Elamur Research⌕
☰
home / papers / 2026 / ArPoMeme: An Annotated Arabic Multimodal Dataset for Political Ideology and Polarization

ArPoMeme: An Annotated Arabic Multimodal Dataset for Political Ideology and Polarization

Paper · 2026 · Wajdi Zaghouani, Kais Attia, Md. Rafiul Biswas et al.

Summary

Arabic political memes are a major channel for ideological expression online. Existing Arabic meme resources, however, concentrate on propaganda and hate, so the spread of ideologies across memes has not been mapped. ArPoMeme fills this gap with about 7,300 political memes collected from public Facebook pages and groups. The memes are sorted into four orientations: Leftist, Islamist, Pan-Arabist and political satire.

The authors do not train a model to infer ideology. Instead, each meme inherits its label from how its source community presents itself through page names, bios and stated positions. Collection is semi-automated:

  • Playwright scrapes the pages.
  • Images are synced to Google Drive.
  • Qwen2.5-VL-7B extracts the overlaid text, which humans then verify.

Using a custom Streamlit tool, annotators label three polarization signals: content that sets an in-group against others, antagonism aimed at opposing groups, and appeals urging people to act. The analysis finds these signals unevenly distributed, with Islamist and satirical memes showing the most hostility and mobilization. The dataset and the tool are released publicly as a reproducible base for Arabic multimodal ideology and polarization research.

Key points

  • About 7,300 Arabic political memes in four classes: Leftist, Islamist, Pan-Arabist and Satirical. Satire is deliberately kept as its own category rather than folded into the others.
  • Ideology labels come from the self-identification of the Facebook pages and groups that produce the memes, not from model predictions.
  • Pipeline: Playwright-based scraping, then Google Drive synchronization, then text extraction with Qwen2.5-VL-7B, then manual verification.
  • Polarization annotation uses three dimensions (Us vs. Them, out-group hostility, calls to action). It is done in a Streamlit interface with distributed labeling, real-time progress tracking and version control.
  • Islamist and satirical memes score highest on hostility and mobilization cues. The authors position the resource against Arabic meme work such as ArMeme and the Propaganda to Hate extension of ArAIEval-2024.

Interesting ideas

  • I1. Labels are taken from the declared identity of the source community, which makes large-scale labeling cheap without a classifier in the loop; this works whenever a source reliably signals the property of its content.
  • I2. A vision–language model does the extraction and humans only verify its output, so people correct a draft instead of writing from scratch; this cuts annotation cost while keeping quality.
  • I3. Satire is treated as its own orientation rather than forced into an ideological slot, since ironic content mocks several sides at once and would otherwise pollute other classes.
  • I4. Polarization is split into three separable, concrete cues (in-group/out-group framing, hostility, mobilizing appeals) instead of one vague score, which makes labels more consistent and easier to explain.
  • I5. Annotation cues are compared across source groups to expose asymmetries, turning the dataset into a diagnostic as well as training material.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Silver data could be tagged by provenance (e.g., texts from sources that declare a genre such as fables or school tests), giving cheap genre labels that witness parsers can be checked against.
  • I2 (learn): An LLM teacher tier could draft frame or tree annotations that a human or a stricter checker only confirms or fixes, logging every correction as a candidate repair rule.
  • I3 (frames): The irony component could output "satirical stance" as a distinct value, instead of mapping ironic utterances onto a literal speaker position, so later steps do not take mockery as endorsement.
  • I4 (frames): Speech-act analysis could flag directives aimed at a group (a call to action) and explicit we/they contrasts as separate, explainable features attached to the frame.
  • I5 (learn): Evaluation could report error rates per text genre or source, revealing where parsers or readers behave unevenly rather than hiding it in one averaged score.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Social Sciences · 744 works Sociology and Political Science · 221 works

