Elamur Research⌕
☰
home / papers / 2025 / MultiMinds at MAHED 2025: Multimodal and Multitask Approaches for Detecting Emotional, Hate, and Offensive Speech in Arabic Content

MultiMinds at MAHED 2025: Multimodal and Multitask Approaches for Detecting Emotional, Hate, and Offensive Speech in Arabic Content

Paper · 2025 · Riddhiman Debnath, Abdul Wadud Shakib and Md Saiful Islam

Summary

The MultiMinds team competed in two subtasks of MAHED 2025, a shared task hosted by ArabicNLP 2025 that addresses emotion, hate and offensive content in Arabic. Task 2 is multitask text classification: each post receives an emotion label (12 classes) plus offensive and hate flags. The team compared several systems on noisy, imbalanced data, which they first cleaned and orthographically normalized:

  • an SVM baseline on TF-IDF features,
  • XGBoost on TF-IDF or on fused AraBERT and XLM-RoBERTa [CLS] embeddings, tuned with Optuna,
  • a fine-tuned AraBERTv2 with three task heads,
  • gpt-oss-20b adapted with LoRA.

The multitask AraBERT gave the best average macro F1, about 0.68. Task 3 is hateful meme detection. Here a lightweight ELU-Net encoder–decoder processes the image, and cross-attention injects DistilBERT-tokenized caption features into its middle layers. With class weighting this reached roughly 75% accuracy. The team placed 10th in Task 2 and 7th in Task 3. The paper's main value is practical: it shows that class weighting and embedding fusion help, and that a small multimodal model is workable. It also shows plainly how severe imbalance can hide very weak minority-class performance.

Key points

  • Task 2 data skew. Offensive posts number 1744 vs 4216 non-offensive, and hateful posts 303 vs 1441. Emotions range from Anger (1551) down to Fear (53). The split was a stratified 70/15/15.

  • Task 2 average macro F1:

    SystemAvg macro F1
    SVM + TF-IDF baseline≈0.51
    Optuna-tuned, class-weighted XGBoost on fused AraBERT+XLM-RoBERTa0.574
    Multitask AraBERTv20.684 (emotion only 0.267)
    DistilBERT0.683
    gpt-oss-20b (post-competition)0.300 (emotion 0.014)

  • Preprocessing. The pipeline strips diacritics, repeated characters, Latin words and digits. It also maps the variant forms of Alif onto a single standard Alif (U+0627).

  • Task 3 data. The meme set has 398 hateful vs 2663 non-hateful items. The baseline used CLIP ViT-B/32 features, first with an SVM and then with a 1D-CNN.

  • Best meme model. ELU-Net + DistilBERT with class weights scored accuracy 0.754 and macro F1 0.500. However, F1 on the hateful class was only 0.165, against 0.858 on the non-hateful class. Training used about 6.2 GB of GPU memory on Colab, and early stopping halted at epoch 5.

Interesting ideas

  • I1. One shared encoder feeding three small task-specific linear heads (emotion, offensive, hate) let related labels share a single representation, and it outperformed separate feature-based classifiers.
  • I2. Headline accuracy (75%) hid a hateful-class F1 of only 0.165, which shows that minority-class and macro scores must always be reported next to accuracy.
  • I3. Folding visually or phonetically interchangeable character variants into one canonical form before tokenization reduces spurious vocabulary splits in noisy social text.
  • I4. Hyperparameters were searched against a class-weighted objective that directly targets macro F1, rather than raw accuracy, which matters whenever labels are heavily skewed.
  • I5. Concatenating the [CLS] vectors of two different pretrained encoders gave a richer input than either encoder alone for a cheap downstream classifier.

For Mova

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

  • I1 (frames): one shared contextual encoder could feed separate light heads for speech act, irony and implicature cues, keeping each decision separately inspectable.
  • I2 (learn): evaluation reports should break scores down per rare category, such as rare UD relations or rare frame roles, so that majority-class wins cannot mask failures.
  • I3 (tokens): the en tokenizer could canonicalize curly vs straight apostrophes and quotes, dash variants and similar Unicode look-alikes before lemmatization, logging every mapping so it stays explainable.
  • I4 (learn): when training the perceptron or gparse on silver data, model selection could optimize macro scores over relation labels, keeping rare dependencies from being sacrificed.
  • I5 (tree): features from two different parser encoders could be fused for an auxiliary classifier, such as a dialect or label-confidence detector, without retraining either encoder.

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

🎓Shahjalal University of Science and TechnologySylhet · Bangladesh2 works · 1 citation

Also on the byline: Athabasca University · Metropolitan University.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 3Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 worksVision & language · 657 works

Cite

Cite as Riddhiman Debnath, Abdul Wadud Shakib and Md Saiful Islam. 2025. MultiMinds at MAHED 2025: Multimodal and Multitask Approaches for Detecting Emotional, Hate, and Offensive Speech in Arabic Content. 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.93

@inproceedings{debnath2025multiminds,
  title = {{MultiMinds at MAHED 2025: Multimodal and Multitask Approaches for Detecting Emotional, Hate, and Offensive Speech in Arabic Content}},
  author = {Riddhiman Debnath, Abdul Wadud Shakib and Md Saiful Islam},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.93},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.93}
}

