Elamur Research⌕
☰
home / papers / 2026 / HCMUS_TheFangs at NakbaVirality Shared Task: The Audience is the Message: Escaping the Deep Learning Trap in Conflict-Domain Virality Prediction

HCMUS_TheFangs at NakbaVirality Shared Task: The Audience is the Message: Escaping the Deep Learning Trap in Conflict-Domain Virality Prediction

Paper · 2026 · Dao Sy, Duy Minh

Summary

The paper describes the first-place system in the Nakba-NLP 2026 NakbaVirality Shared Task. The task is to sort Gaza-conflict social media posts into three ordinal virality levels (Low, Medium, High) using the post text, an image and source metadata. The authors first tried multimodal deep learning with XLM-RoBERTa, CLIP and DINOv2, and these models did very poorly. A t-SNE view of CLIP embeddings showed that conflict imagery looks alike across all virality classes, a pattern they call the Visual Homogeneity Problem. The metadata told a different story: the community a post came from predicted virality far better than its content.

The final system is simple. It uses one Bayesian-smoothed target encoding of the source community, plus TF-IDF unigrams and a few structural counts, fed to XGBoost. It reached a test Macro-F1 of 0.7062, more than 20 points ahead of the next team. A large tuned feature model did worse. The authors conclude that in small, sociologically driven datasets, knowing the audience beats complex models of content. They also report that heavy hashtag use tends to reduce virality.

Key points

  • Data: the NakbaVirality dataset has 1,691 labelled training posts.
    • Each post has a caption, an image and community metadata.
    • About one third of the captions are not in English.
    • The class split is roughly 40% Low, 50% Medium and 10% High, so Macro-F1 is the evaluation metric.
  • Winning system (Exp 32):
    • Community target encoding with smoothing m = 10, computed inside 5-fold stratified cross-validation to avoid leakage.
    • 1,000 TF-IDF unigrams plus character, word and hashtag counts, giving 1,003 features in total.
    • XGBoost with 1,000 trees, learning rate 0.05 and depth 6.
    • Result: test Macro-F1 of 0.7062, Rank #1. The runner-up scored 0.4983.
  • Deep learning failures:
    • XLM-R + CLIP cross-attention fusion: 0.2935.
    • BLIP-2 caption features: 0.2866.
    • Best neural model (CLIP ViT-B/32 + linear head): 0.5230, below TF-IDF + XGBoost at 0.5328.
  • Other community-based models:
    • A GNN community embedding scored 0.6463.
    • "The Singularity," a 219-feature model tuned with Optuna, scored only 0.6806.
  • Community effect and hashtags:
    • High-virality rates differ about tenfold between communities.
    • Hashtag density is negatively correlated with virality (r = −0.153).
    • The authors release their code publicly.

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

🎓Ho Chi Minh City University of ScienceHo Chi Minh City · Vietnam5 works · 2 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSociology and Political Science · 577 works

Cite

Cite as Dao Sy, Duy Minh. 2026. HCMUS_TheFangs at NakbaVirality Shared Task: The Audience is the Message: Escaping the Deep Learning Trap in Conflict-Domain Virality Prediction. Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3bdi84ncsp38

@inproceedings{minh2026hcmus,
  title = {{HCMUS\_TheFangs at NakbaVirality Shared Task: The Audience is the Message: Escaping the Deep Learning Trap in Conflict-Domain Virality Prediction}},
  author = {Dao Sy, Duy Minh},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/3bdi84ncsp38},
  url = {https://doi.org/10.63317/3bdi84ncsp38}
}

References

15 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Annual Meeting of the Association for Computational Linguistics, pages 4040–4054. find
  2. Takuya Akiba, Shotaro Sano, Toshihiro Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 2623–2631.
  3. Firoj Alam, Ferda Ofli, and Muhammad Imran. 2018. CrisisMMD: Multimodal twitter datasets from natural disasters. In Proceedings of the 12th International AAAI Conference on Web and Social Media, volume 12. find
  4. Robert D. Benford and David A. Snow. 2000. Framing processes and social movements: An overview and assessment. Annual Review of Sociology, 26:611–639. find
  5. Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794. find
  6. Justin Cheng, Lada Adamic, P. Alex Dow, Jon Michael Kleinberg, and Jure Leskovec. 2014. Cascades in temporal networks. In Proceedings of the 23rd International Conference on World Wide Web, pages 913–924. find
  7. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… find
  8. Giovanni Da San Martino, Alberto Barrón-Cedeño, Henning Wachsmuth, Rostislav Petrov, and Preslav Nakov. 2020. SemEval-2020 task 11: Detection of propaganda techniques in news articles. In Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1377– 1414.
  9. Michela Del Vicario, Alessandro Bessi, Fabiana Zollo, Fabio Petroni, Antonio Scala, Guido Caldarelli, H. Eugene Stanley, and Walter Quattrociocchi. 2016. The spreading of misinformation online. Proceedings of the National Academy of Sciences, 113(3):554–559. find
  10. Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. 2020. GoEmotions: A dataset of fine-grained emotions. In Proceedings of the 58th Saad Ezzini, Salima Lamsiyah, Shadi Abudalfa, Samir El-Amrany, and Walid Alsafadi. 2026. Nakbavirality: Multimodal viral… find
  11. Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742. PMLR.
  12. Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193.
  13. Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. 2018. CatBoost: Unbiased boosting with categorical features. Advances in Neural Information Processing Systems, 31. find
  14. Nguyen Lam Phu Quy, Pham Phu Hoa, Tran Chi Nguyen, Dao Sy Duy Minh, Nguyen Hoang Minh Ngoc, and Huynh Trung Kiet. 2025. Beyond vision: Contextually enriched image captioning with multi-modal retrieval. In Proceedings of the 3rd Workshop on Event-Centered Information Access (EVENTA), ACM Internationa… — [arXiv:2512.20042](https://arxiv.org/abs/2512.20042) find
  15. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learn…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sy-2026-hcmus-thefangs-nakbavirality-shared-task/