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

# This paper contains examples that may be disturbing to readers.
Source: [https://arxiv.org/html/2607.27393](https://arxiv.org/html/2607.27393)
## AHA\-Memes: A Fine\-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes Warning: This paper contains examples that may be disturbing to readers\.

Mohamed Bayan Kmainasi1, Ali Ezzat Shahroor1, Abul Hasnat2, Md\. Rafiul Biswas3, Wajdi Zaghouani4, Firoj Alam1 1Qatar Computing Research Institute, Qatar,2APAVI\.AI, France 3Hamad Bin Khalifa University, Qatar,4Northwestern University in Qatar, Qatar \{mkmainasi, alsh34060, mbiswas, fialam\}@hbku\.edu\.qa mhasnat@gmail\.com, wajdi\.zaghouani@northwestern\.edu

###### Abstract

Hateful memes are a growing form of multimodal online harm, where hostile intent is often conveyed through the joint interpretation of images, text, cultural references, and implicit targets\. While hateful meme detection has advanced in high\-resource languages, Arabic remains underexplored, with existing meme resources focusing mainly on propaganda or coarse harmful\-content labels\. We introduceAHA\-Memes,111AHA\-Memes:ArabicHAtefulMemeswhich is, to our knowledge, the first large\-scale Arabic hateful meme benchmark with fine\-grained, multi\-label annotations\. The dataset includes 5K manually annotated memes using a taxonomy that captures hate types, i\.e\., attack strategies\. We further provide∼\\sim66K silver\-labeled memes to support future studies\. We benchmark text\-only, image\-only, and late\-fusion multimodal models, as well as few\-shot in\-context learning \(ICL\) and open\- and closed\-weight Vision\-Language Models \(VLMs\) under zero\-shot and fine\-tuning settings\. Our results establish strong baselines and highlight key challenges in culturally grounded Arabic hateful meme detection\. We release the dataset, annotation guidelines, and evaluation scripts to support future research\.222Project resources:[GitHub](https://github.com/MohamedBayan/AHA-Memes)—[Hugging Face](https://huggingface.co/datasets/QCRI/Arabic-Hateful-Memes)

AHA\-Memes: A Fine\-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes Warning: This paper contains examples that may be disturbing to readers\.

Mohamed Bayan Kmainasi1, Ali Ezzat Shahroor1, Abul Hasnat2,Md\. Rafiul Biswas3, Wajdi Zaghouani4, Firoj Alam11Qatar Computing Research Institute, Qatar,2APAVI\.AI, France3Hamad Bin Khalifa University, Qatar,4Northwestern University in Qatar, Qatar\{mkmainasi, alsh34060, mbiswas, fialam\}@hbku\.edu\.qamhasnat@gmail\.com, wajdi\.zaghouani@northwestern\.edu

## 1Introduction

A smiling family photograph may appear harmless on its own, yet the same image can communicate hate when combined with a targeted caption\. Likewise, a national flag paired with an apparently benign proverb may become sectarian or xenophobic depending on the communities, histories, and stereotypes that viewers bring to the meme\. Memes therefore pose a distinctly multimodal and culturally grounded hate detection problemShahrooret al\.\([2026](https://arxiv.org/html/2607.27393#bib.bib32)\)\. Their meaning is often produced through the joint interpretation of image, text, and shared background knowledge, rather than by either modality in isolationSharmaet al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib27)\)\. This dependence on cultural context also complicates human judgment, reduces label consistency, and makes automated detection more challenging\. In a cross\-country study,Buiet al\.\([2025](https://arxiv.org/html/2607.27393#bib.bib37)\)found that annotators from five countries agreed on meme labels only 74% of the time, with agreement falling to 67% between U\.S\. and Indian annotators\. Such disagreement among human annotators suggests that models trained mainly on culturally homogeneous and English\-centric resources are likely to struggle when hate is expressed through culturally specific visual and linguistic cues\. Figure[1](https://arxiv.org/html/2607.27393#S1.F1)illustrates this with two examples fromAHA\-Memes: adoctored portrait that mocks a woman’s appearance \(*Hateful*\), and a television still repurposed as everyday*sarcasm*\(*Not\-Hateful*\), where the hateful or benign reading emerges only from the image–text pairing\.

![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/armeme_examples.png)Figure 1:Examples from theAHA\-Memesdataset across hateful and non\-hateful categories\.Despite growing interest in multimodal hate detection, existing meme benchmarks remain focused in English and often rely on binary or coarse labelsKielaet al\.\([2020](https://arxiv.org/html/2607.27393#bib.bib45)\)\. This limits their ability to capturewho is targeted,what form of hate is expressed, andhow hateful meaning is constructedthrough the interaction of image, text, and cultural context\. Recent multilingual efforts have begun to broaden coverage, but available resources remain limited in scale and granularity\. For example, Multi3Hate provides parallel multilingual annotations but only 300 samples per languageBuiet al\.\([2025](https://arxiv.org/html/2607.27393#bib.bib37)\), while GuardHarMem introduces fine\-grained meme hate labels only for EnglishEl\-amranyet al\.\([2025](https://arxiv.org/html/2607.27393#bib.bib41)\)\. Arabic is especially underserved in this space\. Although Arabic text\-based hate\-speech detection has advanced through shared tasks such as OSACTMubaraket al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib53)\), and Arabic meme research has studied propaganda in ArMemeAlamet al\.\([2024a](https://arxiv.org/html/2607.27393#bib.bib35)\); Kmainasiet al\.\([2025a](https://arxiv.org/html/2607.27393#bib.bib31)\), there is still no dedicated benchmark for fine\-grained hate detection in Arabic memes\.

We address this gap withAHA\-Memes, a manually annotated fine\-grained multimodal dataset for Arabic hateful memes\. The dataset includes hateful and non\-hateful memes and provides multi\-label annotations for hate types\. Figure[2](https://arxiv.org/html/2607.27393#S2.F2)gives an overview of the end\-to\-end construction pipeline\. We conduct detailed experiments on broad set of systems, including text encoders, image encoders, multimodal fusion models, open\-weight VLMs, and closed\-weight VLMs and LLMs\. Our experiments show that fine\-tuned open VLMs perform best for binary detection, that embedded text carries much of the signal while visual information remains complementary, and that fine\-grained hate categorisation remains challenging\. We also evaluate retrieval\-augmented few\-shot in\-context learning as a training\-free alternative\. Our contributions are as follows\.

- •We introduceAHA\-Memes, the first large\-scale Arabic hateful\-meme dataset, consisting of*5K manually annotated*memes with binary hate labels and fine\-grained multi\-label hate types\. In addition, we provide silver labels333By silver labels, we refer to labels generated by an LLM\. For this task we used Gemini\-3\.1\-Pro\.for targeted protected categories and features such as dialect, emotion, and sentiment\.
- •We develop an Arabic\-contextual bilingual annotation scheme for binary hate labels, fine\-grained hate types, and targeted protected categories, which can support future dataset development for Arabic hateful\-meme analysis\.
- •We provide a comprehensive benchmark across five model families, comparing text\-only, image\-only, multimodal fusion, open\-weight VLM, and closed\-weight VLM/LLM systems under zero\-shot, few\-shot, and fine\-tuning settings\.
- •We release an additional∼\\sim66K Arabic meme corpus with silver labels for binary hate, fine\-grained hate type, target category, dialect, emotion, and sentiment, enabling future work on large\-scale Arabic multimodal harmful\-content modelling\.

## 2Related Work

##### Hateful Meme Detection and Resources\.

The Facebook Hateful Memes ChallengeKielaet al\.\([2020](https://arxiv.org/html/2607.27393#bib.bib45)\)established hateful\-meme detection as a multimodal task that requires joint reasoning over image and text\. Its use of benign confounders showed that neither modality alone is sufficient, motivating later work on cross\-modal fusion, visual\-to\-text prompting, and LLM\-based reasoning for meme interpretationKumar and Nandakumar \([2022](https://arxiv.org/html/2607.27393#bib.bib46)\); Caoet al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib38)\); Linet al\.\([2024](https://arxiv.org/html/2607.27393#bib.bib48)\)\. More recent studies further show that external knowledge, cross\-modal alignment, and cultural grounding are important for reliable meme understandingSharmaet al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib27)\); Renet al\.\([2026](https://arxiv.org/html/2607.27393#bib.bib59)\); Kmainasiet al\.\([2025a](https://arxiv.org/html/2607.27393#bib.bib31)\)\. However, most research in this area remains English\-centric, and recent evidence suggests that VLMs may inherit culturally narrow assumptions when applied to non\-English memesWanget al\.\([2026](https://arxiv.org/html/2607.27393#bib.bib64)\)\.

Recent efforts have begun to address under\-resourced languages, including Bengali, Tamil, MalayalamShahrooret al\.\([2026](https://arxiv.org/html/2607.27393#bib.bib32)\)\. However, these resources remain fragmented in scale, language coverage, and annotation depthDas and Mukherjee \([2023](https://arxiv.org/html/2607.27393#bib.bib71)\); Ponnusamyet al\.\([2024](https://arxiv.org/html/2607.27393#bib.bib56)\); Luet al\.\([2024](https://arxiv.org/html/2607.27393#bib.bib51)\)\. In addition, most resources use binary or coarse labels, which obscure*who*is targeted,*what*type of hate is expressed, and*how*hate is rhetorically framed\.

![Refer to caption](https://arxiv.org/html/2607.27393v1/pipelinev2.png)Figure 2:Overview of theAHA\-Memesconstruction pipeline\.
##### Arabic Resources\.

Arabic multimodal disinformation or harmful\-content research has been driven mainly by text\-based datasets and shared tasks\. OSACT advanced offensive\-language and hate\-speech detection in Arabic social mediaMubaraket al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib53)\), while related shared tasks and datasets have addressed propaganda detection and explanation in Arabic and multilingual settingsHasanainet al\.\([2025](https://arxiv.org/html/2607.27393#bib.bib99)\)\. Recent multilingual and multimodal research on Arabic has addressed fact\-checkingAlamet al\.\([2023](https://arxiv.org/html/2607.27393#bib.bib4)\)and the understanding of news and social\-media contentKmainasiet al\.\([2025b](https://arxiv.org/html/2607.27393#bib.bib100)\)\. Arabic multimodal social\-media analysis has focused primarily on propaganda rather than hate\. The ArMeme datasetAlamet al\.\([2024b](https://arxiv.org/html/2607.27393#bib.bib30)\)and ArAIEval shared tasksHasanainet al\.\([2024b](https://arxiv.org/html/2607.27393#bib.bib29)\)introduced Arabic meme propaganda\-technique detection, followed by work on coarse hate labels, explanation\-enhanced modelling, and explainable Arabic\-English propaganda detectionAlamet al\.\([2024a](https://arxiv.org/html/2607.27393#bib.bib35)\); Kmainasiet al\.\([2025a](https://arxiv.org/html/2607.27393#bib.bib31)\); Alamet al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib34)\)\. These resources are valuable for studying persuasion and propaganda, but they do not provide systematic fine\-grained annotations for hateful memes, including hate, and fine\-grained hate type, as presented in Table[1](https://arxiv.org/html/2607.27393#S2.T1)\. We address this gap by developingAHA\-Memes, a dedicated fine\-grained benchmark for Arabic hateful\-meme detection\.

DatasetLang\.SizeMulti\-labelTask granularitySourceArMeme\(Alamet al\.,[2024b](https://arxiv.org/html/2607.27393#bib.bib30)\)AR5,725NoPropaganda statusSocial mediaArAIEval\(Hasanainet al\.,[2024b](https://arxiv.org/html/2607.27393#bib.bib29)\)†AR3,062NoBinary propagandaArMemePropaganda\-to\-Hate\(Alamet al\.,[2024a](https://arxiv.org/html/2607.27393#bib.bib35)\)†AR∼\\sim3\.1KYesHate \+ hate typesArAIEvalMAHED Task 3\(Zaghouaniet al\.,[2025](https://arxiv.org/html/2607.27393#bib.bib3)\)†AR3,562NoBinary hateSocial mediaMemeXplain\(Kmainasiet al\.,[2025a](https://arxiv.org/html/2607.27393#bib.bib31)\)†AR/EN5,725 AR \+ 10K ENNoLabel \+ explanationsArMeme \+ Hateful MemesAraHarMeme\(El\-Amranyet al\.,[2026](https://arxiv.org/html/2607.27393#bib.bib1)\)AR5,313NoThree\-way harmfulnessSocial mediaArPoMeme\(Zaghouaniet al\.,[2026](https://arxiv.org/html/2607.27393#bib.bib2)\)AR∼\\sim7\.3KYesIdeology and polarizationPolitical memesAHA\-MEMES \(ours\)AR5K \+ 66K‡YesHate \+ 8 hate / 3 non\-hate typesfour platforms

Table 1:Comparison of Arabic multimodal meme datasets\.†The resource contains data derived from or overlapping with an earlier Arabic meme dataset\.‡AHA\-Memes contains 5K human\-annotated gold memes and approximately 66K auxiliary silver\-labeled memes\.
##### In\-Context Learning for Multimodality\.

In\-context learning offers a training\-free alternative to fine\-tuning, however, its effectiveness depends strongly on which demonstrations are included in the promptBrownet al\.\([2020](https://arxiv.org/html/2607.27393#bib.bib77)\); Abdelaliet al\.\([2024](https://arxiv.org/html/2607.27393#bib.bib33)\)\. Similarity\-based selection is a common strategy for selecting useful examples in text classificationLiuet al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib73)\)\. In multimodal learning, most ICL work has focused on captioning and visual question answering, often using retrieval from a single modality\. It remains underexplored for multimodal*classification*, especially in culturally grounded and low\-resource settings\. We address this gap by selecting demonstrations through reciprocal\-rank fusion of text and image similaritiesCormacket al\.\([2009](https://arxiv.org/html/2607.27393#bib.bib74)\)\.