References

31 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Malak Abdullah, Mirsad Hadzikadicy, and Samira Shaikhz. 2018. Sedat: sentiment and emotion detection in arabic text using cnn-lstm deep learning. In 2018 17th IEEE international conference on machine learning and applications (ICMLA), pages 835–840. IEEE. find
  2. Lubna Al-Henaki, Hend Al-Khalifa, Abdulmalik Al-Salman, Hajar Alqubayshi, Hind Al-Twailay, Gheeda Alghamdi, and Hawra Aljasim. 2025. Multiprose: A multi-label arabic dataset for propaganda, sentiment, and emotion detection. In International Conference on Applications of Natural Language to Information Systems, pages 156–172. Springer.
  3. 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. In International Conference on Web Information Systems Engineering, pages 380– 390. Springer. Firoj Alam, Abul Hasnat, Fatema Ahmad, Md.
  4. Arid Hasan, and Maram Hasanain. 2024b. ArMeme: Propagandistic content in Arabic memes. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21071–21090, Miami, Florida, USA. Association for Computational Linguistics.
  5. Nouar AlDahoul and Yasir Zaki. 2025. Detecting hope, hate, and emotion in arabic textual speech and multimodal memes using large language models. arXiv preprint arXiv:2508.15810. ↗
  6. Wassen Aldjanabi, Abdelghani Dahou, Mohammed AA Al-Qaness, Mohamed Abd Elaziz, Ahmed Mohamed Helmi, and Robertas Damaševičius. 2021. Arabic offensive and hate speech detection using a cross-corpora multi-task learning model. In Informatics, volume 8, page 69. MDPI. ↗
  7. Safa Alsafari, Samira Sadaoui, and Malek Mouhoub. 2020. Hate and offensive speech detection on arabic social media. Online Social Networks and Media, 19:100096. ↗
  8. Kholoud Alsmearat, Mohammed Shehab, Mahmoud Al-Ayyoub, Riyad Al-Shalabi, and Ghassan Kanaan. 2015. Emotion analysis of arabic articles and its impact on identifying the author’s gender. In 2015 IEEE /ACS 12th International Conference of Computer Systems and Applications (AICCSA), pages 1–6. IEEE. find
  9. Nourah Alswaidan and Mohamed El Bachir Menai. 2020. Hybrid feature model for emotion recognition in arabic text. IEEE Access, 8:37843–37854. ↗
  10. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. Arabert: Transformer-based model for arabic language understanding. arXiv preprint arXiv:2003.00104. ↗
  11. 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, 12:22359–22375. Md. ↗
  12. Rafiul Biswas and Wajdi Zaghouani. 2025a. An annotated corpus of arabic tweets for hate speech analysis. CoRR, abs/2505.11969. Md. ↗
  13. Rafiul Biswas and Wajdi Zaghouani. 2025b. Emohopespeech: An annotated dataset of emotions and hope speech in english and arabic. CoRR, abs/2505.11959. ↗
  14. Hamdy Mubarak, Sabit Hassan, and Shammur Absar Chowdhury. 2023. Emojis as anchors to detect arabic offensive language and hate speech. Natural Language Engineering, 29(6):1436–1457. ↗
  15. Anusha Chhabra and Dinesh Kumar Vishwakarma. 2023. A literature survey on multimodal and multilingual automatic hate speech identification. Multimedia Systems, 29(3):1203–1230. ↗
  16. Andrew Perrin. 2015. Social media usage. Pew research center, 125:52–68. find
  17. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116. ↗
  18. Michael Han Daniel Han and Unsloth team. 2023. Unsloth. find
  19. Yunjiao Deng, Yulei Hou, Jiangtao Yan, and Daxing Zeng. 2022. Elu-net: An efficient and lightweight u-net for medical image segmentation. IEEE Access, 10:35932–35941. ↗
  20. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805. ↗
  21. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR.
  22. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. ↗
  23. Kuljeet Singh, Amit Mahajan, and Vibhakar Mansotra. 2021. 1d-cnn based model for classification and analysis of network attacks. International Journal of Advanced Computer Science and Applications, 12(11):604–613. ↗
  24. Reza Drikvandi and Olamide Lawal. 2023. Sparse principal component analysis for natural language processing. Annals of data science, 10(1):25–41. find
  25. Wajdi Zaghouani, Md Rafiul Biswas, Mabrouka Bessghaier, Shimaa Ibrahim, Georgio Mikros, and Firoj Alam. 2025. Overview on mahed 2025 shared task: Multimodal detection of hope and hate emotions in arabic content. In The Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou. Association for Computational Linguistics. find
  26. Ahmed El-Sayed and Omar Nasr. 2024. Aast-nlp at multimodal hate speech event detection 2024: A multimodal approach for classification of text-embedded images based on clip and bert-based models. In Proceedings of the 7th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE 2024), pages 139–144. Wajdi Zaghouani, Hamdy Mubarak, and Md. find
  27. 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.
  28. Paula Fortuna and Sérgio Nunes. 2018. A survey on automatic detection of hate speech in text. Acm Computing Surveys (Csur), 51(4):1–30. ↗
  29. Amir Jalilifard, Vinicius Fernandes Caridá, Alex Fernandes Mansano, Rogers S Cristo, and Felipe Pen-horate Carvalho da Fonseca. 2021. Semantic sensitive tf-idf to determine word relevance in documents. In Advances in Computing and Network Communications: Proceedings of CoCoNet 2020, Volume 2, pages 327–337. Springer. ↗
  30. Mohamed Bayan Kmainasi, Abul Hasnat, Md Arid Hasan, Ali Ezzat Shahroor, and Firoj Alam. 2025. Memeintel: Explainable detection of propagandistic and hateful memes. arXiv preprint arXiv:2502.16612. ↗
  31. Hongchan Li, Yantong Lu, and Haodong Zhu. 2024. Multi-modal sentiment analysis based on image and text fusion based on cross-attention mechanism. Electronics, 13(11):2069. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/debnath-2025-multiminds-mahed-2025-multimodal-multitask/