Elamur Research⌕
☰
home / papers / 2025 / BZU-AUM@ImageEval2025: An Arabic Image Captioning Dataset for Conflict Narratives with Human Annotation

BZU-AUM@ImageEval2025: An Arabic Image Captioning Dataset for Conflict Narratives with Human Annotation

Paper · 2025 · Mohammed Alkhanafseh, Ola Surakhi and Abdallah Abedaljalill

Summary

The authors address the lack of Arabic image-captioning data for emotionally and historically loaded material. Existing caption corpora such as MS-COCO, Flickr30k and Visual Genome are mostly English and cover neutral, everyday scenes. For the ImageEval 2025 Shared Task, the team built a human-annotated resource of photographs about Palestinian life under occupation, displacement, protest and cultural continuity. Native speakers wrote one Modern Standard Arabic caption per image, aiming for 40–70 words. Two prompting schemes guided the writers. With Chain-of-Thought, they moved from visible elements to actions and then to symbolic or historical meaning. With Role-Based prompting, they wrote as a journalist, eyewitness or humanitarian. Each caption went through two rounds of review. Most captions meet the length target, and they follow a consistent pattern of description first, interpretation second. The authors report that their submission scored best on meaning accuracy and fluency. The resource matters for training and testing Arabic vision-language models on culturally dense content, where literal description is not enough.

Key points

  • The data has 500 images with one caption each, split into two batches of 250 and grouped into five thematic sheets. Topics include the 1948 Nakba, the 1967 war, Jerusalem, cities occupied in 1948, daily life and the Rif resistance.
  • Annotation rules required MSA only, 40–70 words, no speculation and no identification of minors. The two review passes checked length, grammar, register, factual grounding, tone and role.
  • Statistics:
    • mean of 53.99 words and median of 51;
    • 2.72 sentences per caption on average;
    • 79.2% of captions inside the target band;
    • range of 3 to 148 words.
  • Words were counted by whitespace and sentences by Arabic/English punctuation. The data ships as a CSV with image_id, caption and batch_id; prompts and scripts are promised for release after review.
  • The paper contradicts itself on size: the limitations section mentions only 250 images. It also lists exclusion of dialects and possible misuse or retraumatization as limitations.

Interesting ideas

  • I1. Prompting techniques usually meant for LLMs (Chain-of-Thought, role prompts) are applied here to human annotators, which makes the writing more uniform and structured.
  • I2. Each caption is layered: concrete objects and setting first, then events, then cultural or emotional significance. This keeps the grounded part separable from the interpretive part.
  • I3. The same scene is voiced through different stances (objective journalist vs. emotional witness), varying style while the content stays fixed.
  • I4. A ban on speculation is enforced at review time, with a separate check for factual grounding.
  • I5. The paper reports compliance with a target band ("% within 40–70 words"), computed by a deliberately simple, transparent script, instead of reporting only averages.

For Mova

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

  • I1 (learn): when collecting gold or teacher data, give annotators or LLM teachers the same staged checklist, so outputs become comparable across annotators.
  • I2 (back): generation from the meaning graph could follow the same order, realizing entities, then event frames, then inferred implications, so each sentence traces to a known layer.
  • I3 (back): render one world graph in several registers to test that the meaning is preserved and only surface style changes.
  • I4 (frames): mark each frame as directly stated or as implicature, so the reader can tell what the text asserts from what it suggests.
  • I5 (learn): evaluation reports could give the share of outputs meeting each explicit constraint, which exposes tail failures that a mean hides.

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

 Institutions

🎓Birzeit UniversityBirzeit · Palestine38 works · 798 citations🎓American University of MadabaMadaba · Jordan2 works

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 3,066 worksVision & language · 811 works

Cite

Cite as Mohammed Alkhanafseh, Ola Surakhi and Abdallah Abedaljalill. 2025. BZU-AUM@ImageEval2025: An Arabic Image Captioning Dataset for Conflict Narratives with Human Annotation. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.arabicnlp-sharedtasks.54

@inproceedings{alkhanafseh2025bzu,
  title = {{BZU-AUM@ImageEval2025: An Arabic Image Captioning Dataset for Conflict Narratives with Human Annotation}},
  author = {Mohammed Alkhanafseh, Ola Surakhi and Abdallah Abedaljalill},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.54},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.54}
}

References

21 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Muhammad Abdul-Mageed, AbdelRahim Elmadany, Chiyu Zhang, El Moatez Billah Nagoudi, Houda Bouamor, and Nizar Habash. 2023. Nadi 2023: The fourth nuanced arabic dialect identification shared task. arXiv preprint arXiv:2310.16117. ↗
  2. Alfred V. Aho and Jeffrey D. Ullman. 1972. The Theory of Parsing, Translation and Compiling, volume 1. Prentice-Hall, Englewood Cliffs, NJ. find
  3. American Psychological Association. 1983. Publications Manual. American Psychological Association, Washington, DC. find
  4. Rie Kubota Ando and Tong Zhang. 2005. A framework for learning predictive structures from multiple tasks and unlabeled data. Journal of Machine Learning Research, 6:1817–1853. ↗
  5. Galen Andrew and Jianfeng Gao. 2007. Scalable training of L1-regularized log-linear models. In Proceedings of the 24th International Conference on Machine Learning, pages 33–40. Ahlam Bashiti, Alaa Aljabari, Hadi Hamoud, Md. find
  6. Rafiul Biswas, Bilal Shalash, Mustafa Jarrar, Fadi Zaraket, George Mikros, Ehsaneddin Asgari, and Wajdi Zaghouani. 2025. ImageEval 2025: The First Arabic Image Captioning Shared Task. In Proceedings of the Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou, China. Association for Computational Linguistics.
  7. Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, and 1 others. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712. ↗
  8. Ashok K. Chandra, Dexter C. Kozen, and Larry J. Stockmeyer. 1981. Alternation. Journal of the Association for Computing Machinery, 28(1):114–133. find
  9. Jiaxin Ge, Hongyin Luo, Siyuan Qian, Yulu Gan, Jie Fu, and Shanghang Zhang. 2023. Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919. ↗
  10. Dan Gusfield. 1997. Algorithms on Strings, Trees and Sequences. Cambridge University Press, Cambridge, UK. ↗
  11. Mohammed Kharma, Soohyeon Choi, Mohammed AlKhanafseh, and David Mohaisen. 2025. Security and quality in llm-generated code: A multi-language, multi-model analysis. arXiv preprint arXiv:2502.01853. ↗
  12. Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, and 1 others. 2017. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32–73.
  13. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer. ↗
  14. Dong Nguyen and Esther Ploeger. 2025. We need to measure data diversity in nlp–better and broader. arXiv preprint arXiv:2505.20264. ↗
  15. Mohammad Sadegh Rasooli and Joel R. Tetreault. 2015. Yara parser: A fast and accurate dependency parser. Computing Research Repository, arXiv:1503.06733. Version 2. ↗
  16. Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. 2023. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317. ↗
  17. Antonio Torralba and Alexei A Efros. 2011. Unbiased look at dataset bias. In CVPR 2011, pages 1521–1528. IEEE. find
  18. Ning Wang, Jiahao Xie, Jihao Wu, Mingbo Jia, and Linlin Li. 2023. Controllable image captioning via prompting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2617–2625. find
  19. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837.
  20. Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. 2016. Image captioning with semantic attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4651–4659. find
  21. Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. 2023. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168–5191. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alkhanafseh-2025-bzu-aum-imageeval2025-arabic-image/