## 3Dataset

### 3\.1Data Curation

We sourced memes from Facebook, Instagram, Pinterest, and Twitter \(X\)\. For Facebook, Instagram, and Pinterest, we manually selected public groups and pages focused on public figures, politics, and social commentary, contexts in which hateful memes are likely to circulate\. We then collected images from these public sources using a semi\-manual procedure following prior workAlamet al\.\([2024b](https://arxiv.org/html/2607.27393#bib.bib30)\)\. For Twitter \(X\), we collected memes through keyword\-based search\. In Figure[2](https://arxiv.org/html/2607.27393#S2.F2), we provide the data curation pipeline\.

### 3\.2Filtering and OCR Text

Duplicate Removal\.Social\-media collections are highly redundant as users often repost identical or slightly modified content\. We therefore removed exact and near\-duplicate images after initial data curation\. For near\-duplicate filtering, we extracted image embeddings using a model fine\-tuned on social\-media imagesAlamet al\.\([2020](https://arxiv.org/html/2607.27393#bib.bib66)\)following the approach discussed inAlamet al\.\([2024b](https://arxiv.org/html/2607.27393#bib.bib30)\)\. We then computed pairwise Euclidean distances between image embeddings across the full dataset\. Image pairs with a distance of3\.63\.6or lower were treated as near\-duplicates, and only one image from each duplicate cluster was retained\.

OCR Text\.We extracted embedded text using EasyOCR,444[https://github\.com/JaidedAI/EasyOCR](https://github.com/JaidedAI/EasyOCR)which we selected for its broad script coverage and scene\-text detection capabilitiesLiaoet al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib67)\)\. Because our study focuses on multimodal image\-text analysis, we discarded memes with no detectable embedded text\. As a result, every retained meme contains both visual and textual modalities\.

### 3\.3Image Selection for Annotation\.

Harmful, hateful, and misleading content represents a relatively small proportion of online content, however, its societal impact can be substantial, as reported in prior studiesSharmaet al\.\([2022](https://arxiv.org/html/2607.27393#bib.bib27)\)\. To increase the proportion of relevant examples while preserving coverage of non\-hateful memes, we used a weak pre\-selection strategy before manual annotation\. Specifically, we applied Gemma\-3\-12B to assign preliminary binary hateful labels to 71K memes\. We then selected 5K memes for manual annotation from this pre\-selected pool\. This sampling strategy produced a balanced and representative annotation set, consisting of 1,850 hateful memes and 3,150 non\-hateful memes\.

### 3\.4Annotation

##### Manual Annotation Tasks\.

We manually annotate each meme for binary hatefulness and, fine\-grained hierarchical labels\. The annotation scheme is organized as follows\.

- •is\_hatefulis a binary judgment\. Following established definitionsKielaet al\.\([2020](https://arxiv.org/html/2607.27393#bib.bib45)\), a meme is labeled as*Hateful*if it directly or indirectly attacks people on the basis of a protected characteristic, including ethnicity, race, nationality, immigration status, religion, caste, sex, gender identity, sexual orientation, disability, or disease\. We make two boundary cases explicit\.First, attacks against groups that themselves promote or perpetrate hate, such as terrorist organizations, are not treated as hate speech\.Second, content that is offensive, rude, or insulting but does not target a protected category is labeled*Not\-Hateful*\. This separates hatefulness from generic offensiveness\.
- •For memes labeled*Hateful*,hateful\_typecaptures the attack strategy as a multi\-label annotation over*Mocking*,*Incitement*,*Dehumanization*,*Slurs*,*Contempt*,*Inferiority*, and*Exclusion*\. A meme may receive multiple hate\-type labels \(e\.g\.,*Slurs*and*Dehumanization*\)\.
- •For memes labeled*Not\-Hateful*,non\_hateful\_typeassigns one or more labels from*Humor*,*Sarcasm*, and*Other*\.

This hierarchical structure supports binary hate detection at the top level, enables fine\-grained categorization for hateful content, and ensures that fine\-grained labels are consistent with the binary decision by construction\. We provide category definitions and examples in Appendix[B](https://arxiv.org/html/2607.27393#A2)\.

##### Annotation Team\.

Annotation was conducted by a third\-party company\. Three trained native Arabic\-speaking annotators annotated the dataset using bilingual guidelines \(See Appendix[B](https://arxiv.org/html/2607.27393#A2)\) in English and Arabic, which we developed to capture dialectal and culturally specific cuesHasanainet al\.\([2024a](https://arxiv.org/html/2607.27393#bib.bib69)\)\. Before annotation, annotators were informed that the dataset may contain offensive, hateful, or otherwise harmful content\. In addition, they were informed about the use of the data\. We also conducted multiple rounds of training and guideline refinement to improve annotation consistency and quality\. In total,AHA\-Memescontains 5K manually annotated memes, with an annotation cost of∼$​4​K\\sim\\mathdollar 4K\.

##### Annotator Agreement

We measured agreement with Cohen’sκ\\kappaagreement on the annotated dataset\. For each multi\-label task, every subtype was treated as a separate binary label\. Cohen’sκ\\kappawas computed independently for each subtype and each of the three annotator pairs, then macro\-averaged across subtypes and annotator pairs\. Annotation agreement are 0\.91, 0\.75, and 0\.67 foris\_hateful,hateful\_type, andnon\_hateful\_type, respectively, indicating substantial to near\-perfect agreement for all annotation tasksLandis and Koch \([1977](https://arxiv.org/html/2607.27393#bib.bib70)\)\.

### 3\.5Statistics and Data Splits

##### Distribution\.

AHA\-Memesis moderately imbalanced, with 1,850 memes \(37\.0%\) labeled as*Hateful*\. The fine\-grained label distribution is long\-tailed \(Table[2](https://arxiv.org/html/2607.27393#S3.T2)\)\. On the non\-hateful side,*Sarcasm*\(1,393\) and*Humor*\(1,331\) are the most frequent labels\. Among hateful memes,*Mocking*\(1,007\) is the most common attack strategy, whereas*Exclusion*\(17\) is the rarest\. Since fine\-grained labels are multi\-label, the total label count exceeds the number of memes\. The proportion of multi\-label memes also increases from 9\.9% in the train split to 21\.4% in the test split, reflecting the richer co\-occurring strategies captured in the triple\-annotated gold subset\. Per\-split label cardinality is reported in Appendix[C\.3](https://arxiv.org/html/2607.27393#A3.SS3)\.

##### Splits\.

We partitionAHA\-Memesinto train, development, and test splits containing 3,500 \(70%\), 500 \(10%\), and 1,000 \(20%\) memes, respectively\. The splits are stratified by the binary hate label, and we ensure that no meme appears in more than one split\. The label distribution across the train, development, and test sets is reported in Table[2](https://arxiv.org/html/2607.27393#S3.T2)\.

HateFine\-grainedTrainDevTestTotalHatefulMocking706902111,007HatefulIncitement3205185456HatefulDehumanization2474258347HatefulSlurs2524247341HatefulContempt1071850175HatefulInferiority571432103HatefulExclusion104317HatefulOther \(H\)182727Not HatefulOther \(NH\)3805096526Not HatefulSarcasm9341263331,393Not HatefulHumor8631363321,331

Table 2:Fine\-grained sub\-type distribution across splits\. Counts sum to more than the number of memes because the label is multi\-label\.*Humor*,*Sarcasm*, and*Other*are non\-hateful subtypes; the rest are hateful attack types\. H: Hateful; NH: Not\-Hateful\.

### 3\.6Auxiliary Silver Dataset

In addition to the 5K human\-verified memes that constitute the coreAHA\-Memes, we release an auxiliary dataset of∼\\sim66K Arabic memes with LLM\-generated annotations from Gemini 3\.1 Pro\.555[https://deepmind\.google/models/model\-cards/gemini\-3\-1\-pro/](https://deepmind.google/models/model-cards/gemini-3-1-pro/)We refer to this resource as a*silver*datasetWanget al\.\([2024](https://arxiv.org/html/2607.27393#bib.bib11)\); Duet al\.\([2023](https://arxiv.org/html/2607.27393#bib.bib10)\)as its labels are not human\-verified\. The dataset is intended to support weakly supervised learning, semi\-supervised learning, active learning, retrieval\-based analysis, and future annotation expansion\.

The motivation for releasing this auxiliary set is twofold\.First, even without labels, a large in\-domain collection of Arabic memes is valuable for the community\. It captures meme templates, OCR noise, dialectal and code\-mixed text, culturally specific references, visual stereotypes, and rapidly evolving political and social topics\. Such data can support meme\-understanding tasks that require joint reasoning over image content, embedded text, social context, and target entitiesPramanicket al\.\([2021](https://arxiv.org/html/2607.27393#bib.bib15)\)\.Second, the silver labels provide a scalable starting point for methods that explicitly model noisy supervision\. Weak\-supervision frameworks have shown that noisy heuristic labels can be useful when denoised or combined with small amounts of gold data\(Ratneret al\.,[2016](https://arxiv.org/html/2607.27393#bib.bib18),[2017](https://arxiv.org/html/2607.27393#bib.bib19)\)\. Similarly, semi\-supervised and self\-training methods use pseudo\-labels on large unlabeled collections to improve model learning, especially when combined with confidence filtering and consistency\-based training\(Sohnet al\.,[2020](https://arxiv.org/html/2607.27393#bib.bib21)\)\. Recent work on LLM\-assisted annotation further suggests that LLMs can reduce annotation cost and help bootstrap labeled resources, while also emphasizing the importance of distinguishing model\-generated annotations from human\-verified labels\(Tanet al\.,[2024](https://arxiv.org/html/2607.27393#bib.bib24); Heet al\.,[2024](https://arxiv.org/html/2607.27393#bib.bib23)\)\.

We therefore position the∼\\sim66K set as an auxiliary community resource rather than as a core contribution\. All primary results in this paper are reported on the human\-verified 5K set\. Silver annotations are produced with Gemini 3\.1 Pro using two prompts: for the unlabelled∼\\sim66K memes the model predicts the label and subtype together with rich metadata \(topic, dialect, sentiment, target categories, propaganda techniques, etc\.\), whereas for the 5K gold memes it generates the same metadata*conditioned*on the human label, so it never relabels them\. The silver labels indicate that hateful content is relatively rare while the collection remains dialectally and topically diverse\. Full distributions and the prompts are provided in Appendix[E](https://arxiv.org/html/2607.27393#A5.SS0.SSS0.Px2)\.

## 4Experiments

We evaluate models on the two tasks defined by our annotation scheme\. The first is binary hateful\-meme detection, where each meme is classified as*Hateful*or*Not\-Hateful*\. The second is fine\-grained multi\-label hate\-type classification\.

### 4\.1Evaluation Measures

For the binary task, we use macro\-F1 as the primary evaluation measure and additionally report accuracy and recall\. For the fine\-grained hate\-type task, we use macro\-F1 averaged over all labels as the primary measure, with micro\-F1 reported as a complementary measure to account for overall multi\-label performance\.

### 4\.2Models

We conduct experiments with five model groups covering unimodal encoders, multimodal fusion models, and vision\-language models \(VLMs\)\. The text\-only baselines use OCR text with AraBERTv2\(Antounet al\.,[2020](https://arxiv.org/html/2607.27393#bib.bib82)\), MARBERTv2\(Abdul\-Mageedet al\.,[2021](https://arxiv.org/html/2607.27393#bib.bib81)\), CAMeLBERT\-mix\(Inoueet al\.,[2021](https://arxiv.org/html/2607.27393#bib.bib83)\), XLM\-R\-base\(Conneauet al\.,[2020](https://arxiv.org/html/2607.27393#bib.bib84)\), and mBERT\(Devlinet al\.,[2019](https://arxiv.org/html/2607.27393#bib.bib85)\)\. The image\-only baselines use the meme image with ViT\-B/16\(Dosovitskiyet al\.,[2020](https://arxiv.org/html/2607.27393#bib.bib86)\), BEiT\-B/16\(Baoet al\.,[2022](https://arxiv.org/html/2607.27393#bib.bib87)\), Swin\-B\(Liuet al\.,[2021](https://arxiv.org/html/2607.27393#bib.bib88)\), ConvNeXt\-V2\-tiny\(Wooet al\.,[2023](https://arxiv.org/html/2607.27393#bib.bib89)\), and DINOv2\-B\(Oquabet al\.,[2024](https://arxiv.org/html/2607.27393#bib.bib90)\)\. Each unimodal encoder is followed by a linear classification head\. The late\-fusion models concatenate pooled representations from the two strongest text encoders and the two strongest vision encoders\. The concatenated vector is passed to a two\-layer MLP with GELU activation, dropout of 0\.1, and hidden size 512\.

