Elamur Research⌕
☰
home / papers / 2026 / A Multi-Task Learning Framework for Modeling Engagement and Topic-Sensitive Responses in Arabic Women's Discourse

A Multi-Task Learning Framework for Modeling Engagement and Topic-Sensitive Responses in Arabic Women's Discourse

Paper · 2026 · Mabrouka Bessghaier, Md. Rafiul Biswas, Shimaa Ibrahim et al.

Summary

The paper predicts how Facebook audiences collectively respond to Arabic posts about women's issues. It treats this as different from classic sentiment analysis, because readers' reactions often do not match the tone the author wrote in. The authors assembled roughly 158k public Arabic Facebook posts from 2014 to 2024, gathered through CrowdTangle. The posts mix MSA with Egyptian, Levantine, Gulf and Maghrebi dialects. They fine-tune MARBERTv2 with three heads on a shared [CLS] vector:

  • a six-way classifier for the dominant reaction,
  • an 8-dimensional regressor for engagement volume (six reaction counts, comments, shares),
  • a binary detector for posts that drew no reactions at all.

The two auxiliary heads, together with focal loss, are meant to handle two problems: one reaction label dominates, and engagement counts are heavy-tailed. The model reaches a test macro-F1 of 72.4 and a weighted-F1 of 89.1. The work matters because it frames audience interpretation as its own prediction target. It also pairs categorical reactions with behavioural signals, a combination the authors say had not been tried for Arabic social media.

Key points

  • The corpus has about 158k Arabic Facebook posts (2014–2024), collected via the CrowdTangle API from public pages and groups. It covers women's legal rights, feminism debates, gender identity and economic empowerment. It is available on request only, and model weights are not released openly.
  • The label distribution is very skewed:
    • Love: 70.2% (108,338 posts)
    • Haha: 19.9% (30,745 posts)
    • Wow 1.3%, Sad 3.5%, Angry 3.3%, Care 1.8%
  • Regression targets (eight engagement variables) are log(1 + x)-transformed and min–max scaled to [0,1].
  • The backbone is MARBERTv2, chosen for its dialectal Arabic coverage. Task heads sit on the 768-dimensional [CLS] embedding after dropout, and focal loss addresses class imbalance.
  • Headline results: test macro-F1 72.4 and weighted-F1 89.1.

Interesting ideas

  • I1. Each post gets one discrete label (its most frequent reaction), but the full reaction vector is kept as a regression target, so posts that split opinion (e.g., mostly Love with a large Angry share) stay visible instead of being flattened.
  • I2. Behavioural side quantities that are cheap to observe (counts of comments and shares) serve as auxiliary supervision, on the hypothesis that they correlate with rare classes and sharpen their boundaries.
  • I3. A dedicated binary head first decides whether there is any signal at all (zero reactions), separating "silence" from the question of which reaction dominates.
  • I4. The paper sharply separates what the author expresses from how readers take it, illustrated by neutral or negative texts that draw approval and positive ones that provoke anger.
  • I5. Heavy-tailed count targets are tamed with log(1 + x) followed by scaling to [0,1] before joint training, so viral outliers do not dominate the loss.

For Mova

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

  • I1 (learn): when silver trees come from several witness parsers, train on the majority arc but also regress the witnesses' agreement distribution, so contested attachments carry an explicit disagreement signal rather than looking like clean gold.
  • I2 (frames): a rare-label classifier such as irony or a minority speech act could be co-trained with observable auxiliary targets (e.g., presence of negation, quotation, or exclamation counts), giving more gradient to the scarce class while staying explainable.
  • I3 (answer): the readers could get an explicit "the text gives no grounds" decision before choosing an answer, so unanswerable school-test or fable questions are refused rather than forced into a weak guess.
  • I4 (frames): pragmatic frames could hold two distinct slots, the speaker's stated attitude and the likely reader uptake, which would give implicature and irony detection a principled place to record the mismatch.
  • I5 (world): compiled measures and quantities extracted from texts (prices, populations, distances) could be stored with a log-scaled normalized companion value, making magnitude comparisons and learned scorers robust to extreme values.