🧭 Topics 2 Vision & language · 269 works Semitic languages · 1,484 works

Cite

Cite as Wajdi Zaghouani, Kais Attia, Md. Rafiul Biswas and Fadhl Eryani. 2026. ArPoMeme: An Annotated Arabic Multimodal Dataset for Political Ideology and Polarization. Proceedings of the Fifteenth Language Resources and Evaluation Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/5ghumwhwumkz

@inproceedings{zaghouani2026arpomeme,
  title = {{ArPoMeme: An Annotated Arabic Multimodal Dataset for Political Ideology and Polarization}},
  author = {Wajdi Zaghouani, Kais Attia, Md. Rafiul Biswas and Fadhl Eryani},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/5ghumwhwumkz},
  url = {https://doi.org/10.63317/5ghumwhwumkz}
}

References

28 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Firoj Alam, Md. Rafiul Biswas, Uzair Shah, Wajdi Zaghouani, and Georgios Mikros. 2025. Propaganda to hate: A multimodal analysis of arabic memes with multi-agent llms. In Web Information Systems Engineering – WISE 2024, pages 380– 390, Singapore. Springer Nature Singapore. our card
  2. Firoj Alam, Abul Hasnat, Fatema Ahmad, Md Arid Hasan, and Maram Hasanain. 2024. Armeme: Propagandistic content in arabic memes. pages 21071–21090. Association for Computational Linguistics (ACL). find
  3. Marwah Kareem Ali and Wafaa Dahham Mohammed. 2023. The reflection of satire in political cartoons. JOURNAL OF LANGUAGE STUDIES, 5(2):431–440. ↗
  4. Ahmad Alzu’bi, Lojin Bani Younis, Abdelrahman Abuarqoub, and Mohammad Hammoudeh. 2023. Multimodal deep learning with discriminant descriptors for offensive memes detection. J. Data and Information Quality, 15(3). ↗
  5. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923. Ana-Maria Bliuc, John M. Betts, Matteo Vergani, Ayoub Bouguettaya, and Mioara Cristea. A theoretical framework for polarization as the gradual fragmentation of a divided society. 2(1):75. ↗ our card
  6. Anis Charfi, Mabrouka Bessghaier, Raghda Akasheh, Andria Atalla, and Wajdi Zaghouani. 2024a. Hate speech detection with adhar: a multi-dialectal hate speech corpus in arabic. Frontiers in Artificial Intelligence, 7. ↗
  7. Anis Charfi, Mabrouka Bessghaier, Andria Atalla, Raghda Akasheh, Sara Al-Emadi, and Wajdi Zaghouani. 2024b. Stance detection in arabic with a multi-dialectal cross-domain stance corpus. Social Network Analysis and Mining, 14. ↗
  8. Noam Gal, Limor Shifman, and Zohar Kampf. 2016. “it gets better”: Internet memes and the construction of collective identity. New Media & Society, 18(8):1698–1714. find
  9. Heba Heraki and Wajdi Zaghouani. 2025. Analyzing digital polarization on hijab: A dataset of annotated youtube comments. Proceedings of the International AAAI Conference on Web and Social Media, 19:2350–2360. ↗
  10. Eftekhar Hossain, Omar Sharif, and Mohammed Moshiul Hoque. 2022. MemoSen: A multimodal dataset for sentiment analysis of memes. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1542–1554, Marseille, France. European Language Resources Association. Lenka Hrbková, Jakub Macek, and Alena Macková. How does the “us” versus “them” polarization work? capturing political antagonism with the political antagonism scale. 39(1):100–120. find
  11. Michael Johann and Lars Bülow. 2019. One does not simply create a meme: Conditions for the diffusion of internet memes. International journal of communication, 13:23. ↗
  12. Sanaa Laabar and Wajdi Zaghouani. 2024. Multidimensional insights: Annotated dataset of stance, sentiment, and emotion in Facebook comments on Tunisia’s July 25 measures. In Proceedings of the Second Workshop on Natural Language Processing for Political Sciences @ LREC-COLING 2024, pages 22–32, Torino, Italia. ELRA and ICCL. find
  13. Anne Leiser. 2022. Psychological perspectives on participatory culture: Core motives for the use of political internet memes. Journal of Social and Political Psychology, 10(1). ↗