We further benchmark both open\- and closed\-weight vision\-language models \(VLMs\), includingQwen3\-VL\-8B\-Instruct\(Qwen3\-VL\-8B\)Xuet al\.\([2025](https://arxiv.org/html/2607.27393#bib.bib8)\), Fanar\-2\-Oryx\-IVUFANAR TEAMet al\.\([2026](https://arxiv.org/html/2607.27393#bib.bib6)\), Phi\-3\.5\-VisionAbdinet al\.\([2024](https://arxiv.org/html/2607.27393#bib.bib5)\), SmolVLM\-Instruct, Qwen3\-VL\-8B\-Thinking, InternVL3\.5\-8B, Gemini\-2\.5\-Pro, Gemini\-2\.5\-FlashComaniciet al\.\([2025](https://arxiv.org/html/2607.27393#bib.bib9)\), and GPT\-5Singhet al\.\([2026](https://arxiv.org/html/2607.27393#bib.bib7)\)\. Unless otherwise stated, all VLMs are evaluated in the zero\-shot setting\. In addition, we fine\-tune Qwen3\-VL\-8B and evaluate its performance under both fine\-tuning and few\-shot in\-context learning settings\. We conduct few\-shot experiments only with Qwen3\-VL\-8B to examine the effect of few\-shot examples under a controlled setting, rather than exhaustively compare prompting across all VLMs\.

### 4\.3Training and Model Selection

All fine\-tuned encoders and fusion models are trained with AdamW using a weight decay of 0\.01, a linear learning\-rate schedule, and 6% warmup\. For binary hate detection, we use a softmax classification head with cross\-entropy loss\. For fine\-grained hate\-type prediction, we use a sigmoid output layer with binary cross\-entropy loss to support multi\-label classification\. For each model family, we tune the learning rate, number of epochs, and batch size, and select the best checkpoint according to development macro\-F1\. The implementation details are provided in Appendix[C\.2](https://arxiv.org/html/2607.27393#A3.SS2)\.

Open\-weight VLMs \(Qwen3\-VL\-8B and Qwen3\-VL\-2B\) are fine\-tuned with low\-rank adaptation using LoRA with rank 16 andα=32\\alpha\{=\}32, applied to all linear layers while keeping the vision tower frozen\. These models are trained for 3 epochs with a learning rate of1×10−41\{\\times\}10^\{\-4\}, a cosine schedule, andbfloat16precision\. For all supervised settings, models are trained ontrain, selected ondevby macro\-F1, and evaluated once on the held\-outtestsplit using the single best configuration\. For fine\-grained hate\-type prediction, the sigmoid decision threshold is tuned on thedevset\. Zero\-shot models are evaluated directly ontest\.

## 5Results and Discussion

### 5\.1Results Across Model Families

In Table[3](https://arxiv.org/html/2607.27393#S5.T3), we report performance across all experimental settings including random and majority\-class baselines\.

BinaryFine\-grainedFamilyModelmacro\-F1AccRecHatemacro\-F1micro\-F1BaselineMajority0\.3990\.6630\.0000\.0500\.295Random0\.5000\.5530\.3370\.1250\.233Text \(FT\)MARBERTv20\.7090\.7430\.5960\.2630\.390CAMeLBERT\-mix0\.6890\.7260\.5670\.2520\.375AraBERTv20\.6780\.7060\.6110\.2950\.411XLM\-R\-base0\.6500\.7160\.4180\.2590\.415mBERT0\.6370\.6670\.5640\.2450\.364Image \(FT\)ConvNeXt\-V2\-tiny0\.6570\.7000\.5130\.2460\.418Swin\-base0\.6570\.6860\.5850\.2580\.365BEiT\-base0\.6510\.7080\.4510\.2550\.385ViT\-base0\.6450\.7080\.4270\.2680\.424DINOv2\-base0\.5890\.6810\.3090\.2330\.349Fusion \(FT\)MARBERTv2\+\+ViT0\.7300\.7570\.6500\.2740\.428MARBERTv2\+\+BEiT0\.7240\.7500\.6560\.3180\.444AraBERTv2\+\+ViT0\.7060\.7410\.5910\.2790\.416AraBERTv2\+\+BEiT0\.6980\.7390\.5490\.2980\.426Open VLM\(zero\-shot\)Qwen3\-VL\-8B\-Instruct0\.6430\.7430\.3180\.1760\.352Fanar\-2\-Oryx\-IVU0\.5350\.7000\.1540\.1150\.288Phi\-3\.5\-vision0\.4860\.6820\.0950\.0440\.159SmolVLM\-Instruct0\.4730\.5590\.2290\.1520\.341Qwen3\-VL\-8B\-Thinking0\.4690\.6760\.0770\.1910\.382InternVL3\.5\-8B0\.4580\.6760\.0620\.1640\.369Open VLM\(fine\-tuned\)Qwen3\-VL\-8B0\.7680\.7970\.6560\.3340\.475Qwen3\-VL\-2B0\.7390\.7820\.5580\.2840\.437Open VLM\(few\-shot\)Qwen3\-VL\-8B0\.7080\.7710\.4540\.2410\.405Closed VLM\(zero\-shot\)Gemini\-2\.5\-pro0\.7110\.7740\.4570\.3400\.465GPT\-50\.6280\.7410\.2820\.3010\.464Gemini\-3\.5\-flash0\.4990\.6940\.1040\.2710\.486Table 3:Results on theAHA\-Memestest split\. FT: Fine\-tuned Bold marks the best value per column\. RecHateis recall of the*Hateful*class, and fine\-grained hate\-type macro\-F1 is averaged over all ten categories\.Fine\-tuned VLMs are strongest for binary task\.Qwen3\-VL\-8B with LoRA achieves the best binary macro\-F1 of 0\.768\. It outperforms the best late\-fusion model, the best zero\-shot closed VLM, and the strongest text encoder\. Qwen3\-VL\-2B also performs well, reaching 0\.739 macro\-F1, which is close to the best fusion models\. This suggests that supervised multimodal model is the most effective approach when labeled data are available\.

Zero\-shot VLMs show low hateful\-class recall\.Qwen3\-VL\-8B\-Instruct is the strongest zero\-shot open VLM, reaching 0\.643 binary macro\-F1, followed by Fanar\-2\-Oryx\-IVU\. The remaining open VLMs perform close to the majority baseline on binary and show very low*Hateful*recall\. The best zero\-shot open VLM falls behind Gemini\-2\.5\-pro by about 7 macro\-F1 and the fine\-tuned Qwen3\-VL\-8B by about 13 points\. Fine\-tuning substantially changes this pattern, improving Qwen3\-VL\-8B from 0\.643 to 0\.768 on binary and from 0\.176 to 0\.334 on fine\-grained hate\-type prediction\.

Fine\-grained hate\-type classification remains difficult\.The best fine\-grained hate\-type macro\-F1 is 0\.340, achieved by Gemini\-2\.5\-pro, with fine\-tuned Qwen3\-VL\-8B close behind at 0\.334\. All model groups remain below 0\.35 macro\-F1\. The gap between macro\-F1 and micro\-F1 suggests that models perform better on frequent labels such as*Humor*and*Sarcasm*than on rare hate types\. Fine\-tuning improves the performance, however, it does not fully mitigate the effect of label imbalance\.

Fusion helps, but text provides the strongest unimodal signal\.Late fusion improves over the best text\-only encoder on binary and increases*Hateful*recall, with MARBERTv2\+\+BEiT matching the highest recall of 0\.656\. Image\-only encoders are competitive, however, they are less reliable as standalone models\. This pattern suggests that embedded Arabic text carries much of the task signal, while visual content provides complementary context that is most useful when combined with text\.

Arabic\-specific encoders provide strong text baselines\.Among text\-only models, MARBERTv2 achieves the best binary macro\-F1, while AraBERTv2 achieves the best fine\-grained hate\-type macro\-F1\. Both Arabic\-specific encoders outperform mBERT and XLM\-R\-base on binary, highlighting the value of Arabic\-centric pre\-training for dialect\-rich meme text\.

### 5\.2Effect of Prompting

We comparedirect promptingwithrationale\-style promptingin English and Arabic for closed VLMs\. Full results are reported in Appendix Table[8](https://arxiv.org/html/2607.27393#A3.T8)\. Two trends are consistent\.First, model choice has a larger effect than prompt format\. Gemini\-2\.5\-pro performs best across all prompting settings, followed by GPT\-5 and Gemini\-3\.5\-flash\.Second, rationale\-style prompting yields limited and model\-dependent gains\. It has little effect on binary performance, improves fine\-grained hate\-type prediction for the larger models, and hurts Gemini\-3\.5\-flash when rationales are requested in Arabic\. This suggests thatArabic rationale generation can introduce additional errorsfor lower\-capacity models\.

### 5\.3Retrieval\-Augmented Few\-Shot ICL

![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/fewshot/fig_kscaling.png)Figure 3:Few\-shot learning for Qwen3\-VL\-8B\. Binary performance improves with more samples, however, it falls below zero\-shot atK=1K\{=\}1\. Fine\-grained hate\-type performance improves modestly and saturates early\. RRF is the strongest retrieval strategy\.We evaluate retrieval\-augmented few\-shot ICL\. For each test meme, we retrieveKKlabeled examples from thetrainsplit and place them before the query as image\-text demonstrations\. We compare random demonstrations with text\-based retrieval, image\-based retrieval, and reciprocal rank fusion \(RRF\) over jina\-clip\-v2 embeddings\(Koukounaset al\.,[2024](https://arxiv.org/html/2607.27393#bib.bib76); Cormacket al\.,[2009](https://arxiv.org/html/2607.27393#bib.bib74)\)\. We evaluateK∈1,3,5K\\in\{1,3,5\}using Qwen3\-VL\-8B and Qwen3\-VL\-2B\. Details results are reported in Appendix[D](https://arxiv.org/html/2607.27393#A4)\.

##### Few\-shot ICL gives moderate gains over zero\-shot prompting, but remains below fine\-tuning\.

As shown in Figure[3](https://arxiv.org/html/2607.27393#S5.F3), RRF is the strongest retrieval strategy\. The best setting is Qwen3\-VL\-8B with RRF andK=5K\{=\}5, which improves binary macro\-F1 from 0\.643 to 0\.708 and fine\-grained hate\-type macro\-F1 from 0\.176 to 0\.241 over the zero\-shot Qwen3\-VL\-8B baseline\. However, this remains below full fine\-tuning, which reaches 0\.768 binary macro\-F1 and 0\.334 fine\-grained hate\-type macro\-F1\. These results show that retrieval selects more informative demonstrations but still falls short of supervised adaptation\.

##### Retrieval mainly improves hateful\-class recall\.

RRF withK=5K\{=\}5raises*Hateful*recall from 0\.318 to 0\.454, reducing the zero\-shot model’s bias toward non\-hateful predictions\. Gains for fine\-grained hate\-type prediction are smaller and are driven mostly by frequent labels such as*Humor*and*Mocking*, while rare hateful types remain difficult\. A single demonstration is unstable and hurts binary performance, whereasK=3K\{=\}3andK=5K\{=\}5provide more reliable context\.

### 5\.4Error Analysis

##### Low recall is the main zero\-shot failure\.

Zero\-shot VLMs show low*Hateful*recall on binary\. They under\-predict the*Hateful*class and therefore miss many hateful memes, making this the most important safety failure for moderation\. Fine\-tuning mitigates this problem, Qwen3\-VL\-8B and the best fusion model recover roughly two\-thirds of hateful memes\. This suggests that the low\-recall pattern is mainly a limitation of the zero\-shot setting rather than a limitation of the model family\.

Fine\-grained hate\-type predictions are biased toward frequent labels\.The zero\-shot closed VLMs predict fine\-grained hate\-type distributions that differ considerably from the gold label distribution, as shown in Appendix Table[9](https://arxiv.org/html/2607.27393#A3.T9)\. Gemini\-3\.5\-flash over\-predicts frequent labels such as*Humor*and*Sarcasm*, which improves micro\-F1, however, weakens macro\-F1\. Gemini\-2\.5\-pro over\-predicts hateful types such as*Mocking*,*Contempt*, and*Slurs*, increasing recall at the expense of precision\. GPT\-5 is more cautious\.*Exclusion*has only three test instances and is almost never predicted correctly\.

Longer meme text increases difficulty\.Binary errors increase with the length of the extracted Arabic text for all models, as shown in Appendix Table[10](https://arxiv.org/html/2607.27393#A3.T10)\. The trend is steepest for Gemini\-3\.5\-flash\. Longer text often require more context, pragmatic inference, and target identification, whereas memes with short text are more likely to contain direct lexical cues\. Larger models are more robust, however, the trend remains visible across systems\.

Implicit hate is harder to classify\.Qualitative inspection of Gemini\-2\.5\-pro with Arabic rationale prompting shows that false negatives are not caused by OCR or image\-understanding failures\. The model identifies the visual template and reads dialectal Arabic correctly but treats implicit sectarian, religious, or misogynistic attacks as humor or satire and predicts*Not\-Hateful*\. False positives are offensive dialectal jokes that annotators judged not to be hateful\. These cases show that the hardest boundary is the normative judgment of implicit hate rather than multimodal prediction\. This motivates the explicit taxonomy and annotation guidelines described in §[3\.4](https://arxiv.org/html/2607.27393#S3.SS4.SSS0.Px3)\.

## 6Conclusion and Future Work

We presentedAHA\-Memes, a fine\-grained Arabic hateful\-meme benchmark comprising 5K human\-annotated memes with a hierarchical annotation scheme covering hatefulness, attack subtype, and target category\. We also release a separate auxiliary corpus of∼\\sim66K silver\-labeled memes with model\-generated labels and metadata for weak supervision rather than evaluation\. Experiments show that fine\-tuning an open\-weight VLM gives the best binary performance, while late fusion and Arabic text encoders remain strong baselines\. Text provides the strongest unimodal signal, while visual context offers complementary information\. Fine\-grained classification remains challenging because models struggle to recover rare and overlapping hate categories\. Zero\-shot VLMs make conservative predictions and often miss hateful memes, while few\-shot prompting provides only modest gains\. Future work will focus on rare and co\-occurring hate categories, target\-aware modeling, clearer separation between offensive humor and protected\-group hate, and broader coverage of dialects and emerging meme formats\.

