---
title: "AraMinds at MAHED 2025 — paper by Zaytoon et al. (2025)"
description: "The paper describes the AraMinds team's entry for subtask 3 of the MAHED 2025 shared task. 27 references, cited by 1 paper here."
authors: ["Mohamed Zaytoon", "Ahmed Mahmoud Salem", "Ahmed Sakr", "Hossam Elkordi"]
year: 2025
venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks"
source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.85/"
doi: "10.18653/v1/2025.arabicnlp-sharedtasks.85"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/zaytoon-2025-araminds-mahed-2025-leveraging-vision/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zaytoon-2025-araminds-mahed-2025-leveraging-vision/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/13/) / AraMinds at MAHED 2025: Leveraging Vision-Language Models and Contrastive Multi-task Learning for Multimodal Hate Speech Detection

# AraMinds at MAHED 2025: Leveraging Vision-Language Models and Contrastive Multi-task Learning for Multimodal Hate Speech Detection

**Authors:** [Mohamed Zaytoon](https://md.research.elamur.ai/free/en/authors/mohamed-zaytoon/), [Ahmed Mahmoud Salem](https://md.research.elamur.ai/free/en/authors/ahmed-mahmoud-salem/), [Ahmed Sakr](https://md.research.elamur.ai/free/en/authors/ahmed-sakr/) and [Hossam Elkordi](https://md.research.elamur.ai/free/en/authors/hossam-elkordi/) · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.85/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.85.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.85](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.85) · [OpenAlex](https://openalex.org/W4416037801)

**Citations:** 1 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/zaytoon-2025-araminds-mahed-2025-leveraging-vision/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper describes the AraMinds team's entry for subtask 3 of the MAHED 2025 shared task. The task is a binary hate / no-hate decision on Arabic memes. Memes are difficult because irony and humour can hide hostile intent, Arabic dialects vary widely, topics range across politics, religion and gender, and non-hateful examples dominate the data.

The authors built three systems:
- **Dual-encoder with late fusion.** A text encoder (MARBERTv2 or Qwen2.5) is paired with a CNN image encoder (ResNet / ResNeXt). Their features are concatenated and trained with BCE plus an in-batch contrastive loss.
- **Multi-task variant.** It adds an auxiliary head that predicts nine fine-grained categories supplied with the dataset.
- **Fine-tuned VLM.** Gemma3-4B is fine-tuned with an instruction prompt and generates the label as text.

A simple majority-vote ensemble of the three reached macro-F1 0.74 on the test set and placed third. The work suggests that cheap ingredients can make a small multimodal dataset go further: auxiliary labels, a carefully placed contrastive objective, and voting across different architectures.

## Key points
- **Data:** memes collected from Facebook, Instagram and Pinterest, then de-duplicated and sampled from about 6k candidates. Splits are 2,452 train, 606 validation and 500 test.
- **Multi-tasking** with the nine-way fine-grained head raised validation macro-F1 from 0.689 to 0.702 (Qwen2.5-1.5B + ResNet-18).
- **Text encoder size and type:**
  - Moving from Qwen2.5-0.5B to Qwen2.5-1.5B gave +0.039.
  - Encoder-only MARBERTv2 was slightly better still (0.705).
  - A larger image encoder (ResNeXt-101 vs ResNet-18) brought almost no gain.
- **Placement of the contrastive loss:**
  - On text embeddings only: 0.702.
  - On fused embeddings: 0.666.
  - On text + fused: 0.687.
  - The fine-tuned Gemma3-4B VLM scored 0.692.
- **Test set:** single systems scored 0.71–0.72, and the max-voting ensemble scored 0.74. All training ran on one RTX-3090.

## Interesting ideas
- When a dataset includes a fine-grained label that correlates with the main label, a second classification head can use it as extra supervision at almost no cost, which helps when data is scarce.
- Applying a contrastive objective to only one representation, and testing which one, can matter more than adding it everywhere, since contrasting the fused vector actually hurt here.
- Fine-tuning a generative model while computing loss only on the short answer tokens turns an instruction-following model into a clean classifier with a constrained output space.
- Equal-weight majority voting over architecturally different models (encoder, decoder, VLM) beat every single member, because their errors differ.
- The error analysis showed that one channel can mask another: obvious sarcasm in the text hid hostility in the image, and a cartoon style hid a hateful religious stance.

