Elamur Research⌕
☰
home / papers / 2026 / Xin1212 at NakbaVirality Shared Task: Frozen CLIP with Residual Adapter for Multimodal Virality Classification

Xin1212 at NakbaVirality Shared Task: Frozen CLIP with Residual Adapter for Multimodal Virality Classification

Paper · 2026 · Xinyan Zhang and Bingzhou Yang

Summary

This system paper addresses the NakbaVirality shared task. The task asks participants to predict how viral a social-media post will be (Low, Medium or High) from its text and attached image. The posts come from X and Reddit and are multilingual and noisy. The labelled data is small and class-imbalanced, so Macro-F1 is the main metric.

The authors keep a LAION CLIP encoder completely frozen and concatenate its text and image embeddings. The fused vector passes through a residual two-layer MLP adapter that learns a task-specific correction, and a small MLP head with dropout assigns the final label. Training uses class-weighted cross-entropy with label smoothing.

The authors tested a versioned series of models, from Baseline to V9, on a stratified development split. V9 performed best there with 0.5492 Macro-F1, and it scored 0.4559 F1 on the blind test set. The work shows that a lightweight design with a frozen backbone can improve minority-class detection in low-resource multimodal virality prediction, at some cost to plain accuracy.

Key points

  • Backbone: the encoder is laion/CLIP-ViT-B-32-laion2B-s34B-b79K, kept frozen.
    • Text and image embeddings are joined by late fusion (concatenation).
    • A residual adapter then refines the fused vector (h̃ = h + A(h)).
  • Data: 1,691 labelled instances were split 80/20 (stratified, seed 42), giving a 339-example development set. The official blind test set has 872 posts.
  • Development results (V9 vs. Baseline):
    • Macro-F1 rose from 0.4798 to 0.5492.
    • Virality-weighted F1 rose from 0.3924 to 0.5252.
    • Accuracy fell from 0.6401 to 0.5959.
  • Official test scores: F1 0.4559, accuracy 0.6089, precision 0.6000, recall 0.6089.
  • Training setup:
    • AdamW optimizer, batch size 16, 20 epochs, learning rate 1e-3.
    • Cosine schedule with warmup.
    • Label smoothing with ε = 0.1, plus class-weighted loss for the three classes.

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

🎓Dalian University of TechnologyDalian · China20 works · 46 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksComputer Vision and Pattern Recognition · 3,486 worksSocial Sciences · 1,954 worksSociology and Political Science · 626 works

🧭 Topics 1Research subjects and languages the work is about.Vision & language · 811 works

Cite

Cite as Xinyan Zhang and Bingzhou Yang. 2026. Xin1212 at NakbaVirality Shared Task: Frozen CLIP with Residual Adapter for Multimodal Virality Classification. 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 10.63317/3myy3h8imvfz

@inproceedings{zhang2026xin1212,
  title = {{Xin1212 at NakbaVirality Shared Task: Frozen CLIP with Residual Adapter for Multimodal Virality Classification}},
  author = {Xinyan Zhang and Bingzhou Yang},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/3myy3h8imvfz},
  url = {https://doi.org/10.63317/3myy3h8imvfz}
}

References

6 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mohammed Al-alshaqi, Danda B Rawat, and Chun-mei Liu. 2025. A bert-based multimodal framework for enhanced fake news detection using text and image data fusion. Computers, 14(6):237. find
  2. Despoina Antonakaki and Sotiris Ioannidis. 2025. Cross-platform digital discourse analysis of the israel-hamas conflict: Sentiment, topics, and event dynamics. arXiv preprint arXiv:2601.02367. — [arXiv:2601.02367](https://arxiv.org/abs/2601.02367) find
  3. Sedat Dogan, Nina Dethlefs, and Debarati Chakraborty. 2025. Early multimodal prediction of cross-lingual meme virality on reddit: A time-window analysis. arXiv preprint arXiv:2510.05761. — [arXiv:2510.05761](https://arxiv.org/abs/2510.05761) find
  4. Saad Ezzini, Salima Lamsiyah, Shadi Abudalfa, Samir El-Amrany, and Walid Alsafadi. 2026. The nakbavirality shared task on multimodalvirality prediction in high-stakes discourse. In Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources (Nakba-NLP 2026), co-located wi… find
  5. Guangyu Mu, Ying Chen, Xiurong Li, Li Dai, and Jiaxiu Dai. 2025. Semantic enhancement and cross-modal interaction fusion for sentiment analysis in social media. PLOS ONE, 20(4):e0321011. find
  6. D Vamsidhar, Parth Desai, Aniket K Shahade, Shruti Patil, and Priyanka V Deshmukh. 2025. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-xin1212-nakbavirality-shared-task-frozen/