Predicting how audiences react to Arabic social media posts requires reasoning beyond textual sentiment: reactions emerge from collective interpretation moderated by engagement dynamics and topical context. We present a multi-task learning (MTL) framework that jointly learns (i) audience reaction classification (Love, Haha, Angry, Sad, Care, Wow), (ii) engagement magnitude regression (six reactions, comments, shares), and (iii) non-engagement detection. On a corpus of 158k Arabic Facebook posts spanning women’s rights, gender debates, and economic empowerment, our model achieves a test macro-F1 of 72.4 and weighted-F1 of 89.1.

We present ArabMisinfo, a corpus of 274,837 public Facebook Page posts in Arabic spanning 2014-2024 and focused on news and public affairs. Each record includes page-level metadata, UTC timestamps, link domains, and platform-native engagement signals (likes, comments, shares, and the seven post-2016 reaction types). We accompany the corpus with reproducible diagnostics for engagement and polarization, the Polarization Index (PI), Share-to-Like Anomaly (SLA), Risk Triage Index (RTI), and Engagement Anomaly Rule (EPR). It provides scripts that deterministically reproduce all tables and figures. To address the absence of content-level labels, we contribute a discourse role annotation of 10,000 posts (3.6% of the corpus) using Fanar-C-2-27B, classifying each post into one of six roles. The annotation reveals that Arabic Facebook misinformation discourse is dominated by counter-information activity: 30.6% of posts are Rumor Denial or Official Correction, outpacing Unverified Claim posts (12.3%) by a 2.5:1 ratio, while 19.3% represent incidental keyword matches, directly quantifying the sampling noise introduced by keyword-based collection. To comply with platform terms and copyright, we publicly release Post IDs, Page IDs, and full metadata and derived indices; message text is available only under a research data-use agreement (DUA). ArabMisinfo enables engagement-aware analyses of Arabic misinformation ecosystems and, to our knowledge, provides the one of the first discourse-role-annotated Arabic Facebook corpora at this scale.

We present the first cross-platform computational analysis of AI-sustainability discourse across LinkedIn, X, and Bluesky, using 52,102 posts collected over 15 months. Combining RoBERTa-based sentiment classification with BERTopic topic modeling, we identify 36 topics and compare sentiment, topic distributions, and engagement across platforms. Overall, 43.98% of posts are positive, 40.16% neutral, and 15.86% negative. Sentiment varies by topic rather than by platform: corporate and AI-potential themes are mostly positive, while resource-cost topics carry the highest negative sentiment. Topic distributions differ significantly across platforms: LinkedIn focuses on corporate sustainability, X on hardware and cryptocurrency, and Bluesky on environmental resource costs. Niche topics consistently attract more engagement than high-volume ones. These findings have implications for how organizations and policymakers communicate about AI-sustainability, as topic framing and platform audience composition shape public discourse more than platform selection alone. We release the annotated dataset to support future research.

This study evaluates three approaches—instruction prompting of large language models (LLMs), instruction fine-tuning of LLMs, and transformer-based pretrained models on emotion detection in Arabic social media text. We compare pretrained transformer models like AraBERT, CaMelBERT, and XLM-RoBERTa with instruction prompting with advanced LLMs like GPT-4o, Gemini, Deepseek, and Fanar, and instruction fine-tuning approaches with LLMs like Llama 3.1, Mistral, and Phi. With a highly preprocessed dataset of 10,000 labeled Arabic tweets with overlapping emotional labels, our findings reveal that transformer-based pretrained models outperform instruction prompting and instruction fine-tuning approaches. Instruction prompts leverage general linguistic skills with maximum efficiency but fall short in detecting subtle emotional contexts. Instruction fine-tuning is more specific but trails behind pretrained transformer models. Our findings establish the need for optimized instruction-based approaches and underscore the important role played by domain-specific transformer architectures in accurate Arabic emotion detection.

This paper presents the MAHED 2025 Shared Task on Multimodal Detection of Hope and Hate Emotions in Arabic Content, comprising three subtasks: (1) text-based classification of Arabic content into hate and hope, (2) multi-task learning for joint prediction of emotions, offensive content, and hate speech, and (3) multimodal detection of hateful content in Arabic memes. We provide three high-quality datasets totaling over 22,000 instances sourced from social media platforms, annotated by native Arabic speakers with Cohen’s Kappa exceeding 0.85. Our evaluation attracted 46 leaderboard submissions from participants, with systems leveraging Arabic-specific pre-trained language models (AraBERT, MARBERT), large language models (GPT-4, Gemini), and multimodal fusion architectures combining CLIP vision encoders with Arabic text models. The best-performing systems achieved macro F1-scores of 0.723 (Task 1), 0.578 (Task 2), and 0.796 (Task 3), with top teams employing ensemble methods, class-weighted training, and OCR-aware multimodal fusion. Analysis reveals persistent challenges in dialectal robustness, minority class detection for hope speech, and highlights key directions for future Arabic content moderation research.