## Limitations

AHA\-Memesis designed as a focused benchmark for Arabic hateful\-meme understanding rather than a comprehensive representation of Arabic meme culture\. We collected data four public platforms, Facebook, Instagram, Pinterest, and Twitter/X, which gives broad public coverage, however, our collection may not capture all dialects, regions, platform communities, or private and ephemeral content\. Some annotation boundaries are inherently nuanced, especially between offensive humor, satire, political criticism, and protected\-group hate\. We address this detailed annotation guideline, and training\. The auxiliary silver corpus broadens the resource and supports weakly supervised research, but its labels are model\-generated and not human\-verified\.

## Ethics and Broader Impact

AHA\-Memessupports research on Arabic multimodal hate detection and safer content moderation\. It is developed from publicly available memes and includes hateful, offensive, and sensitive material solely for research and evaluation\. We do not collect or release personal information, which helps minimize privacy risks\. As models can miss harmful content or over\-flag legitimate content, they should be used with human oversight, clear documentation, and ongoing monitoring\. Annotation was performed by trained, fairly compensated native\-speaker annotators using guidelines that distinguish protected\-group hate from generic offensiveness and legitimate criticism\. To further reduce privacy and misuse risks, we release OCR text, labels, metadata, and image references\.

## Acknowledgments

The work was supported by NPRP grant 14C\-0916\-210015 from the Qatar National Research Fund, part of the Qatar Research Development and Innovation Council \(QRDI\)\. The findings reported herein are solely the responsibility of the authors\.