## For Mova

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

- **I1** (tree): The parser could add an auxiliary head that predicts fine-grained relation subtypes or construction classes, so coarse UD label decisions get a richer training signal from small treebanks.
- **I2** (learn): When adding contrastive or metric losses to the tagger or parser, run an ablation over which layer receives the loss rather than assuming the deepest shared vector is best.
- **I3** (learn): LLM teacher tiers could be tuned to emit only a fixed label vocabulary (for example a frame or relation name), so their silver outputs are easier to compare across witnesses.
- **I4** (learn): Silver-data agreement among the perceptron parser, gparse and other witnesses works best when the witnesses are deliberately diverse, and the vote can serve as a confidence signal.
- **I5** (frames): Irony and implicature detection should check whether a surface tone such as joking conflicts with the propositional content, and flag the mismatch explicitly rather than letting tone decide.

*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

- [Alexandria University](https://md.research.elamur.ai/free/en/universities/alexandria-university/) — University, Alexandria, Egypt · 39 works · 287 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,772 works · [🤝 Sociology and Political Science](https://md.research.elamur.ai/free/en/shelves/social-sciences/sociology-and-political-science/) · 577 works

🧭 Topics (4): [😊 Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) · 972 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works · [🖼️ Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) · 657 works · [🛡️ Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/) · 517 works

## Cite

Cite as: Mohamed Zaytoon, Ahmed Mahmoud Salem, Ahmed Sakr and Hossam Elkordi. 2025. AraMinds at MAHED 2025: Leveraging Vision-Language Models and Contrastive Multi-task Learning for Multimodal Hate Speech Detection. 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.

```bibtex
@inproceedings{zaytoon2025araminds,
  title = {{AraMinds at MAHED 2025: Leveraging Vision-Language Models and Contrastive Multi-task Learning for Multimodal Hate Speech Detection}},
  author = {Mohamed Zaytoon, Ahmed Mahmoud Salem, Ahmed Sakr and Hossam Elkordi},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.85},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.85}
}
```

- [← previous Arabizi vs LLMs: Can the Genie Understand the Language of Aladdin?](https://md.research.elamur.ai/free/en/papers/almaoui-2025-arabizi-vs-llms-can-genie/)
- [next → AraTable: Benchmarking LLMs' Reasoning and Understanding of Arabic Tabular Data](https://md.research.elamur.ai/free/en/papers/alshaikh-2025-aratable-benchmarking-llms-reasoning-understanding/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/13/)

