." authors: ["Mohamed Bayan Kmainasi", "Ali Ezzat Shahroor", "Abul Hasnat", "Md. Rafiul Biswas", "Wajdi Zaghouani", "Firoj Alam"] year: 2026 venue: "arXiv:2607.27393" source: "https://arxiv.org/abs/2607.27393" doi: "10.48550/arxiv.2607.27393" arxiv: "2607.27393" licence: "CC BY 4.0 (arXiv)" licence_url: "http://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/kmainasi-2026-aha-memes-fine-grained-multimodal/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/kmainasi-2026-aha-memes-fine-grained-multimodal/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/) / AHA-Memes: A Fine-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes # AHA-Memes: A Fine-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes **Authors:** [Mohamed Bayan Kmainasi](https://md.research.elamur.ai/free/en/authors/mohamed-bayan-kmainasi/), [Ali Ezzat Shahroor](https://md.research.elamur.ai/free/en/authors/ali-ezzat-shahroor/), [Abul Hasnat](https://md.research.elamur.ai/free/en/authors/abul-hasnat/), [Md. Rafiul Biswas](https://md.research.elamur.ai/free/en/authors/md-rafiul-biswas/), [Wajdi Zaghouani](https://md.research.elamur.ai/free/en/authors/wajdi-zaghouani/), [Firoj Alam](https://md.research.elamur.ai/free/en/authors/firoj-alam/) · **Year:** 2026 · **Venue:** arXiv:2607.27393 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2607.27393) · [PDF](https://arxiv.org/pdf/2607.27393) · [DOI 10.48550/arxiv.2607.27393](https://doi.org/10.48550/arxiv.2607.27393) · [arXiv 2607.27393](https://arxiv.org/abs/2607.27393) · [OpenAlex](https://openalex.org/W7171922447) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) · [1](#cited-by) in our library, of the 140884 papers whose bibliographies we read ## Summary Hateful memes express hostility through the combination of picture, caption and shared cultural background. Arabic has had no dedicated resource for studying this at a fine level, since earlier Arabic meme collections focused on propaganda or broad harm labels. The authors build AHA-Memes from four social platforms. Their pipeline removes near-duplicate images using embedding distance, keeps only memes whose text OCR can detect, and uses a mid-sized VLM to pre-screen the pool so hateful items are more common. Native Arabic speakers then label the selected memes using bilingual guidelines. Each meme gets a binary hate decision and multi-label hate types. A further pool of roughly 66K memes carries LLM-generated silver labels intended for weak supervision, not evaluation. The benchmark covers text encoders, image encoders, late fusion, and open- and closed-weight VLMs under zero-shot, retrieval-based few-shot and fine-tuning settings. A fine-tuned open VLM does best on binary detection. Embedded text carries most of the signal and images add to it. Fine-grained typing remains difficult, and zero-shot models miss many hateful memes. The resource opens culturally grounded multimodal moderation research for Arabic. ## Key points - The gold set has 5K memes (1,850 hateful, 3,150 non-hateful). Annotation was done by three trained native-speaker annotators through a vendor and cost about $4K. The taxonomy has 8 hate types and 3 non-hate types. - Sources are Facebook, Instagram, Pinterest (manually chosen public pages) and Twitter (X) (keyword search). Near-duplicates were images with Euclidean embedding distance ≤ 3.6. Text came from EasyOCR, and memes without text were dropped. - To pick memes for annotation, Gemma-3-12B weakly pre-labeled about 71K memes. Silver labels for the ~66K auxiliary corpus came from Gemini-3.1-Pro and cover hate, hate type, target, dialect, emotion and sentiment. - Fine-tuned Qwen3-VL-8B reaches 0.768 binary macro-F1 and 0.334 hate-type macro-F1. Its zero-shot scores are 0.643 and 0.176. Arabic-centric text encoders beat mBERT and XLM-R-base on the binary task. - Few-shot ICL with RRF retrieval at K=5 gives 0.708 binary and 0.241 hate-type macro-F1, and raises Hateful recall from 0.318 to 0.454. A single demonstration (K=1) performs worse than zero-shot on binary. ## Interesting ideas - Few-shot demonstrations are chosen by merging two separate similarity rankings, one from text and one from image, with reciprocal rank fusion; this beats either ranking alone and costs no training. — For Mova (learn): When building prompts for LLM teacher tiers, retrieve example sentences by fusing ranks from lexical similarity and UD-tree-shape similarity, so the teacher sees cases that resemble the query both in wording and in structure. - A cheap model screens a large unlabeled pool so the rare class is better represented before paying for human labels, while negatives stay in the sample. — For Mova (learn): Use disagreement among the witness parsers to pick constructions that are rare or contested, and send those to human or top-tier review instead of sampling uniformly. - Gold data is kept strictly for evaluation, and the much larger LLM-labeled pool is explicitly limited to weak supervision. — For Mova (learn): Keep a hard boundary between silver treebank data used for training and a human-checked test set, so self-induced repair rules are never scored on data they helped create. - Errors are broken down by input text length, which shows difficulty rising with the amount of context and inference required, not only with lexical cues. — For Mova (answer): Report reader and QA accuracy bucketed by passage and sentence length to tell which stages fail once multi-step pragmatic inference is needed. - The hardest failures came from models that understood the content but read implicit attacks on a group as harmless humor, so the real difficulty is a normative boundary rather than perception. — For Mova (frames): The irony and implicature module should explicitly fill a target slot (who is mocked, and whether it is a group) separately from the humor flag, so that a joke and an attack can coexist in the frame. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) ## Cited by Cited by 1 paper of our library cites it (of the 140884 whose bibliographies we read); each has a page here. - [ArGuard Shared Task: Harmful Content Detection in Arabic Memes and LLM Prompts](https://md.research.elamur.ai/free/en/papers/alam-2026-arguard-shared-task-harmful-content/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{kmainasi2026aha, title = {{AHA-Memes: A Fine-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes}}, author = {Kmainasi, Mohamed Bayan and Shahroor, Ali Ezzat and Hasnat, Abul and Biswas, Md. Rafiul and Zaghouani, Wajdi and Alam, Firoj}, year = {2026}, doi = {10.48550/arxiv.2607.27393}, eprint = {2607.27393}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2607.27393} } ``` - [← previous AgriChain: Visually-Grounded Expert-Verified Reasoning for Interpretable Agricultural Vision–Language Models](https://md.research.elamur.ai/free/en/papers/mahmood-2026-agrichain-visually-grounded-expert-verified/) - [next → AI Corpus Linguist: More than a Year of Experience](https://md.research.elamur.ai/free/en/papers/milicka-2026-ai-corpus-linguist-more-than/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/) ## References 1 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Ahmed Abdelali, Hamdy Mubarak, Shammur Absar Chowdhury, Maram Hasanain, Basel Mousi et al. 2024. LAraBench: Benchmarking Arabic AI with large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 487–520, St. Julian’s, Malta. Association for Computational Linguistics. [↗](https://arxiv.org/abs/2404.14219) [our card](https://md.research.elamur.ai/free/en/papers/abdin-2024-phi-3-technical-report-highly/)