## References

- A\. Abdelali, H\. Mubarak, S\. A\. Chowdhury, M\. Hasanain, B\. Mousi, S\. Boughorbel, S\. Abdaljalil, Y\. El Kheir, D\. Izham, F\. Dalvi, M\. Hawasly, N\. Nazar, Y\. Elshahawy, A\. Ali, N\. Durrani, N\. Milic\-Frayling, and F\. Alam \(2024\)LAraBench: benchmarking Arabic AI with large language models\.InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),Y\. Graham and M\. Purver \(Eds\.\),St\. Julian’s, Malta,pp\. 487–520\.External Links:[Link](https://aclanthology.org/2024.eacl-long.30/),[Document](https://dx.doi.org/10.18653/v1/2024.eacl-long.30)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px3.p1.1)\.
- M\. Abdin, J\. Aneja, H\. Awadalla, A\. Awadallah, A\. A\. Awan, N\. Bach, A\. Bahree, A\. Bakhtiari, J\. Bao, H\. Behl,et al\.\(2024\)Phi\-3 technical report: a highly capable language model locally on your phone\.ArXiv preprintarXiv:2404\.14219\.External Links:[Link](https://arxiv.org/abs/2404.14219)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p2.1)\.
- ARBERT & MARBERT: deep bidirectional transformers for Arabic\.Online,pp\. 7088–7105\.External Links:[Link](https://aclanthology.org/2021.acl-long.551),[Document](https://dx.doi.org/10.18653/v1/2021.acl-long.551)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- F\. Alam, A\. Barrón\-Cedeño, G\. S\. Cheema, G\. K\. Shahi, S\. Hakimov, M\. Hasanain, C\. Li, R\. Míguez, H\. Mubarak, W\. Zaghouani,et al\.\(2023\)Overview of the clef\-2023 checkthat\! lab task 1 on check\-worthiness in multimodal and multigenre content\.\.InCLEF \(Working Notes\),pp\. 219–235\.Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1)\.
- F\. Alam, M\. R\. Biswas, U\. Shah, W\. Zaghouani, and G\. Mikros \(2024a\)Propaganda to hate: a multimodal analysis of arabic memes with multi\-agent llms\.InInternational Conference on Web Information Systems Engineering,pp\. 380–390\.Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p2.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2607.27393#S2.T1.2.2.2.1)\.
- F\. Alam, A\. Hasnat, F\. Ahmad, Md\. A\. Hasan, and M\. Hasanain \(2024b\)ArMeme: propagandistic content in Arabic memes\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 21071–21090\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.1173/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.1173)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2607.27393#S2.T1.7.7.9.1),[§3\.1](https://arxiv.org/html/2607.27393#S3.SS1.p1.1),[§3\.2](https://arxiv.org/html/2607.27393#S3.SS2.p1.1)\.
- F\. Alam, H\. Mubarak, W\. Zaghouani, G\. Da San Martino, and P\. Nakov \(2022\)Overview of the wanlp 2022 shared task on propaganda detection in arabic\.InProceedings of the Seventh Arabic Natural Language Processing Workshop \(WANLP\),pp\. 108–118\.Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1)\.
- F\. Alam, F\. Ofli, M\. Imran, T\. Alam, and U\. Qazi \(2020\)Deep learning benchmarks and datasets for social media image classification for disaster response\.pp\. 151–158\.Cited by:[§3\.2](https://arxiv.org/html/2607.27393#S3.SS2.p1.1)\.
- W\. Antoun, F\. Baly, and H\. Hajj \(2020\)AraBERT: transformer\-based model for arabic language understanding\.pp\. 9\.Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- H\. Bao, L\. Dong, S\. Piao, and F\. Wei \(2022\)BEiT: BERT pre\-training of image transformers\.External Links:[Link](https://openreview.net/forum?id=p-BhZSz59o4)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- T\. B\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell, S\. Agarwal, A\. Herbert\-Voss, G\. Krueger, T\. Henighan, R\. Child, A\. Ramesh, D\. M\. Ziegler, J\. Wu, C\. Winter, C\. Hesse, M\. Chen, E\. Sigler, M\. Litwin, S\. Gray, B\. Chess, J\. Clark, C\. Berner, S\. McCandlish, A\. Radford, I\. Sutskever, and D\. Amodei \(2020\)Language models are few\-shot learners\.pp\. 1877–1901\.Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px3.p1.1)\.
- M\. D\. Bui, K\. von der Wense, and A\. Lauscher \(2025\)Multi3Hate: multimodal, multilingual, and multicultural hate speech detection with vision\-language models\.InProceedings of the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational LinguisticsProceedings of the 2022 Conference on Empirical Methods in Natural Language ProcessingProceedings of the 15th International Workshop on Semantic Evaluation \(SemEval\-2021\)Proceedings of the 16th International Workshop on Semantic Evaluation \(SemEval\-2022\)Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision \(WACV\)Proceedings of the 2024 Conference on Empirical Methods in Natural Language ProcessingAdvances in Neural Information Processing SystemsProceedings of the Second Workshop on NLP for Positive Impact \(NLP4PI\)Findings of the Association for Computational Linguistics: EMNLP 2023Proceedings of the ACM Web Conference 2024Proceedings of the 2022 Conference on Empirical Methods in Natural Language ProcessingProceedings of the 63rd Annual Meeting of the Association for Computational LinguisticsAdvances in Neural Information Processing SystemsProceedings of the 4th Workshop on Open\-Source Arabic Corpora and Processing Tools \(OSACT\)Proceedings of the 5th Workshop on Open\-Source Arabic Corpora and Processing Tools \(OSACT\)Proceedings of the ACM Web Conference 2026Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and EvaluationFindings of the Association for Computational Linguistics: EMNLP 2021Proceedings of the 2025 IEEE International Conference on Big Data \(BigData\)Proceedings of the Fourteenth Workshop on Semantic EvaluationCompanion Proceedings of the ACM Web Conference 2025Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying \(TRAC\)Proceedings of the ACM Web Conference 2026 \(WWW\)Proceedings of the IEEE conference on computer vision and pattern recognition2020 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining \(ASONAM\)Proceedings of the International AAAI Conference on Web and Social MediaProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING\)Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)Proceedings of Deep Learning Inside Out \(DeeLIO\): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning ArchitecturesProceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information RetrievalProceedings of the IEEE/CVF International Conference on Computer Vision \(ICCV\)Advances in Neural Information Processing Systems 33 \(NeurIPS\)Advances in Neural Information Processing Systems 35 \(NeurIPS\)Proceedings of the AAAI Conference on Artificial IntelligenceProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\)LREC 2020 Workshop Language Resources and Evaluation Conference 11–16 May 2020Proceedings of the Sixth Arabic Natural Language Processing WorkshopProceedings of the 58th Annual Meeting of the Association for Computational LinguisticsProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long and Short Papers\)International Conference on Learning RepresentationsWeb Information Systems Engineering – WISE 2024Findings of the Association for Computational Linguistics: EMNLP 2025Findings of the Association for Computational Linguistics: NAACL 2025Companion Proceedings of the ACM Web Conference 2026,L\. Biester, D\. Demszky, Z\. Jin, M\. Sachan, J\. Tetreault, S\. Wilson, L\. Xiao, J\. Zhao, A\. Herbelot, X\. Zhu, A\. Palmer, N\. Schneider, J\. May, E\. Shutova, D\. Jurafsky, J\. Chai, N\. Schluter, J\. Tetreault, J\. Burstein, C\. Doran, T\. Solorio, M\. Barhamgi, H\. Wang, X\. Wang, C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, V\. Peng, L\. Chiruzzo, A\. Ritter, and L\. Wang \(Eds\.\),NAACL ’25EMNLP ’22EMNLP ’24NeurIPS ’20EMNLP ’23WWW ’24EMNLP ’22ACL ’25NeurIPS ’24WWW ’26LREC\-COLING ’24WWW ’25CVPR ’16WWW Companion ’26, Vol\.12020\-December1371536,pp\. 9714 – 9731\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.490)Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p1.1),[§1](https://arxiv.org/html/2607.27393#S1.p2.1)\.
- R\. Cao, R\. K\. Lee, W\. Chong, and J\. Jiang \(2022\)Prompting for multimodal hateful meme classification\.Conference paperpp\. 321 – 332\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.emnlp-main.22)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1)\.
- G\. Comanici, E\. Bieber, M\. Schaekermann, I\. Pasupat, N\. Sachdeva, I\. Dhillon, M\. Blistein, O\. Ram, D\. Zhang, E\. Rosen,et al\.\(2025\)Gemini 2\.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities\.ArXiv preprintarXiv:2507\.06261\.External Links:[Link](https://arxiv.org/abs/2507.06261)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p2.1)\.
- A\. Conneau, K\. Khandelwal, N\. Goyal, V\. Chaudhary, G\. Wenzek, F\. Guzmán, E\. Grave, M\. Ott, L\. Zettlemoyer, and V\. Stoyanov \(2020\)Unsupervised cross\-lingual representation learning at scale\.Online,pp\. 8440–8451\.External Links:[Link](https://aclanthology.org/2020.acl-main.747/),[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.747)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- G\. V\. Cormack, C\. L\. A\. Clarke, and S\. Buettcher \(2009\)Reciprocal rank fusion outperforms condorcet and individual rank learning methods\.pp\. 758–759\.Cited by:[§D\.1](https://arxiv.org/html/2607.27393#A4.SS1.p2.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px3.p1.1),[§5\.3](https://arxiv.org/html/2607.27393#S5.SS3.p1.2)\.
- M\. Das and A\. Mukherjee \(2023\)BanglaAbuseMeme: a dataset for Bengali abusive meme classification\.Singapore,pp\. 15498–15512\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.959/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.959)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p2.1)\.
- J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova \(2019\)BERT: pre\-training of deep bidirectional transformers for language understanding\.Minneapolis, Minnesota,pp\. 4171–4186\.External Links:[Link](https://aclanthology.org/N19-1423/),[Document](https://dx.doi.org/10.18653/v1/N19-1423)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- A\. Dosovitskiy, L\. Beyer, A\. Kolesnikov, D\. Weissenborn, X\. Zhai, T\. Unterthiner, M\. Dehghani, M\. Minderer, G\. Heigold, S\. Gelly, J\. Uszkoreit, and N\. Houlsby \(2020\)An image is worth 16x16 words: transformers for image recognition at scale\.CoRRabs/2010\.11929\.External Links:[Link](https://arxiv.org/abs/2010.11929),2010\.11929Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- W\. Du, L\. Advani, Y\. Gambhir, D\. Perry, P\. Shiralkar, Z\. Xing, and A\. Colak \(2023\)Effective proxy for human labeling: ensemble disagreement scores in large language models for industrial NLP\.InProceedings of the Third Workshop on Natural Language Generation, Evaluation, and Metrics \(GEM\),S\. Gehrmann, A\. Wang, J\. Sedoc, E\. Clark, K\. Dhole, K\. R\. Chandu, E\. Santus, and H\. Sedghamiz \(Eds\.\),Singapore,pp\. 53–61\.External Links:[Link](https://aclanthology.org/2023.gem-1.5/)Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p1.1)\.
- S\. El\-Amrany, A\. El Khaiari, A\. El Bekkali, S\. Lamsiyah, M\. R\. Brust, K\. Lavangnananda, J\. E\. Pecero, and P\. Bouvry \(2026\)AraHarMeme: a multimodal benchmark for harmful meme detection in arabic\.In2026 23rd International Joint Conference on Computer Science and Software Engineering \(JCSSE\),pp\. 564–569\.External Links:[Document](https://dx.doi.org/10.1109/JCSSE68839.2026.11597005)Cited by:[Table 1](https://arxiv.org/html/2607.27393#S2.T1.7.7.10.1)\.
- S\. El\-amrany, S\. Lamsiyah, M\. R\. Brust, and P\. Bouvry \(2025\)GuardHarMem and HarMDetect: a multimodal dataset and benchmark model for fine\-grained harmful meme classification\.Social Network Analysis and Mining15\(1\),pp\. 63\.External Links:[Document](https://dx.doi.org/10.1007/s13278-025-01475-2)Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p2.1)\.
- FANAR TEAM, U\. Abbas, M\. S\. Ahmad, M\. Ahmad, A\. Al\-Homaid, A\. Al\-Nuaimi, E\. Altinisik, E\. Asgari, S\. Chawla, S\. Chowdhury, F\. Dalvi, K\. Darwish, N\. Durrani, M\. Elfeky, A\. Elmagarmid, M\. Eltabakh, A\. Ersoy, M\. Fatehkia, M\. Q\. Hashim, M\. Hawasly, M\. Hefeeda, M\. Husaini, K\. Isufaj, S\. Jung, H\. Lachemat, J\. K\. Lucas, A\. Mohamed, T\. Mohiuddin, B\. Mousi, H\. Mubarak, A\. Musleh, M\. Ouzzani, A\. Sadeghi, H\. T\. Sencar, M\. Shinoy, O\. Sinan, and Y\. Zhang \(2026\)Fanar 2\.0: Arabic generative AI stack\.ArXiv preprintarXiv:2603\.16397\.External Links:[Link](https://arxiv.org/abs/2603.16397)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p2.1)\.
- M\. Hasanain, F\. Ahmad, and F\. Alam \(2024a\)Can GPT\-4 identify propaganda? annotation and detection of propaganda spans in news articles\.Torino, Italy,pp\. 2724–2744\.External Links:[Link](https://aclanthology.org/2024.lrec-main.244/)Cited by:[§3\.4](https://arxiv.org/html/2607.27393#S3.SS4.SSS0.Px2.p1.1)\.
- M\. Hasanain, M\. A\. Hasan, M\. B\. Kmainasi, E\. Sartori, A\. E\. Shahroor, G\. Da San Martino, and F\. Alam \(2025\)PropXplain: can LLMs enable explainable propaganda detection?\.Suzhou, China,pp\. 23855–23863\.External Links:[Link](https://aclanthology.org/2025.findings-emnlp.1296/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.1296),ISBN 979\-8\-89176\-335\-7Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1)\.
- M\. Hasanain, Md\. A\. Hasan, F\. Ahmad, R\. Suwaileh, Md\. R\. Biswas, W\. Zaghouani, and F\. Alam \(2024b\)ArAIEval shared task: propagandistic techniques detection in unimodal and multimodal Arabic content\.InProceedings of the Second Arabic Natural Language Processing Conference,N\. Habash, H\. Bouamor, R\. Eskander, N\. Tomeh, I\. Abu Farha, A\. Abdelali, S\. Touileb, I\. Hamed, Y\. Onaizan, B\. Alhafni, W\. Antoun, S\. Khalifa, H\. Haddad, I\. Zitouni, B\. AlKhamissi, R\. Almatham, and K\. Mrini \(Eds\.\),Bangkok, Thailand,pp\. 456–466\.External Links:[Link](https://aclanthology.org/2024.arabicnlp-1.44/),[Document](https://dx.doi.org/10.18653/v1/2024.arabicnlp-1.44)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2607.27393#S2.T1.1.1.1.1)\.
- X\. He, Z\. Lin, Y\. Gong, A\. Jin, H\. Zhang, C\. Lin, J\. Jiao, S\. M\. Yiu, N\. Duan, and W\. Chen \(2024\)AnnoLLM: making large language models to be better crowdsourced annotators\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 6: Industry Track\),Mexico City, Mexico,pp\. 165–190\.External Links:[Link](https://aclanthology.org/2024.naacl-industry.15/),[Document](https://dx.doi.org/10.18653/v1/2024.naacl-industry.15)Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p2.1)\.
- G\. Inoue, B\. Alhafni, N\. Baimukan, H\. Bouamor, and N\. Habash \(2021\)The interplay of variant, size, and task type in Arabic pre\-trained language models\.Kyiv, Ukraine \(Online\)\.Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- D\. Kiela, H\. Firooz, A\. Mohan, V\. Goswami, A\. Singh, P\. Ringshia, and D\. Testuggine \(2020\)The hateful memes challenge: detecting hate speech in multimodal memes\.Conference paperpp\.\.External Links:[Link](https://proceedings.neurips.cc/paper/2020/hash/1b84c4cee2b8b3d823b30e2d604b1878-Abstract.html)Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p2.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1),[1st item](https://arxiv.org/html/2607.27393#S3.I1.i1.p1.1)\.
- M\. B\. Kmainasi, A\. Hasnat, M\. A\. Hasan, A\. E\. Shahroor, and F\. Alam \(2025a\)MemeIntel: explainable detection of propagandistic and hateful memes\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 30263–30279\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.1539/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.1539),ISBN 979\-8\-89176\-332\-6Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p2.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2607.27393#S2.T1.5.5.5.1)\.
- M\. B\. Kmainasi, A\. E\. Shahroor, M\. Hasanain, S\. R\. Laskar, N\. Hassan, and F\. Alam \(2025b\)LlamaLens: specialized multilingual LLM for analyzing news and social media content\.Albuquerque, New Mexico,pp\. 5642–5664\.External Links:[Link](https://aclanthology.org/2025.findings-naacl.313/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-naacl.313),ISBN 979\-8\-89176\-195\-7Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1)\.
- A\. Koukounas, G\. Mastrapas, S\. Eslami, B\. Wang, M\. K\. Akram, M\. Günther, I\. Mohr, S\. Sturua, N\. Wang, and H\. Xiao \(2024\)jina\-clip\-v2: multilingual multimodal embeddings for text and images\.ArXiv preprintarXiv:2412\.08802\.Cited by:[§D\.1](https://arxiv.org/html/2607.27393#A4.SS1.p1.1),[§5\.3](https://arxiv.org/html/2607.27393#S5.SS3.p1.2)\.
- G\. K\. Kumar and K\. Nandakumar \(2022\)Hate\-CLIPper: multimodal hateful meme classification based on cross\-modal interaction of CLIP features\.Abu Dhabi, United Arab Emirates \(Hybrid\),pp\. 171–183\.External Links:[Link](https://aclanthology.org/2022.nlp4pi-1.20/),[Document](https://dx.doi.org/10.18653/v1/2022.nlp4pi-1.20)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1)\.
- J\. R\. Landis and G\. G\. Koch \(1977\)The measurement of observer agreement for categorical data\.Biometrics33\(1\),pp\. 159–174\.Cited by:[§3\.4](https://arxiv.org/html/2607.27393#S3.SS4.SSS0.Px3.p1.2)\.
- M\. Liao, Z\. Zou, Z\. Wan, C\. Yao, and X\. Bai \(2022\)Real\-time scene text detection with differentiable binarization and adaptive scale fusion\.IEEE transactions on pattern analysis and machine intelligence45\(1\),pp\. 919–931\.Cited by:[§3\.2](https://arxiv.org/html/2607.27393#S3.SS2.p2.1)\.
- H\. Lin, Z\. Luo, W\. Gao, J\. Ma, B\. Wang, and R\. Yang \(2024\)Towards explainable harmful meme detection through multimodal debate between large language models\.Conference paperpp\. 2359 – 2370\.External Links:[Document](https://dx.doi.org/10.1145/3589334.3645381)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Liu, D\. Shen, Y\. Zhang, B\. Dolan, L\. Carin, and W\. Chen \(2022\)What makes good in\-context examples for GPT\-3?\.pp\. 100–114\.Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px3.p1.1)\.
- Z\. Liu, Y\. Lin, Y\. Cao, H\. Hu, Y\. Wei, Z\. Zhang, S\. Lin, and B\. Guo \(2021\)Swin transformer: hierarchical vision transformer using shifted windows\.External Links:2103\.14030,[Link](https://arxiv.org/abs/2103.14030)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- J\. Lu, B\. Xu, X\. Zhang, H\. Wang, H\. Zhu, D\. Zhang, L\. Yang, and H\. Lin \(2024\)Towards comprehensive detection of Chinese harmful memes\.Conference paperpp\.\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2024/hash/17fc467c11997914127c001fdc801bea-Abstract-Datasets_and_Benchmarks_Track.html)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p2.1)\.
- H\. Mubarak, H\. Al\-Khalifa, and A\. Al\-Thubaity \(2022\)Overview of OSACT5 Shared Task on Arabic Offensive Language and Hate Speech Detection\.Marseille, France,pp\. 162–166\.External Links:[Link](https://aclanthology.org/2022.osact-1.20/)Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p2.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px2.p1.1)\.
- M\. Oquab, T\. Darcet, T\. Moutakanni, H\. Vo, M\. Szafraniec, V\. Khalidov, P\. Fernandez, D\. Haziza, F\. Massa, A\. El\-Nouby, M\. Assran, N\. Ballas, W\. Galuba, R\. Howes, P\. Huang, S\. Li, I\. Misra, M\. Rabbat, V\. Sharma, G\. Synnaeve, H\. Xu, H\. Jegou, J\. Mairal, P\. Labatut, A\. Joulin, and P\. Bojanowski \(2024\)DINOv2: learning robust visual features without supervision\.External Links:2304\.07193,[Link](https://arxiv.org/abs/2304.07193)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- R\. Ponnusamy, K\. Pannerselvam, S\. Rajiakodi, P\. K\. Kumaresan, S\. Thavareesan, B\. Sivagnanam, K\.A\. Anshid, S\. S\. Kumar, P\. Buitelaar, and B\. R\. Chakravarthi \(2024\)From laughter to inequality: annotated dataset for misogyny detection in Tamil and Malayalam memes\.Conference paperpp\. 7480 – 7488\.External Links:[Link](https://aclanthology.org/2024.lrec-main.660/)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p2.1)\.
- S\. Pramanick, S\. Sharma, D\. Dimitrov, Md\. S\. Akhtar, P\. Nakov, and T\. Chakraborty \(2021\)MOMENTA: a multimodal framework for detecting harmful memes and their targets\.InFindings of the Association for Computational Linguistics: EMNLP 2021,Punta Cana, Dominican Republic,pp\. 4439–4455\.External Links:[Link](https://aclanthology.org/2021.findings-emnlp.379/),[Document](https://dx.doi.org/10.18653/v1/2021.findings-emnlp.379)Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p2.1)\.
- A\. Ratner, S\. H\. Bach, H\. Ehrenberg, J\. Fries, S\. Wu, and C\. Ré \(2017\)Snorkel: rapid training data creation with weak supervision\.Proceedings of the VLDB Endowment11\(3\),pp\. 269–282\.External Links:[Document](https://dx.doi.org/10.14778/3157794.3157797)Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p2.1)\.
- A\. J\. Ratner, C\. M\. De Sa, S\. Wu, D\. Selsam, and C\. Ré \(2016\)Data programming: creating large training sets, quickly\.InAdvances in Neural Information Processing Systems,Vol\.29,pp\. 3567–3575\.Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p2.1)\.
- C\. Ren, D\. Jeong, M\. Wu, Y\. Huang, Y\. Gao, and Y\. Li \(2026\)A survey of multimodal hate meme detection\.Expert Systems with Applications323,pp\. 132507\.External Links:[Document](https://dx.doi.org/10.1016/j.eswa.2026.132507)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1)\.
- A\. E\. Shahroor, M\. B\. Kmainasi, A\. Hasnat, D\. Dimitrov, G\. Da San Martino, P\. Nakov, and F\. Alam \(2026\)MemeLens: multilingual multitask VLMs for memes\.arXiv: 2601\.12539\.External Links:[Link](https://arxiv.org/abs/2601.12539)Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p1.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p2.1)\.
- S\. Sharma, F\. Alam, Md\. S\. Akhtar, D\. Dimitrov, G\. Da San Martino, H\. Firooz, A\. Halevy, F\. Silvestri, P\. Nakov, and T\. Chakraborty \(2022\)Detecting and understanding harmful memes: a survey\.InProceedings of the Thirty\-First International Joint Conference on Artificial Intelligence, IJCAI\-22,L\. D\. Raedt \(Ed\.\),pp\. 5597–5606\.Note:Survey TrackExternal Links:[Link](https://doi.org/10.24963/ijcai.2022/781)Cited by:[§1](https://arxiv.org/html/2607.27393#S1.p1.1),[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1),[§3\.3](https://arxiv.org/html/2607.27393#S3.SS3.p1.1)\.
- A\. Singh, A\. Fry, A\. Perelman, A\. Tart, A\. Ganesh, A\. El\-Kishky, A\. McLaughlin, A\. Low, A\. Ostrow, A\. Ananthram,et al\.\(2026\)OpenAI GPT\-5 system card\.ArXiv preprintarXiv:2601\.03267\.External Links:[Link](https://arxiv.org/abs/2601.03267)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p2.1)\.
- K\. Sohn, D\. Berthelot, N\. Carlini, Z\. Zhang, H\. Zhang, C\. A\. Raffel, E\. D\. Cubuk, A\. Kurakin, and C\. Li \(2020\)FixMatch: simplifying semi\-supervised learning with consistency and confidence\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 596–608\.Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p2.1)\.
- Z\. Tan, D\. Li, S\. Wang, A\. Beigi, B\. Jiang, A\. Bhattacharjee, M\. Karami, J\. Li, L\. Cheng, and H\. Liu \(2024\)Large language models for data annotation and synthesis: a survey\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 930–957\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.54/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.54)Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p2.1)\.
- J\. Wang, T\. Wang, and Z\. Zeng \(2024\)On the use of silver standard data for zero\-shot classification tasks in information extraction\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),N\. Calzolari, M\. Kan, V\. Hoste, A\. Lenci, S\. Sakti, and N\. Xue \(Eds\.\),Torino, Italia,pp\. 12423–12434\.External Links:[Link](https://aclanthology.org/2024.lrec-main.1088/)Cited by:[§3\.6](https://arxiv.org/html/2607.27393#S3.SS6.p1.1)\.
- M\. Wang, K\. Ren, P\. Jalan, A\. Ashraf, T\. V\. Vu, R\. Seetharaman, S\. Nawaz, and U\. Naseem \(2026\)From native memes to global moderation: cross\-cultural evaluation of vision–language models for hateful meme detection\.Dubai, United Arab Emirates,pp\. 9788–9799\.External Links:[Document](https://dx.doi.org/10.1145/3774904.3793007)Cited by:[§2](https://arxiv.org/html/2607.27393#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Woo, S\. Debnath, R\. Hu, X\. Chen, Z\. Liu, I\. S\. Kweon, and S\. Xie \(2023\)ConvNeXt V2: co\-designing and scaling convnets with masked autoencoders\.CoRRabs/2301\.00808\.External Links:[Link](https://doi.org/10.48550/arXiv.2301.00808),[Document](https://dx.doi.org/10.48550/arXiv.2301.00808),2301\.00808Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p1.1)\.
- J\. Xu, Z\. Guo, H\. Hu, Y\. Chu, X\. Wang, J\. He, Y\. Wang, X\. Shi, T\. He, X\. Zhu,et al\.\(2025\)Qwen3\-Omni technical report\.ArXiv preprintarXiv:2509\.17765\.External Links:[Link](https://arxiv.org/abs/2509.17765)Cited by:[§4\.2](https://arxiv.org/html/2607.27393#S4.SS2.p2.1)\.
- W\. Zaghouani, K\. Attia, Md\. R\. Biswas, and F\. Eryani \(2026\)ArPoMeme: an annotated arabic multimodal dataset for political ideology and polarization\.InProceedings of the Fifteenth Language Resources and Evaluation Conference \(LREC 2026\),Palma, Mallorca, Spain,pp\. 6597–6608\.External Links:[Link](https://lrec.elra.info/lrec2026-main-524),[Document](https://dx.doi.org/10.63317/5ghumwhwumkz),ISBN 978\-2\-493814\-49\-4Cited by:[Table 1](https://arxiv.org/html/2607.27393#S2.T1.6.6.6.2)\.
- W\. Zaghouani, Md\. R\. Biswas, M\. Bessghaier, S\. Ibrahim, G\. Mikros, A\. Hasnat, and F\. Alam \(2025\)MAHED shared task: multimodal detection of hope and hate emotions in Arabic content\.InProceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks,K\. Darwish, A\. Ali, I\. Abu Farha, S\. Touileb, I\. Zitouni, A\. Abdelali, S\. Al\-Ghamdi, S\. Alkhereyf, W\. Zaghouani, S\. Khalifa, B\. AlKhamissi, R\. Almatham, I\. Hamed, Z\. Alyafeai, A\. Alowisheq, G\. Inoue, K\. Mrini, and W\. Alshammari \(Eds\.\),Suzhou, China,pp\. 560–574\.External Links:[Link](https://aclanthology.org/2025.arabicnlp-sharedtasks.75/),[Document](https://dx.doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.75),ISBN 979\-8\-89176\-356\-2Cited by:[Table 1](https://arxiv.org/html/2607.27393#S2.T1.4.4.4.1)\.
- X\. Zhai, B\. Mustafa, A\. Kolesnikov, and L\. Beyer \(2023\)Sigmoid loss for language image pre\-training\.pp\. 11975–11986\.Cited by:[§D\.1](https://arxiv.org/html/2607.27393#A4.SS1.p1.1)\.