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

🎓Northwestern University in QatarDoha · Qatar24 works · 148 citations

 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 2Research subjects and languages the work is about.Semitic languages · 2,908 worksDiscourse analysis · 466 works

Cite

Cite as Mabrouka Bessghaier, Md. Rafiul Biswas, Shimaa Ibrahim and Wajdi Zaghouani. 2026. A Multi-Task Learning Framework for Modeling Engagement and Topic-Sensitive Responses in Arabic Women's Discourse. Findings of the Association for Computational Linguistics: EACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-eacl.253

@inproceedings{bessghaier2026multi,
  title = {{A Multi-Task Learning Framework for Modeling Engagement and Topic-Sensitive Responses in Arabic Women's Discourse}},
  author = {Mabrouka Bessghaier, Md. Rafiul Biswas, Shimaa Ibrahim and Wajdi Zaghouani},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.253},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.253}
}

References

16 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, and El Moatez Billah Nagoudi. 2021. Arbert & marbert: Deep bidirectional transformers for arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7088– 7105, Online. Association for Computational Linguistics. ↗
  2. Mohammed Al-Badawi. 2024. English and arabic language trends in social media marketing: Analyzing linguistic patterns and their impact on audience engagement in zarqa, jordan market. In Frontiers of Human Centricity in the Artificial Intelligence-Driven Society 5.0, pages 969–974. Springer. ↗
  3. Noura AlShenaifi, Nora Al-Twairesh, and Hend Al-Khalifa. 2024a. Fine-tuning marbert for arabic stance detection. In Proceedings of ArabicNLP 2024, pages 856–861. find
  4. Noura AlShenaifi, Nora Al-Twairesh, and Hend Al-Khalifa. 2024b. Llm-in-the-loop active learning for arabic sentiment analysis. arXiv preprint arXiv:2509.23515. ↗
  5. Md Rafiul Biswas, Shimaa Ibrahim, Mabrouka Bessghaier, and Wajdi Zaghouani. 2025. Evaluation of pretrained and instruction-based pretrained models for emotion detection in arabic social media text. In Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing-Natural Language Processing in the Generative AI Era, pages 158–165.
  6. Anis Charfi, Mabrouka Ben-Sghaier, Andria Samy Raouf Atalla, Raghda Akasheh, Sara Al-Emadi, and Wajdi Zaghouani. 2024a. Marasta: A multi-dialectal arabic cross-domain stance corpus. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11060–11069.
  7. Anis Charfi, Mabrouka Besghaier, Raghda Akasheh, Andria Atalla, and Wajdi Zaghouani. 2024b. Hate speech detection with adhar: a multi-dialectal hate speech corpus in arabic. Frontiers in Artificial Intelligence, 7:1391472. ↗
  8. Rehab Duwairi and Iman Qarqaz. 2017. Collecting and processing arabic facebook comments for sentiment analysis. International Journal of Computer Applications, 170(8):1–6. find
  9. Lisa Graziani, Stefano Melacci, and Marco Gori. 2019. Jointly learning to detect emotions and predict facebook reactions. In International Conference on Artificial Neural Networks, pages 185–195. Springer. ↗
  10. Shimaa Amer Ibrahim, Mabrouka Bessghaier, and Wajdi Zaghouani. 2025. Ahasis shared task: Hybrid lexicon-augmented arabert model for sentiment detection in arabic dialects. In Proceedings of the Shared Task on Sentiment Analysis for Arabic Dialects, pages 29–34.
  11. Yunwoo Kim and Junhyuk Hwang. 2025. Predicting social media engagement from emotional and temporal features. arXiv preprint arXiv:2508.21650. ↗
  12. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988. ↗
  13. Flor Miriam Plaza-del Arco, Sercan Halat, Sebastian Padó, and Roman Klinger. 2021. Multi-task learning with sentiment, emotion, and target detection to recognize hate speech and offensive language. arXiv preprint arXiv:2109.10255. ↗
  14. Chris Pool and Malvina Nissim. 2016. Distant supervision for emotion detection using facebook reactions. In Proceedings of the Workshop on Computational Modeling of People’s Opinions, Personality, and Emotions in Social Media (PEOPLES), pages 30–39, Osaka, Japan. ↗
  15. The COLING 2016 Organizing Committee. find
  16. Wajdi Zaghouani, Md Rafiul Biswas, Mabrouka Bessghaier, Shimaa Ibrahim, George Mikros, Abul Hasnat, and Firoj Alam. 2025. Mahed shared task: Multimodal detection of hope and hate emotions in arabic content. In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 560–574.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bessghaier-2026-multi-task-learning-framework-modeling/