## References

27 references, 11 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, and El Moatez Billah Nagoudi. 2021. [ARBERT & MARBERT: Deep bidirectional transformers for Arabic](https://md.research.elamur.ai/free/en/papers/abdul-mageed-2021-arbert-marbert-deep-bidirectional-transformers/). In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing. Association for Computational Linguistics. [↗](https://doi.org/10.48550/arxiv.2101.01785)
2. Ibrahim Ahmed, Mostafa Abbas, Rany Hatem, Andrew Ihab, and Mohamed Waleed Fahkr. 2022. Fine-tuning arabic pre-trained transformer models for egyptian-arabic dialect offensive language and hate speech detection and classification. In 2022 20th International Conference on Language Engineering (ESOLEC). [find](https://scholar.google.com/scholar?q=Ibrahim+Ahmed%2C+Mostafa+Abbas%2C+Rany+Hatem%2C+Andrew+Ihab%2C+and+Mohamed+Waleed+Fahkr%2E+2022%2E+Fine%2Dtuning+arabic+pre%2Dtrained+transformer+models+for+egyptian%2Darabic+dialect+offensive+language+and+hate+speech+)
3. Samar Al-Saqqa, Arafat Awajan, and Bassam Hammo. 2024. A survey of hate speech detection for arabic social media: Methods and datasets. Procedia Computer Science. 15th International Conference on Emerging Ubiquitous Systems and Pervasive Networks / 14th International Conference on Current and Future Trends of Information and Communication Technologies in Healthcare EUSPN/ICTH 2024. [find](https://scholar.google.com/scholar?q=Samar+Al%2DSaqqa%2C+Arafat+Awajan%2C+and+Bassam+Hammo%2E+2024%2E+A+survey+of+hate+speech+detection+for+arabic+social+media%3A+Methods+and+datasets%2E+Procedia+Computer+Science%2E+15th+International+Conference+on+Emer)
4. Firoj Alam, Md Rafiul Biswas, Uzair Shah, Wajdi Zaghouani, and Georgios Mikros. 2024a. [Propaganda to hate: A multimodal analysis of arabic memes with multi-agent llms](https://md.research.elamur.ai/free/en/papers/alam-2024-propaganda-hate-multimodal-analysis-arabic/). In International Conference on Web Information Systems Engineering. Springer.
5. Firoj Alam, Abul Hasnat, Fatema Ahmed, Md Arid Hasan, and Maram Hasanain. 2024b. [Armeme: Propagandistic content in arabic memes](https://md.research.elamur.ai/free/en/papers/alam-2024-armeme-propagandistic-content-arabic-memes/). arXiv preprint arXiv:2406.03916. [↗](https://arxiv.org/abs/2406.03916)
6. Greeshma Arya, Mohammad Kamrul Hasan, Ashish Bagwari, Nurhizam Safie, Shayla Islam, Fatima Rayan Awad Ahmed, Aaishani De, Muhammad Attique Khan, and Taher M Ghazal. 2024. Multimodal hate speech detection in memes using contrastive language-image pre-training. IEEe Access. [↗](https://doi.org/10.1109/access.2024.3361322)
7. Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C. Li, Adrien Bardes, Suzanne Petryk, Oscar Mañas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, Mark Ibrahim, Melissa Hall, Yunyang Xiong, Jonathan Lebensold, Candace Ross, Srihari Jayakumar, Chuan Guo, Diane Bouchacourt, Haider Al-Tahan, and 22 others. 2024. An introduction to vision-language modeling. ArXiv. [find](https://scholar.google.com/scholar?q=Florian+Bordes%2C+Richard+Yuanzhe+Pang%2C+Anurag+Ajay%2C+Alexander+C%2E+Li%2C+Adrien+Bardes%2C+Suzanne+Petryk%2C+Oscar+Ma%C3%B1as%2C+Zhiqiu+Lin%2C+Anas+Mahmoud%2C+Bargav+Jayaraman%2C+Mark+Ibrahim%2C+Melissa+Hall%2C+Yunyang+Xiong%2C+J)
8. Hossam Elkordi, Ahmed Sakr, Marwan Torki, and Nagwa El-Makky. 2024. [AlexuNLP24 at AraFinNLP2024: Multi-dialect Arabic intent detection with contrastive learning in banking domain](https://md.research.elamur.ai/free/en/papers/elkordi-2024-alexunlp24-arafinnlp2024-multi-dialect-arabic/). In Proceedings of the Second Arabic Natural Language Processing Conference, Bangkok, Thailand. Association for Computational Linguistics. [↗](https://doi.org/10.13140/rg.2.2.14741.74729)
9. Nizar Y Habash. 2010. Introduction to Arabic natural language processing. Morgan & Claypool Publishers. [↗](https://doi.org/10.1007/978-3-031-02139-8)
10. Maram Hasanain, Md Arid Hasan, Fatema Ahmed, Reem Suwaileh, Md Rafiul Biswas, Wajdi Zaghouani, and Firoj Alam. 2024. [Araieval shared task: propagandistic techniques detection in unimodal and multimodal arabic content](https://md.research.elamur.ai/free/en/papers/hasanain-2024-araieval-shared-task-propagandistic-techniques/). arXiv preprint arXiv:2407.04247. [↗](https://arxiv.org/abs/2407.04247)
11. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. [Deep residual learning for image recognition](https://md.research.elamur.ai/free/en/papers/he-2015-deep-residual-learning-image-recognition/). In Proceedings of the IEEE conference on computer vision and pattern recognition. [↗](https://doi.org/10.48550/arxiv.1512.03385)