## Appendix AData and Resource Release

We will release666[anonymous\.com](https://arxiv.org/html/2607.27393v1/anonymous.com)the supporting code, documentation, and dataset under the CC BY\-NC\-SA 4\.0 license to facilitate adoption, reproducibility, and future extensions\.

## Appendix BAnnotation Guidelines

### B\.1English Annotation Guideline

##### Purpose\.

The purpose of this annotation is:\(i\)to identify whether a meme is*Hateful*or*Not\-Hateful*;\(ii\)for hateful memes, to identify the attack type\(s\) \(multi\-label\) and the targeted protected category\(ies\) \(multi\-label\); and\(iii\)for non\-hateful memes, to assign one or more non\-hateful subtypes \(multi\-label\)\. The annotation design is*hierarchical*and*conditional*: the fine\-grained tasks are completed based on the binary label decision, discussed below\.

##### Definitions\.

We define*hateful*content as a direct or indirect attack on people on the basis of a protected characteristic, including ethnicity, race, nationality, immigration status, religion, caste, sex, gender identity, sexual orientation, and disability or disease\. An*attack*includes violent or dehumanizing speech, such as comparing people to non\-human entities, statements of inferiority, calls for exclusion or segregation, or mockery of a hate crime\. We apply two boundary rules:

- •Attacking groups that themselves promote or carry out hate, such as terrorist organisations, is*not*considered hate speech\.
- •A meme that is insulting or offensive but does not target a protected category is labelled*Not\-Hateful*under this scheme\.

#### B\.1\.1Task 1: Binary Classification \(is\_hateful; Hateful vs\. Not\-Hateful\)

The task is to select exactly one of*Hateful*or*Not\-Hateful*\. The choice determines the subsequent annotation task: if*Hateful*, complete Tasks 2 if*Not\-Hateful*, complete Task 3\.

#### B\.1\.2Task 2: Attack Type \(hateful\_type; Hateful only; multi\-label\)

The task is to select*all*attack types that apply; a meme may exhibit several at once \(e\.g\.,*Slurs*\+\+*Dehumanization*, or*Mocking*\+\+*Inferiority*\)\.

- •Dehumanization:explicitly or implicitly describing a group as subhuman, such as by comparing an ethnic group to animals or using “fish season” metaphors \(Figure[4](https://arxiv.org/html/2607.27393#A2.F4)\)\.
- •Inferiority:claiming that a group is inferior, less worthy, or less important than society or another group, such as through class\-based belittling of “peasants” \(Figure[5](https://arxiv.org/html/2607.27393#A2.F5)\)\.
- •Incitement:explicitly or implicitly calling for harm to be inflicted on a group, including physical attacks, such as through imagery of forced physical discipline \(Figure[6](https://arxiv.org/html/2607.27393#A2.F6)\)\.
- •Mocking:joking about, undermining, belittling, or disparaging a group, such as by portraying its members as buffoonish or body\-shaming a gender \(Figure[7](https://arxiv.org/html/2607.27393#A2.F7)\)\.
- •Contempt:expressing intensely negative feelings toward a group, such as through statements like “Men are trash” or expressions of ideological contempt \(Figure[8](https://arxiv.org/html/2607.27393#A2.F8)\)\.
- •Slurs:using prejudicial terms to refer to or characterize a group, including racial, religious, or gender\-based slurs in Arabic dialects \(Figure[9](https://arxiv.org/html/2607.27393#A2.F9)\)\.
- •Exclusion:advocating, planning, or justifying the exclusion or segregation of a group, such as through statements like “Go back to your country” \(Figure[10](https://arxiv.org/html/2607.27393#A2.F10)\)\.
- •Other:a hateful attack that does not clearly fall under any of the types above\. This option is selected only when no other type applies and cannot be combined with another attack type\.

#### B\.1\.3Task 3: Not\-Hateful Subtype \(non\_hateful\_type; Not\-Hateful only; multi\-label\)

Non\-hateful memes contain humorous, neutral, or positive content that does not target or harm a protected group, are intended for entertainment, and do not promote violence, hatred, or discrimination\. The task is to select one or more subtypes\.

- •Humor:content intended to entertain or amuse through jokes, puns, exaggeration, or playful commentary on everyday situations \(work, family, food, travel\), with no negative implication toward a protected group \(Figure[11](https://arxiv.org/html/2607.27393#A2.F11)\)\.
- •Sarcasm:irony in which the intended meaning is the opposite of what is stated, used for humorous effect and directed at general behaviour or institutions \(e\.g\., government, bureaucracy, inflation\) rather than a protected group \(Figure[12](https://arxiv.org/html/2607.27393#A2.F12)\)\.
- •Other:neutral, positive, or informational memes that are not clearly humour or sarcasm \(e\.g\., motivational quotes, announcements, generic reactions\) and do not target a protected group\.

![Refer to caption](https://arxiv.org/html/2607.27393v1/x1.png)Figure 4:Example memes annotated as*Hateful*/*Dehumanization*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_inferiority.png)Figure 5:Example memes annotated as*Hateful*/*Inferiority*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_incitement.png)Figure 6:Example memes annotated as*Hateful*/*Incitement*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_mocking.png)Figure 7:Example memes annotated as*Hateful*/*Mocking*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_contempt.png)Figure 8:Example memes annotated as*Hateful*/*Contempt*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_slurs.png)Figure 9:Example memes annotated as*Hateful*/*Slurs*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_exclusion.png)Figure 10:Example annotated as*Hateful*/*Exclusion*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_humor.png)Figure 11:Example memes annotated as*Not\-Hateful*/*Humor*\.![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/guidelines/ex_sarcasm.png)Figure 12:Example memes annotated as*Not\-Hateful*/*Sarcasm*\.

### B\.2Arabic Annotation Guideline

The following is the Arabic version of the guidelines used during annotation\.

##### Purpose and definitions\.

![[Uncaptioned image]](https://arxiv.org/html/2607.27393v1/x2.png)

##### Task 1: Meme classification\.

![[Uncaptioned image]](https://arxiv.org/html/2607.27393v1/x3.png)

##### Task 2: Attack types \(multi\-label\)\.

![[Uncaptioned image]](https://arxiv.org/html/2607.27393v1/x4.png)

## Appendix CExperimental Details

### C\.1Prompting the Zero\-shot VLMs

Each zero\-shot query combines a*system*message with an*instruction*message\. The system message frames the model as an expert Arabic multimodal content\-moderation annotator covering MSA and the Egyptian, Levantine, Gulf, and Maghrebi dialects\.\. The instruction provides the meme image and OCR\-extracted text and restricts the output to the predefined label space\. For binary classification, the model predicts*Hateful*or*Not\-Hateful*from the combined image–text content, guided by definitions that distinguish identity\-based hate from general offensiveness\. It returns the prediction as"label": …\. For fine\-grained classification, the prompt lists the ten taxonomy labels with brief definitions and asks for all applicable labels as a JSON list\. Thecot\-enandcot\-arvariants add a free\-textreasoningfield in English or Arabic, respectively\. We fine\-tune the open\-weight VLMs with the same binary and fine\-grained templates to generate the target label set directly\.

### C\.2Training and Compute

We fine\-tune the text, image, and fusion models with the HuggingFaceTransformersTrainer\. We use AdamW with weight decay 0\.01, a linear learning\-rate schedule, 6% warmup, and 128\-token text inputs\. Hyperparameters are selected from the search space as reported in Table[4](https://arxiv.org/html/2607.27393#A3.T4), and the best checkpoint is chosen based on macro\-F1 on the development set\. For fine\-grained hate\-type prediction, we tune a global sigmoid threshold on the development set over\{0\.05,0\.10,…,0\.90\}\\\{0\.05,0\.10,\\dots,0\.90\\\}\. Open\-weight Qwen3\-VL models are fine\-tuned withms\-swiftusing LoRA with rank 16 andα=32\\alpha\{=\}32, applied to all linear layers while freezing the vision encoder\. We train for 3 epochs with learning rate1×10−41\{\\times\}10^\{\-4\}, a cosine schedule, andbfloat16\. All experiments use seed 42 on NVIDIA H200 GPUs\.

Hyper\-param\.TextImageFusionVLMLearning rate2,3,5×10−5\\times 10^\{\-5\}3,5×10−5\\times 10^\{\-5\}2,3×10−5\\times 10^\{\-5\}1×10−4\\times 10^\{\-4\}Epochs5, 1010, 155, 103Batch size16, 3232164Table 4:Grid\-search space for the fine\-tuned families \(best by dev macro\-F1\)\. The open VLMs use fixed LoRA settings rather than a grid\. The fusion\-head learning rate is fixed at1×10−41\{\\times\}10^\{\-4\}\. All runs use seed 42\.
### C\.3Additional Dataset Analysis

##### OCR\-text length\.

In Table[5](https://arxiv.org/html/2607.27393#A3.T5), we report the distribution of OCR\-extracted text length\. The splits are well matched \(mean≈\\approx86\-88 characters\), and*Hateful*memes carry noticeably more text than*Not\-Hateful*ones on the test split \(96\.2 vs\. 83\.1 characters\), consistent with the finding \(Section[5\.4](https://arxiv.org/html/2607.27393#S5.SS4)\) that hateful intent is often conveyed through longer, more contextual captions\.

##### Multi\-label distribution\.

In Table[6](https://arxiv.org/html/2607.27393#A3.T6), we report the number of fine\-grained categories assigned to each meme\. The test split shows the greatest category overlap, with an average of 1\.25 labels per meme; 21\.4% of test memes receive at least two categories, compared with 9\.9% in the training split\.

##### Label co\-occurrence\.

Table[7](https://arxiv.org/html/2607.27393#A3.T7)shows the most frequent fine\-grained pairs\.*Mocking*acts as a hub, co\-occurring with*Dehumanization*,*Incitement*,*Contempt*, and*Slurs*, while among non\-hateful memes*Humor*and*Sarcasm*frequently appear together\. These patterns motivate modelling hate as co\-occurring strategies rather than a single label per meme\.