  14. Liam McLoughlin and Rosalynd Southern. 2021. By any memes necessary? small political acts, incidental exposure and memes during the 2017 uk general election. The British Journal of Politics and International Relations, 23(1):60–84. ↗
  15. Mihaela-Georgiana Mihăilescu. 2024. Never mess with the “memers”: How meme creators are redefining contemporary politics. Social Media + Society, 10(4):20563051241296256. ↗
  16. Ahd Mohamed and Wajdi Zaghouani. 2024. Expression of depression among arab twitter users using arabic corpus analysis. Procedia Computer Science, 244:76–85. 6th International Conference on AI in Computational Linguistics. Buddhika Nettasinghe, Allon G Percus, and Kristina Lerman. How out-group animosity can shape partisan divisions: A model of affective polarization. 4(3):pgaf082. find
  17. Francisco Rangel, Paolo Rosso, Wajdi Zaghouani, and Anis Charfi. 2020. Fine-grained analysis of language varieties and demographics. Natural Language Engineering, 26(6):641–661. Publisher Copyright: © Cambridge University Press 2020. ↗
  18. Vanessa Romo. 2023-04-25. ’dark brandon’ meme makes an appearance on biden’s new campaign website. NPR. find
  19. Elad Segev, Asaf Nissenbaum, Nathan Stolero, and Limor Shifman. 2015. Families and networks of internet memes: The relationship between cohesiveness, uniqueness, and quiddity concreteness. Journal of Computer-Mediated Communication, 20(4):417–433. ↗
  20. Uzair Shah, Md Biswas, Marco Agus, Mowafa Househ, and Wajdi Zaghouani. 2024. Mememind at araieval shared task: Generative augmentation and feature fusion for multimodal propaganda detection in arabic memes through advanced language and vision models. pages 467– 472. Association for Computational Linguistics (ACL). find
  21. Anatolii Shestakov and Wajdi Zaghouani. 2024. Analyzing conflict through data: A dataset on the digital framing of sheikh jarrah evictions. In Proceedings of the Second Workshop on Natural Language Processing for Political Sciences @ LREC-COLING 2024, pages 55–67, Torino, Italia. ELRA and ICCL. find
  22. LIMOR SHIFMAN. 2014. Memes in Digital Culture. The MIT Press. ↗
  23. Chereen Shurafa and Wajdi Zaghouani. 2024. Sentiment analysis and emotion annotation of a large-scale arabic youtube trauma corpus. pages 1–7. find
  24. Rika Suryaningsih. 2025. Memes as ideological texts: A multimodal discourse analysis of political semiotics in digital cultures. International Journal of Humanities Science Innovations and Management Studies IJHSIMS, 2(2). ↗
  25. Shardul Suryawanshi, Bharathi Raja Chakravarthi, Mihael Arcan, and Paul Buitelaar. 2020. Multimodal meme dataset (MultiOFF) for identifying offensive content in image and text. In Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying, pages 32–41, Marseille, France. European Language Resources Association (ELRA). our card
  26. Winda Widyaningsih, Rizky Ramadhani, and Edy Prihantoro. 2025. Political polarization through political memes for new voters in the 2024 presidential election. Bricolage : Jurnal Magister Ilmu Komunikasi, 11:165. ↗
  27. Bo Xu, Tingting Li, Junzhe Zheng, Mehdi Naseriparsa, Zhehuan Zhao, Hongfei Lin, and Feng Xia. 2022. Met-meme: A multimodal meme dataset rich in metaphors. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’22, page 2887–2899, New York, NY, USA. Association for Computing Machinery. Wajdi Zaghouani, Hamdy Mubarak, and Md. find
  28. Rafiul Biswas. 2024. So hateful! building a multi-label hate speech annotated Arabic dataset. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 15044–15055, Torino, Italia. ELRA and ICCL. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zaghouani-2026-arpomeme-annotated-arabic-multimodal-dataset/