Elamur Research⌕
☰
home / papers / 2026 / I2C At MultiPRIDE: Transformers And Generative LLMs For Multilingual Reclamation Classification

I2C At MultiPRIDE: Transformers And Generative LLMs For Multilingual Reclamation Classification

Paper · 2026 · Laura Vázquez-Ramos, Jacinto Mata-Vázquez and Victoria Pachón-Álvarez

Summary

This paper describes the I2C team's systems for MultiPRIDE 2026 Task 1 at EVALITA 2026. The task asks whether an LGBTQ+-related term in a social media post is used with reclamatory (in-group, positive) intent or not. The authors worked in the text-only setting for Italian, Spanish and English. The organisers' datasets are small and skewed toward the non-reclamatory class, so the team balanced each training split with contextual word embedding augmentation (bert-base-uncased). They also built a merged multilingual pool.

Two paradigms were compared, one per allowed run:

  • Run 1: a generative LLM tuned with QLoRA instructions.
  • Run 2: a discriminative Transformer classifier based on XLM-RoBERTa-base.

The Transformer gave the best results in every language. The best macro-F1 scores were highest for Italian and lowest for English. An error analysis traced most failures to quotation, mention-versus-use confusion, irony and politically charged posts. This shows how hard intent-sensitive moderation remains without conversational context.

Key points

  • Best official test macro-F1: 0.843 (IT), 0.711 (ES), 0.571 (EN). XLM-RoBERTa-base outperformed the LLM run across all subtasks.
  • Data and class imbalance:
    • IT: 1,086 instances, 19.06% reclamatory.
    • ES: 876 instances, 15.18% reclamatory.
    • EN: 1,026 instances, 8.58% reclamatory.
    • Merged pool: 2,988 instances, treated as an unconstrained setup.
  • Balancing: CWE-based augmentation was applied only to training data. It reached 1,500 examples per class (3,000 total) after trimming the majority class. Untouched 20% held-out test splits were kept.
  • LLM run: QLoRA fine-tuning of Qwen2.5-0.5B, with teknium/OpenHermes-2.5-Mistral-7B used for Spanish. The best hyperparameters included a learning rate of 1e-4 and a linear scheduler.
  • Proposed future work:
    • Targeted augmentation for irony, activism and quoted mentions.
    • Explicit stance and mention-detection steps.
    • Use of thread or user context where allowed.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksMultilingual & cross-lingual NLP · 828 works

Cite

Cite as Laura Vázquez-Ramos, Jacinto Mata-Vázquez and Victoria Pachón-Álvarez. 2026. I2C At MultiPRIDE: Transformers And Generative LLMs For Multilingual Reclamation Classification. Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{vazquezramos2026i2c,
  title = {{I2C At MultiPRIDE: Transformers And Generative LLMs For Multilingual Reclamation Classification}},
  author = {Laura Vázquez-Ramos, Jacinto Mata-Vázquez and Victoria Pachón-Álvarez},
  booktitle = {Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026)},
  year = {2026},
  url = {https://aclanthology.org/2026.evalita-1.26/}
}

References

15 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. M.-C. Schreuder, Safe spaces, agency, and resistance: A metasynthesis of lgbtq language use, Journal of LGBT Youth 18 (2021) 256–272. find
  2. C. Ferrando, L. Draetta, M. Madeddu, M. Sosto, V. Patti, P. Rosso, C. Bosco, J. Mata, E. Gualda, Multipride at evalita 2026: Overview of the multilingual automatic detection of slur reclamation in the lgbtq+ context task, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processin… find
  3. E. Sezerer, S. Tekir, A survey on neural word embeddings, 2021. URL: https://arxiv.org/abs/2110. 01804. doi: 10.48550/arXiv.2110.01804 . arXiv:2110.01804 . — [doi:10.48550/arxiv.2110.01804](https://doi.org/10.48550/arxiv.2110.01804) ↗
  4. J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2019), Associa…
  5. A. Zubiaga, Natural language processing in the era of large language models, 2024. find
  6. S. Islam, H. Elmekki, A. Elsebai, J. Bentahar, N. Drawel, G. Rjoub, W. Pedrycz, A comprehensive survey on applications of transformers for deep learning tasks, Expert Systems with Applications 241 (2024) 122666. find
  7. T. Dettmers, A. Pagnoni, A. Holtzman, L. Zettlemoyer, Qlora: Efficient finetuning of quantized llms, Advances in neural information processing systems 36 (2023) 10088–10115.
  8. Q. Team, et al., Qwen2 technical report, arXiv preprint arXiv:2407.10671 2 (2024). find
  9. M. Vangeli, Large language models as advanced data preprocessors: Transforming unstructured text into fine-tuning datasets, 2024. find
  10. X. Ou, H. Li, Ynu@ dravidian-codemix-fire2020: Xlm-roberta for multi-language sentiment analysis., in: FIRE (Working Notes), 2020, pp. 560–565. find
  11. Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019).
  12. M. Grandini, E. Bagli, G. Visani, Metrics for multi-class classification: an overview, arXiv preprint arXiv:2008.05756 (2020). doi: 10.48550/arXiv.2008.05756 . — [doi:10.48550/arxiv.2008.05756](https://doi.org/10.48550/arxiv.2008.05756) find
  13. J. V. Carter, J. Pan, S. N. Rai, S. Galandiuk, Roc-ing along: Evaluation and interpretation of receiver operating characteristic curves, Surgery 159 (2016) 1638–1645. doi: 10.1016/j.surg.2015.12. 029 . — [doi:10.1016/j.surg.2015.12](https://doi.org/10.1016/j.surg.2015.12) find
  14. D. Chicco, G. Jurman, The advantages of the matthews correlation coefficient (mcc) over f1 score and accuracy in binary classification evaluation, BMC genomics 21 (2020) 6. doi: 10.1186/ s12864-019-6413-7 . find
  15. A. P. Bradley, The use of the area under the roc curve in the evaluation of machine learning algorithms, Pattern recognition 30 (1997) 1145–1159. doi: 10.1016/S0031-3203(96)00142-2 . — [doi:10.1016/s0031-3203(96)00142-2](https://doi.org/10.1016/s0031-3203%2896%2900142-2) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ramos-2026-i2c-multipride-transformers-generative-llms/