Confidence-aware system combines Arabic text and images to flag hateful memes

An open-source content moderation system that combines Arabic-language analysis with image recognition achieved an area under the receiver operating characteristic curve of 0.78 and a macro F1 score of 0.71 when tested on a held-out set of Arabic-translated hateful memes. The software pairs the MARBERTv2 Arabic language encoder with a frozen CLIP ViT-B/32 visual encoder, then routes uncertain cases to a vision-language model for secondary review. Its developers describe the result as a proof-of-concept for confidence-aware moderation rather than a production-ready system, noting that the evaluation was conducted on machine-translated English benchmark data rather than memes originally written in Arabic.

The work targets a form of online abuse that can evade systems examining only one part of a meme. A meme combines an image with a short caption, and its discriminatory meaning may depend on the relationship between them. A caption that appears ordinary in isolation, for example, can become harmful when placed over an image depicting a particular religion, ethnicity, gender or nationality. In the Facebook Hateful Memes benchmark used here, models based solely on text or images have performed near chance, illustrating why researchers designed the dataset with benign confounders whose meaning changes when image and text are considered together.

Arabic presents additional complications for automated analysis. The language has substantial differences between Modern Standard Arabic and regional dialects, and memes may combine dialectal vocabulary, transliterated foreign words and irregular spelling. MARBERTv2 was selected to represent the textual side of the system because it is a 163-million-parameter BERT-style model pretrained on one billion Arabic social-media posts covering both Modern Standard Arabic and dialectal varieties. The model accepts sequences of up to 128 tokens and converts them into a 768-dimensional representation through attention-masked mean pooling, providing a language-specific signal rather than relying on an English text encoder.

The image branch uses the visual half of CLIP ViT-B/32, a contrastively pretrained vision transformer associated with 400 million image-text pairs. Images are resized to 224 by 224 pixels, center-cropped and normalized with CLIP-specific values. The encoder produces a 512-dimensional representation, and its weights remain frozen during training to reduce overfitting on the relatively small dataset and preserve representations learned during large-scale pretraining. Separate linear layers, layer normalization and GELU activations project the 768-dimensional text vector and 512-dimensional image vector into a common 512-dimensional space. Concatenating the two projections yields a 1,024-dimensional representation for a three-layer multilayer perceptron that outputs a binary hate probability.

For training, the researchers translated the captions of roughly 9,000 examples from the Facebook Hateful Memes dataset into Arabic with the 600-million-parameter NLLB-200-distilled-600M translation model. The translated data were divided by stratified sampling into 6,300 training examples, 1,800 validation examples and 900 test examples, preserving class proportions. The model was optimized with AdamW, using a learning rate of 2 × 10−5 for MARBERTv2 and 1 × 10−3 for the projection and classification layers, along with weight decay, gradient clipping and a short warmup schedule. Training could run for up to 50 epochs, with early stopping based on validation macro F1.

The test results at a decision threshold of 0.53 showed that the system favored precision on the non-hate class and recall on the hate class, while performing more evenly across the two classes overall. The non-hate class had a precision of 0.77, recall of 0.82 and F1 score of 0.79. For the hate class, precision was 0.66, recall was 0.59 and F1 was 0.62. The macro averages were 0.72 for precision, 0.70 for recall and 0.71 for F1. In a same-split ablation, a text-only MARBERTv2 classifier obtained a macro F1 of 0.62, nine percentage points below the fused image-and-text model, indicating that the visual branch supplied additional discriminative information within this translated-data setting.

The published comparisons place the system among several distinct experimental groups rather than treating all scores as directly equivalent. On the original English Hateful Memes benchmark, the reported BERT text-only baseline achieved an AUROC of 0.65 and macro F1 of 0.59, while ViLBERT reached an AUROC of 0.70 and macro F1 of 0.63. Hate-CLIPper, using CLIP for both image and text, reported 0.858 AUROC and 0.811 macro F1 on the English data. A separate ResNet-RoBERTa fusion system with SMOTE and incremental PCA reported a macro F1 of 0.8153. These English results are not a controlled comparison with the Arabic-translated experiment, but they show the scale of the gap between the new pipeline and stronger approaches developed on the original dataset.

Native Arabic benchmarks provide additional context but cannot establish a direct ranking. The MAHED 2025 shared task included more than 22,000 Arabic instances annotated by native speakers, and its best Task 3 system reported a macro F1 of 0.796. YassirEA, another participant, paired MARBERTv2 with the larger CLIP ViT-L/14 encoder and combined four fusion strategies, reporting a macro F1 of 0.75 on MAHED. The new system was not evaluated on that native-Arabic dataset, so its 0.71 score on translated Hateful Memes cannot be treated as a like-for-like comparison. The authors identify evaluation on MAHED and incorporation of natively authored Arabic memes as important future steps.

Beyond the classifier, the software includes a bilingual web interface that demonstrates confidence-aware routing. EasyOCR first extracts Arabic or Latin-script text from a flagged image. The fusion model then assigns a hate probability: scores of at least 0.53 produce a REMOVE decision, scores below 0.40 produce ALLOW, and intermediate scores are sent to a hosted Llama 4 Scout vision-language model for a structured REMOVE, WARN or ALLOW decision with a one-sentence explanation. In one illustrative case, a score of 41.7% fell inside the escalation band and the vision-language judge returned REMOVE despite the first-stage model’s low hate probability. If the external API fails, the system defaults to REMOVE. Decisions, confidence scores, timestamps and explanations are stored in an in-memory moderation log.

Several limitations constrain the findings. Native Arabic speakers informally scanned all translated captions and closely reviewed 100 examples selected to include slang, sarcasm and idiomatic language, finding no systematic gross mistranslation, but this was not a blind annotation protocol or an inter-annotator agreement study. The authors also did not formally relabel the dataset to verify that every hate or non-hate label remained valid after translation. Only captions were translated, leaving image-text alignment unassessed, and the system was not tested on natively authored Arabic meme content. At inference, EasyOCR may struggle with stylized fonts, low-resolution images and mixed Arabic-Latin scripts, creating a gap between clean translated training captions and noisy extracted text. English OCR output is also passed directly to MARBERTv2 without translation. The released code supports the inference pipeline, but the training and translation scripts are absent, latency and computational costs were not benchmarked, and the escalation workflow was not evaluated under real-world traffic. The result therefore demonstrates a reusable technical route for multimodal Arabic moderation, while leaving its performance on culturally specific Arabic memes unestablished.

Subject of Research: Confidence-aware multimodal detection of hateful memes using Arabic text and image encoders

Article Title: An agentic-style, confidence-aware multimodal system for Arabic hate speech detection in memes

Article References: Boulaamail, M. A., Hanine, M., Cheikh Tourad, M., Alomran, A., & Alshuhri, S. (2026). An agentic-style, confidence-aware multimodal system for Arabic hate speech detection in memes. SoftwareX, 36, Article 103134. https://doi.org/10.1016/j.softx.2026.103134

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103134

Keywords: Arabic language, hate speech detection, multimodal models, meme moderation, machine translation, confidence routing, vision-language model, content moderation, open-source software, agentic-style, confidence-aware, multimodal