12. Fatemah Husain and Ozlem Uzuner. 2022. Investigating the effect of preprocessing arabic text on offensive language and hate speech detection. Transactions on Asian and Low-Resource Language Information Processing. [↗](https://doi.org/10.1145/3501398)
13. Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. 2020. [The hateful memes challenge: Detecting hate speech in multimodal memes](https://md.research.elamur.ai/free/en/papers/kiela-2020-hateful-memes-challenge-detecting-hate/). Advances in neural information processing systems. [↗](https://doi.org/10.48550/arxiv.2005.04790)
14. PM Ashok Kumar and K Varalakshmi. 2021. Hate speech detection using text and image tweets based on bi-directional long short-term memory. In 2021 International conference on disruptive technologies for multi-disciplinary research and applications (CENT-CON). IEEE. [find](https://scholar.google.com/scholar?q=PM+Ashok+Kumar+and+K+Varalakshmi%2E+2021%2E+Hate+speech+detection+using+text+and+image+tweets+based+on+bi%2Ddirectional+long+short%2Dterm+memory%2E+In+2021+International+conference+on+disruptive+technologies+fo)
15. Phillip Lippe, Nithin Holla, Shantanu Chandra, Santhosh Rajamanickam, Georgios Antoniou, Ekaterina Shutova, and Helen Yannakoudakis. 2020. A multimodal framework for the detection of hateful memes. arXiv preprint arXiv:2012.12871. [↗](https://arxiv.org/abs/2012.12871)
16. Yuyang Chen and Feng Pan. 2022. Multimodal detection of hateful memes by applying a vision-language pre-training model. Plos one. [↗](https://doi.org/10.1371/journal.pone.0274300)
17. Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. In International Conference on Learning Representations. [↗](https://doi.org/10.48550/arxiv.1711.05101)
18. Anusha Chhabra and Dinesh Kumar Vishwakarma. 2023. A literature survey on multimodal and multilingual automatic hate speech identification. Multimedia Systems. [↗](https://doi.org/10.1007/s00530-023-01051-8)
19. Hala Mulki, Hatem Haddad, Chedi Bechikh Ali, and Halima Alshabani. 2019. [L-hsab: A levantine twitter dataset for hate speech and abusive language](https://md.research.elamur.ai/free/en/papers/mulki-2019-l-hsab-levantine-twitter-dataset/). In Proceedings of the third workshop on abusive language online.
20. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. [Bert: Pre-training of deep bidirectional transformers for language understanding](https://md.research.elamur.ai/free/en/papers/devlin-2019-bert-pre-training-deep-bidirectional/). In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies.
21. Flor Miriam Plaza-del Arco, M Dolores Molina-González, L Alfonso Urena-López, and M Teresa Martín-Valdivia. 2021. Comparing pre-trained language models for spanish hate speech detection. Expert Systems with Applications. [↗](https://doi.org/10.1016/j.eswa.2020.114120)
22. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. [↗](https://arxiv.org/abs/2503.19786)
23. Qwen Team. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671. [↗](https://arxiv.org/abs/2407.10671)
24. Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. 2017. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition. [find](https://scholar.google.com/scholar?q=Saining+Xie%2C+Ross+Girshick%2C+Piotr+Doll%C3%A1r%2C+Zhuowen+Tu%2C+and+Kaiming+He%2E+2017%2E+Aggregated+residual+transformations+for+deep+neural+networks%2E+In+Proceedings+of+the+IEEE+conference+on+computer+vision+and+p)
25. Wajdi Zaghouani, Md Rafiul Biswas, Mabrouka Bessghaier, Shimaa Ibrahim, Georgio Mikros, Abul Hasnat, and Firoj Alam. 2025. Overview of [mahed shared task: Multimodal detection of hope and hate emotions in arabic content](https://md.research.elamur.ai/free/en/papers/zaghouani-2025-mahed-shared-task-multimodal-detection/). In The Third Arabic Natural Language Processing Conference (ArabicNLP 2025). Association for Computational Linguistics.
26. Mohamed Zaytoon, Nagwa M El-Makky, and Marwan Torki. 2024. [Alexunlp-mz at araieval shared task: contrastive learning, llm features extraction and multi-objective optimization for arabic multi-modal meme propaganda detection](https://md.research.elamur.ai/free/en/papers/zaytoon-2024-alexunlp-mz-araieval-shared-task/). In Proceedings of The Second Arabic Natural Language Processing Conference.
27. Ziqi Zhang, David Robinson, and Jonathan Tepper. 2018. Detecting hate speech on twitter using a convolution-gru based deep neural network. In European semantic web conference. Springer. [find](https://scholar.google.com/scholar?q=Ziqi+Zhang%2C+David+Robinson%2C+and+Jonathan+Tepper%2E+2018%2E+Detecting+hate+speech+on+twitter+using+a+convolution%2Dgru+based+deep+neural+network%2E+In+European+semantic+web+conference%2E+Springer%2E)

*Updated 1 November 2026*