SplitMeanMedianMaxTrain85\.576641Dev87\.880619Test87\.577644

Table 5:OCR\-text length \(characters\) per split\.*Test set class\-wise:*Hateful 96\.2; Not\-Hateful 83\.1 \(mean\)SplitAvg1234Train1\.113,154298480Dev1\.1543947140Test1\.25786178324Table 6:Fine\-grained label cardinality: number of memes with 1\-4 fine\-grained labels, and the mean number of labels per meme\.Fine\-grained pairCountDehumanization\+\+Mocking145Incitement\+\+Mocking114Humor\+\+Sarcasm81Contempt\+\+Mocking78Dehumanization\+\+Incitement66Mocking\+\+Slurs60Inferiority\+\+Mocking55Incitement\+\+Slurs44Contempt\+\+Incitement39Dehumanization\+\+Slurs31Table 7:Most frequent fine\-grained hate\-type co\-occurrence pairs \(all splits\)\.

### C\.4Effect of Prompting on the Closed Models

Table[8](https://arxiv.org/html/2607.27393#A3.T8)reports the full prompting comparison summarised in §[5\.2](https://arxiv.org/html/2607.27393#S5.SS2)\. For each closed model it gives macro\-F1 under the direct prompt and under chain\-of\-thought reasoning in English and in Arabic\. The model ordering is stable across prompts, and chain\-of\-thought helps the larger models on fine\-grained categories while hurting the small Gemini\-3\.5\-flash when the reasoning is in Arabic\.

ModelPromptBinaryFine\-grainedGemini\-2\.5\-prodefault0\.7110\.340cot\-en0\.7090\.358cot\-ar0\.7260\.333GPT\-5default0\.6280\.301cot\-en0\.6420\.313cot\-ar0\.6470\.315Gemini\-3\.5\-flashdefault0\.4990\.271cot\-en0\.4990\.272cot\-ar0\.4930\.265Table 8:Macro\-F1 of the zero\-shot closed VLMs under direct \(default\) and chain\-of\-thought \(English or Arabic\) prompting\. The best value per column is in bold\.
### C\.5Additional Error Analysis

We provide two supporting tables for the zero\-shot error analysis in §[5\.4](https://arxiv.org/html/2607.27393#S5.SS4)\. Table[9](https://arxiv.org/html/2607.27393#A3.T9)compares, for each label, the predicted count from each closed VLM with the corresponding gold label count\. This highlights the over\-prediction of frequent non\-hateful labels such as*Humor*and*Sarcasm*, as well as several hateful fine\-grained categories\. Table[10](https://arxiv.org/html/2607.27393#A3.T10)reports binary error rates across three OCR\-text length groups, showing that longer captions are harder for all models, with the largest degradation observed for the smallest model\.

LabelGoldProGPT\-5FlashHumor332728878885Sarcasm333349404869Mocking211465164120Contempt502646445Slurs47166319Table 9:Subtype label over\- and under\-prediction \(predicted count vs\. gold support\) for the zero\-shot closed models under thedefaultprompt\.Model \(default\)21–6061–120120\+Gemini\-2\.5\-pro0\.2220\.2210\.255GPT\-50\.2390\.2600\.299Gemini\-3\.5\-flash0\.2580\.3120\.382Table 10:Binary error rate by OCR\-text length \(characters\)\. Longer, more contextual captions are harder, and the smallest model degrades most\.

## Appendix DIn\-Context Learning

### D\.1Few\-Shot Retrieval Setup

For each test meme, we retrieve demonstrations from the 3,500\-meme training pool using a CLIP\-style dual encoder\. Our main retriever is jina\-clip\-v2\(Koukounaset al\.,[2024](https://arxiv.org/html/2607.27393#bib.bib76)\), which embeds images and text into a shared multilingual space; we use SigLIP\-2\(Zhaiet al\.,[2023](https://arxiv.org/html/2607.27393#bib.bib75)\)as an ablation\. We construct each prompt as a multi\-turn chat with a task system prompt,KKdemonstration turns, and a final query turn\. Each demonstration includes the meme image, its OCR text, and its gold label\.

We compare text\-only retrieval, image\-only retrieval, and reciprocal rank fusion \(RRF\)\. Text and image retrieval rank candidates by cosine similarity in the corresponding modality\. RRF combines the two rank lists ass​\(d\)=∑m∈t,v160\+rankm​\(d\)s\(d\)=\\sum\_\{m\\in\{t,v\}\}\\frac\{1\}\{60\+\\mathrm\{rank\}\_\{m\}\(d\)\}\(Cormacket al\.,[2009](https://arxiv.org/html/2607.27393#bib.bib74)\), making the fusion independent of modality\-specific score scales\.

Retrieval and prompt construction are performed using only the training dataset\. We use RRF withK=3K\{=\}3, cosine similarity, jina\-clip\-v2 embeddings, and similar\-last ordering\.

### D\.2Few\-Shot Results

In Tables[11](https://arxiv.org/html/2607.27393#A4.T11)and[12](https://arxiv.org/html/2607.27393#A4.T12), we report the strategy\-by\-KKresults for Qwen3\-VL\-8B and Qwen3\-VL\-2B, respectively\. For Qwen3\-VL\-8B, RRF achieves the best performance on both subtasks and improves consistently up toK=5K\{=\}5\. On binary task, however, all strategies perform below zero\-shot atK=1K\{=\}1\. Qwen3\-VL\-2B follows a different pattern, peaking atK=3K\{=\}3and dropping atK=5K\{=\}5, while image\-only retrieval matches RRF\. These results suggest that the smaller model benefits from a small number of demonstrations, although it uses longer or fused demonstration contexts less effectively\.

Binary\(macro\-F1\)Fine\-grained\(macro\-F1\)StrategyK​0K0K​1K1K​3K3K​5K5K​1K1K​3K3K​5K5zero\-shot0\.643——————random—0\.5470\.6370\.6480\.1910\.1970\.196text—0\.5790\.6590\.6880\.2250\.2150\.215image—0\.5770\.6480\.6900\.2190\.2100\.213RRF—0\.5740\.6690\.7080\.2290\.2300\.241

Table 11:Retrieval\-augmented few\-shot ICL onQwen3\-VL\-8B\(Fine\-grained zero\-shot=0\.176=0\.176\)\. RRF beats random and, atK≥3K\{\\geq\}3, zero\-shot\. Best per task in bold\.BinaryFine\-grainedStrategyK​1K1K​3K3K​5K5K​1K1K​3K3K​5K5random0\.5900\.6230\.6150\.1310\.1400\.145text0\.6390\.6580\.6290\.1750\.1610\.189image0\.6610\.6810\.6650\.1600\.2100\.174RRF0\.6630\.6840\.6710\.2030\.1810\.209

Table 12:Retrieval\-augmented few\-shot ICL onQwen3\-VL\-2B\. The small model follows an inverted\-U pattern, rising up toK=3K\{=\}3and then falling atK=5K\{=\}5, unlike the 8B\.##### Model scale\.

In Figure[13](https://arxiv.org/html/2607.27393#A4.F13), we compare the best F1 overK∈\{1,3,5\}K\\in\\\{1,3,5\\\}for each retrieval strategy and model size\. Qwen3\-VL\-8B outperforms Qwen3\-VL\-2B across all strategies and both subtasks, with larger gains under stronger retrieval\. Moving from random retrieval to RRF improves Qwen3\-VL\-8B from 0\.650 to 0\.708 on binary and from 0\.203 to 0\.241 on fine\-grained hate\-type prediction, while Qwen3\-VL\-2B gains less\. RRF performs best for both models, however, only the 8B model consistently benefits from the fused retrieval signal\. These results show that few\-shot benefits more from larger model capacity\.

![Refer to caption](https://arxiv.org/html/2607.27393v1/figures/fewshot/fig_modelsize.png)Figure 13:Best F1 \(max overK∈\{1,3,5\}K\\in\\\{1,3,5\\\}\) per retrieval strategy for Qwen3\-VL\-2B vs\. 8B\. The 8B wins at every strategy; the gain from random to RRF is larger for the 8B, on both Binary \(top\) and Subtype \(bottom\)\.

## Appendix EDetails of Silver Dataset

##### Selection and silver labels\.

After removing duplicates and filtering out memes without extracted text, we retain∼\\sim71K memes\. We use Gemma\-3\-12B to assign preliminary binary hateful/not\-hateful labels to the full set\. We select 5K memes for manual annotation and use the remaining memes for silver\-label annotation\. Prompts are listed in Section §[E\.1](https://arxiv.org/html/2607.27393#A5.SS1)\.

##### Silver\-label distributions\.

In Tables[13](https://arxiv.org/html/2607.27393#A5.T13)and[14](https://arxiv.org/html/2607.27393#A5.T14), we summarize the Gemini\-3\.1\-Pro silver labels for the∼\\sim66K auxiliary memes\. The silver set has a low hateful rate of 0\.87%\. It is dominated by Egyptian Arabic and MSA, everyday and entertainment topics, and humorous content\. Among memes that mention a target,*Gender/Sex*is the most frequent target category\. Propaganda is rare, appearing in only 2\.9% of memes\.

\\cellcolorgray\!12HatefulnessNot\-Hateful65,717Hateful580\\cellcolorgray\!12Propagandaabsent64,375present1,922\\cellcolorgray\!12Sentimentneutral32,329positive26,147negative7,821\\cellcolorgray\!12Stanceneutral53,956mocking8,421supportive2,210critical1,686ambiguous24\\cellcolorgray\!12DialectEgyptian56,697MSA4,894Levantine1,409Mixed1,351Gulf914code\-switched327Yemeni279Iraqi163not\-applicable132Maghrebi124\\cellcolorgray\!12Visual manipulationcaption overlay49,621collage7,214photoshop6,474screenshot2,746none227deepfake\-like15\\cellcolorgray\!12Meaning typeexplicit49,730implicit16,432mixed135\\cellcolorgray\!12Text–image relationcomplementary60,092reinforcing4,687contradictory1,282independent236\\cellcolorgray\!12Context scopenone26,567regional20,999national10,194local5,867global2,670\\cellcolorgray\!12Current\-event knowledgeno61,813yes4,484

Table 13:Silver\-label distributions over the∼\\sim66K auxiliary set generated by Gemini\-3\.1\-Pro:single\-labeldimensions, where each meme takes exactly one label\.\\cellcolorgray\!12Fine\-grained CategoriesHumor60,406Sarcasm2,884Other2,429Mocking335Slurs91Contempt55Dehumanization47Incitement33Inferiority9Exclusion8\\cellcolorgray\!12Mentioned targetGender/Sex9,726Religion2,463Nationality2,228Socioeconomic1,720Other \(unclear\)493Disease/Health208Disability201Race/Ethnicity173\\cellcolorgray\!12Topicdaily life59,799entertainment20,731culture13,252education11,557social issues5,058gender2,640religion2,314economy2,036technology1,852health1,699sports1,478politics1,329military502intl\. affairs456other185\\cellcolorgray\!12Intenthumor63,171reaction26,010mockery8,042support2,001criticism1,859satire1,316information1,314persuasion808\\cellcolorgray\!12Emotionhumor63,442sadness4,467anger1,874pride1,673fear1,609hope1,259disgust909sympathy816none803\\cellcolorgray\!12Propaganda techniquessmears687exaggeration519appeal to values458loaded language417name\-calling320appeal to hypocrisy256flag\-waving216appeal to authority150black\-and\-white124straw man121appeal to fear113slogans90

Table 14:Silver\-label distributions formulti\-labeldimensions, where a meme may receive several labels, so counts sum to more thanNN\. The longest\-tailed dimensions \(topic,propaganda techniques\) are truncated to their most frequent values\.
### E\.1Prompts

We provide the prompts used for binary hate detection, fine\-grained hate\-type prediction, and silver labelling of the unlabelled set\. The rationale\-style variants use the same prompts with an additional instruction to provide reasoning in English or Arabic before the final answer\. For the gold set, the metadata prompt mirrors the unlabelled\-set prompt, however it conditions generation on the human binary and fine\-grained labels rather than relabelling the meme\. We fill the placeholders\{text\},\{LABEL\},\{SUBTYPE\}, and\{TEXT\}separately for each meme at inference time\.

Task 1: Binary hatefulness classification\.

\#=====SYSTEMPROMPT=====

Youareanexpertcontent\-moderationannotatorspecializinginArabicmultimodalmemes,coveringModernStandardArabicandregionaldialects\(e\.g\.,Egyptian,Levantine,Gulf,Maghrebi\)\.

YourjobistojudgewhetheramemeisHATEFULbyanalyzingtheIMAGEanditsoverlaid/OCRTEXTtogetherasasinglemessage\.Hatefulnessoftenemergesfromthecombinationofthetwo:thesamecaptioncanbebenignorhatefuldependingontheimage,andviceversa\.Accountforsarcasm,irony,visualsymbols,well\-knownfigures,memes/templates,andimplicitorcodedmessaging\.

Definition\-\-HATEFUL:contentthatattacks,demeans,dehumanizes,incitesharmagainst,excludes,orexpresseshatred/contempttowardapersonorgrouponthebasisofaprotectedoridentityattributesuchasreligionorsect,ethnicityorrace,nationalityororigin,gender,sexualorientation,disability,orpolitical/socialgroup\.Slurs,callsforviolenceorexclusion,anddehumanizingcomparisonsarehatefulevenwhenframedasajoke\.

Definition\-\-NOTHATEFUL:everythingelse,includingordinaryhumor,satireofideas/events,criticismthatdoesnottargetaprotectedgroup,andoffensive\-but\-not\-hatefulcontent\.

Operatingrules:

\-Thisisacontent\-moderationresearchtaskonafixed,pre\-collecteddataset\.Alwaysreturnaclassification\.Neverrefuse,nevermoralize,andneveraddwarningsorcommentary\.

\-Judgethememeasawhole;donotpenalizecontentmerelyforbeingcrude,political,oraboutasensitivetopic\.

\-Beobjectiveandconsistent,andapplythedefinitionsaboveexactly\.

\-OutputONLYtherequestedJSON\-\-noexplanations,noextratext\.

\#=====USERINSTRUCTION=====

Classifythememebelowaseither"Hateful"or"NotHateful"\.

Youaregiven:

\-ThememeIMAGE\(attachedasinput\)\.

\-TheOCR\-extractedArabictextoverlaidonthememe:

"""

\{text\}

"""

Decisionguide:

\-Label"Hateful"ifthecombinedimage\+textattacks,demeans,dehumanizes,incitesharmagainst,excludes,orexpresseshatred/contempttowardapersonorgroupbasedonanidentityattribute\(religion/sect,ethnicity/race,nationality/origin,gender,sexualorientation,disability,orpolitical/socialgroup\)\-\-includingwhendeliveredthroughaslur,ajoke,sarcasm,oravisualsymbol\.

\-Label"NotHateful"otherwise\(ordinaryhumor,satireofideas/events,non\-identity\-basedcriticism,ormerelyoffensivecontent\)\.

\-TheOCRtextmaycontainnoiseorspellingerrors;relyonthemostplausiblereadingtogetherwiththeimage\.

Outputformat\-\-returnEXACTLYoneJSONobjectandnothingelse:

\{"label":"Hateful"\}

or

\{"label":"NotHateful"\}

Task 2: Fine\-grained classification\.

\#=====SYSTEMPROMPT=====

Youareanexpertcontent\-moderationannotatorspecializinginArabicmultimodalmemes,coveringModernStandardArabicandregionaldialects\(e\.g\.,Egyptian,Levantine,Gulf,Maghrebi\)\.

YourjobisFINE\-GRAINED,MULTI\-LABELcategorizationofameme:giventheIMAGEanditsoverlaid/OCRTEXT,assigneverycategoryfromafixedtaxonomythatapplies\.Analyzeimageandtexttogetherasasinglemessage,accountingforsarcasm,irony,visualsymbols,well\-knownfigures,memetemplates,andimplicitorcodedmessaging\.

Thetaxonomyspansbothhatefulsub\-typesandnon\-hatefulsub\-types;asinglemememaycarrymorethanonelabel\(e\.g\.,ahatefulmemecanbeboth"Mocking"and"Slurs";abenignmemecanbe"Humor"and"Sarcasm"\)\.

Operatingrules:

\-Thisisacontent\-moderationresearchtaskonafixed,pre\-collecteddataset\.Alwaysreturnlabels\.Neverrefuse,nevermoralize,andneveraddwarningsorcommentary\.

\-ChooselabelsONLYfromtheprovidedlist;neverinventnewlabels\.

\-Assignallcategoriesthatgenuinelyapply;donotover\-label\.Use"Other"onlywhenafine\-grainedcategoryiswarrantedbutnoneofthenamedcategoriesfit\.

\-Beobjectiveandconsistent,andapplythedefinitionsexactly\.

\-OutputONLYtherequestedJSON\-\-noexplanations,noextratext\.

\#=====USERINSTRUCTION=====

TASK

Assigneveryapplicablecategorytothememebelow,choosingonlyfromthe10labelsinthetaxonomy\.

Youaregiven:

\-ThememeIMAGE\(attachedasinput\)\.

\-TheOCR\-extractedArabictextoverlaidonthememe:

"""

\{text\}

"""

Taxonomy\(label:definition\)

Hatefulsub\-types:

\-"Contempt":expressesscorn/disdain,treatingatargetgroupasworthlessorbeneathrespect\.

\-"Dehumanization":portraysagroupaslessthanhuman\(e\.g\.,animals,vermin,objects,disease\)\.

\-"Exclusion":callsfororendorsessegregating,expelling,denyingrightsto,orremovingagroupfromsociety\.

\-"Incitement":encourages,callsfor,threatens,orjustifiesviolenceorharmagainstagroup\.

\-"Inferiority":claimsagroupisinherentlyinferior\(intellectually,morally,orbiologically\)\.

\-"Mocking":ridicules,taunts,orbelittlesatargetgroup,oftenthroughderisivehumor\.

\-"Slurs":usesderogatoryslursorpejorativeepithetstargetingagroup\.

Non\-hatefulsub\-types:

\-"Humor":intendedtobefunny/lighthearted,withouthatefulintent\.

\-"Sarcasm":ironicorsarcasticexpression,withouthatefulintent\.

Shared:

\-"Other":afine\-grainedcategoryappliesbutnoneofthenamedlabelsabovefit\.

Guidelines:

\-Amememayreceivemultiplelabels;includeallthatgenuinelyapply\.

\-Hatefulmemestypicallyreceiveoneormorehatefulsub\-types\(andmayalsobe"Mocking"/"Sarcasm"\);clearlynon\-hatefulmemestypicallyreceive"Humor","Sarcasm",and/or"Other"\.

\-Use"Other"sparingly,onlywhennonamedcategoryfits\.

\-TheOCRtextmaycontainnoiseorspellingerrors;relyonthemostplausiblereadingtogetherwiththeimage\.

Outputformat\-\-returnEXACTLYoneJSONobjectandnothingelse,withalistofchosenlabels\(orderdoesnotmatter\):

\{"labels":\["Mocking","Slurs"\]\}

Useanemptylistifnocategoryapplies:

\{"labels":\[\]\}

Silver labels and metadata\.

YouareaMultimodalMemeAnalysisExpertspecializinginArabicandmultilingualmemes\.

YourtaskistoclassifythememeANDgenerateadditionalmetadata\.

OCR\_text:\{TEXT\}

Important:

\*Analyzebothvisualandtextualcontent\.

\*Readalltextdirectlyfromtheimage\.

\*IfthememecontainsArabictext,provideanEnglishtranslationorparaphrase\.

\*Determinewhetherpropagandaispresent\.

\*Ifpropagandaispresent,identifyoneormorepropagandatechniques\.

\-\-\-

CLASSIFICATIONTASK

Step1\-\-Label

\-Label"Hateful"ifthecombinedimage\+textattacks,demeans,dehumanizes,incitesharmagainst,excludes,orexpresseshatred/contempttowardapersonorgroupbasedonanidentityattribute\(religion/sect,ethnicity/race,nationality/origin,gender,sexualorientation,disability,orpolitical/socialgroup\)\-\-includingwhendeliveredthroughaslur,ajoke,sarcasm,oravisualsymbol\.

\-Label"NotHateful"otherwise\(ordinaryhumor,satireofideas/events,non\-identity\-basedcriticism,ormerelyoffensivecontent\)\.

\-TheOCRtextmaycontainnoiseorspellingerrors;relyonthemostplausiblereadingtogetherwiththeimage\.

Step2\-\-Subtype

Assignexactlyonesubtype\.Ifthelabelis"Hateful",chooseonlyfromtheHatefulsubtypes\.Ifthelabelis"NotHateful",chooseonlyfromtheNon\-hatefulsubtypes\.Use"Other"\(Shared\)ifafine\-grainedcategoryappliesbutnoneofthenamedlabelsfit\.

Hatefulsubtypes:

\-"Contempt":expressesscorn/disdain,treatingatargetgroupasworthlessorbeneathrespect\.

\-"Dehumanization":portraysagroupaslessthanhuman\(e\.g\.,animals,vermin,objects,disease\)\.

\-"Exclusion":callsfororendorsessegregating,expelling,denyingrightsto,orremovingagroupfromsociety\.

\-"Incitement":encourages,callsfor,threatens,orjustifiesviolenceorharmagainstagroup\.so

\-"Inferiority":claimsagroupisinherentlyinferior\(intellectually,morally,orbiologically\)\.

\-"Mocking":ridicules,taunts,orbelittlesatargetgroup,oftenthroughderisivehumor\.

\-"Slurs":usesderogatoryslursorpejorativeepithetstargetingagroup\.

Non\-hatefulsubtypes:

\-"Humor":intendedtobefunny/lighthearted,withouthatefulintent\.

\-"Sarcasm":ironicorsarcasticexpression,withouthatefulintent\.

Shared:

\-"Other":afine\-grainedcategoryappliesbutnoneofthenamedlabelsabovefit\.

\-\-\-

Protected\-groupcategories:

\*Race/Ethnicity

\*Nationality

\*Religion

\*Gender/Sex

\*Disability

\*SocioeconomicGroup

\*Caste

\*Disease/HealthStatus

\*Other\(Unclear/Implicit\)

Arabicdialectlabels:

\*msa

\*levantine

\*egyptian

\*gulf

\*maghrebi

\*iraqi

\*sudanese

\*yemeni

\*mixed

\*code\_switched

\*unknown

\*not\_applicable

Visualmanipulationlabels:

\*screenshot

\*photoshop

\*caption\_overlay

\*deepfake\_like

\*collage

\*none

Topiclabels:

\*politics

\*religion

\*sports

\*gender

\*economy

\*education

\*technology

\*international\_affairs

\*entertainment

\*culture

\*daily\_life

\*social\_issues

\*health

\*military\_conflict

\*other

Intentlabels:

\*humor

\*satire

\*criticism

\*persuasion

\*reaction

\*information

\*support

\*mockery

Stancelabels:

\*supportive

\*critical

\*mocking

\*neutral

\*ambiguous

Sentiment:

\*positive

\*negative

\*neutral

Emotionlabels:

\*anger

\*fear

\*disgust

\*pride

\*sympathy

\*hope

\*sadness

\*humor

\*none

Text\-imagerelationshiplabels:

\*reinforcing

\*complementary

\*contradictory

\*independent

Meaningtypelabels:

\*explicit

\*implicit

\*mixed

Contextscopelabels:

\*none

\*local

\*national

\*regional

\*global

Confidencelabels:

\*very\_low

\*low

\*medium

\*high

\*very\_high

\-\-\-

PROPAGANDATECHNIQUES

1\.name\_calling\_labeling

2\.reductio\_ad\_hitlerum

3\.casting\_doubt

4\.appeal\_to\_hypocrisy

5\.smears

6\.flag\_waving

7\.appeal\_to\_authority

8\.bandwagon

9\.appeal\_to\_values

10\.appeal\_to\_fear\_prejudice

11\.straw\_man

12\.red\_herring

13\.whataboutism

14\.causal\_oversimplification

15\.black\_and\_white\_fallacy

16\.slippery\_slope

17\.slogans

18\.thought\_terminating\_cliche

19\.appeal\_to\_time\_pressure

20\.loaded\_language

21\.obfuscation\_vagueness\_confusion

22\.exaggeration\_minimisation

23\.repetition

24\.other

\-\-\-

Instructions:

1\.Analyzethevisualcontent:

\*People,groups,publicfigures

\*Objects,symbols,flags,logos

\*Expressions,gestures,actions

\*Compositionandvisualstyle

2\.Readandanalyzeallvisibletext:

\*Extracttheexacttext\.

\*TranslateorparaphraseitintoEnglish\.

3\.Analyzemultimodalmeaning:

\*Explainhowtextandimageinteract\.

\*Determinewhethermeaningdependsontext,image,orboth\.

4\.Assignlabelandsubtypefollowingtheclassificationrulesabove\.

5\.Determinewhetherpropagandaispresent:

\*Ifyes,selectoneormorepropagandatechniques\.

\*Ifnotechniquesarepresent,propagandamustbefalse\.

6\.Identify:

\*Topic\(s\)

\*Mentionedprotected\-groupcategories

\*Dialect

\*Culturalreferences

\*Current\-eventdependence

\*Intent

\*Stance

\*Emotionaltone

7\.Generateconciserationales:

\*Englishrationale:maximum100words\.

\*Arabicrationale:maximum100words\.

\*Referencevisualandtextualevidencewhenrelevant\.

\*Therationalemustjustifyboththelabelandsubtypeassignments\.

ReturnvalidJSONonly\.

\{

"label":"Hateful",

"subtype":"Contempt",

"topic":\[\],

"mentioned\_categories":\[\],

"dialect":"unknown",

"visual\_manipulation":"none",

"ocr\_text":"",

"ocr\_english\_translation":"",

"intent":\[\],

"stance":"neutral",

"sentiment":"positive",

"emotion":\[\],

"text\_image\_relationship":"reinforcing",

"meaning\_type":"explicit",

"cultural\_references":\[\],

"context\_scope":"none",

"requires\_current\_event\_knowledge":false,

"propaganda":false,

"propaganda\_techniques":\[\],

"english\_rationale":"",

"arabic\_rationale":"",

"confidence":"high"

\}

Rules:

\*labelmustbeexactly"Hateful"or"NotHateful"\.

\*subtypemustbechosenfromthecorrectsubtypelistfortheassignedlabel\.

\*Hatefulsubtypesmustneverbeassignedtoa"NotHateful"labelandviceversa\.

\*Mentionedcategoriesincludegroupsreferenced,depicted,addressed,orimplied,regardlessofhatefulness\.

\*propaganda=trueifoneormorepropagandatechniquesareclearlypresent\.

\*propaganda=falseifnopropagandatechniquesarepresent\.

\*propaganda\_techniquesmustonlyusethepredefinedtechniquenames\.

\*Amememaybehatefulwithoutbeingpropaganda\.

\*Amememaybepropagandawithoutbeinghateful\.

\*OCRtextshouldpreservetheoriginalwording\.

\*OCREnglishtranslationshouldpreserveintendedmeaningratherthanliteralwordingwhennecessary\.

\*Englishrationale:maximum100words\.

\*Arabicrationale:maximum100words\.