Elamur Research⌕
☰
home / papers / 2026 / LlaNa at MultiPRIDE: A Fine-Tuning Approach with Cross-Lingual Augmentation

LlaNa at MultiPRIDE: A Fine-Tuning Approach with Cross-Lingual Augmentation

Paper · 2026 · Alessio Mercurio, Giorgio Talluto, Irene Siragusa et al.

Summary

The paper addresses a hard case for hate-speech classifiers: distinguishing reclaimed LGBTQ+ slurs, where community members use a derogatory word to express identity or pride, from slurs used as insults. The work was done for the MultiPRIDE shared task at EVALITA 2026. The LlaNa team took part in the Italian subtasks A1 (tweet text only) and B1 (tweet plus the author's user bio). They identified the shortage of positive reclamation examples as the main obstacle and chose a data-centric fix. Reclamation-labelled tweets from the English and Spanish MultiPRIDE sets were machine-translated into Italian with OPUS-MT. The authors preferred OPUS-MT over LLMs because it was less likely to soften or drop slurs. One XLM-RoBERTa-base model with a compact classification head and a class-weighted loss served both subtasks. Augmentation raised internal validation macro F1 in both settings. On the official test set, the system ranked first in B1 and fifth in A1. The results suggest that low-cost cross-lingual augmentation, combined with author-profile context, can help classifiers separate reclamation from abuse when data is scarce.

Key points

  • Data: The Italian training set has 1,086 tweets, only 207 positive (reclamation) versus 879 negative, with an 80/20 internal split for development. A1 augmentation drew on English and Spanish positives; B1 augmentation used only the Spanish B2 data.
  • Model and training: XLM-RoBERTa-base with 256-token inputs feeds the CLS embedding to a 512-unit dense layer, normalization, ReLU and dropout 0.3, then a binary output. Training used AdamW (lr 1e-5, weight decay 0.05), weighted BCE, batch size 8, up to 8 epochs with early stopping, on a single RTX 3060.
  • Validation gains from augmentation (macro F1): A1 rose from 0.8817 to 0.9131 (+0.0314); B1 rose from 0.9033 to 0.9196 (+0.0163).
  • Official test results: B1 reached 0.9021, ranking 1st against a baseline of 0.8981. A1 reached 0.8835, ranking 5th against a baseline of 0.8731. In A1 the positive class showed precision of about 0.86 but recall of only about 0.76.
  • Error analysis: Misclassifications clustered in tweets mixing conflicting sentiments, such as struggle followed by pride. They also occurred in tweets that merely quote or report someone else's use of a slur, which the model wrongly labelled as reclamation.

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 · 18,614 worksArtificial Intelligence · 17,965 works

🧭 Topics 1Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 882 works

Cite

Cite as Alessio Mercurio, Giorgio Talluto, Irene Siragusa and Roberto Pirrone. 2026. LlaNa at MultiPRIDE: A Fine-Tuning Approach with Cross-Lingual Augmentation. 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{mercurio2026llana,
  title = {{LlaNa at MultiPRIDE: A Fine-Tuning Approach with Cross-Lingual Augmentation}},
  author = {Alessio Mercurio, Giorgio Talluto, Irene Siragusa and Roberto Pirrone},
  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-1.29/}
}

References

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

  1. 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
  2. A. Galinsky, C. Wang, J. Whitson, E. Anicich, K. Hugenberg, G. Bodenhausen, The reappropriation of stigmatizing labels: The reciprocal relationship between power and self-labeling, Psychological science 24 (2013). doi: 10.1177/0956797613482943 . — [doi:10.1177/0956797613482943](https://doi.org/10.1177/0956797613482943) find
  3. C. Bosco, F. Dell’Orletta, F. Poletto, M. Sanguinetti, M. Tesconi, Overview of the EVALITA 2018 Hate Speech Detection Task, in: Proceedings of the 6th evaluation campaign of Natural Language Processing and Speech tools for Italian (EVALITA 2018), CEUR-WS.org, 2018.
  4. D. Nozza, E. Fersini, V. Patti, F. Rangel, P. Rosso, HODI at EVALITA 2023: Overview of the Homotransphobia Detection Task, in: Proceedings of the Eighth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian (EVALITA 2023), CEUR-WS.org, Parma, Italy, 2023. find
  5. F. Cutugno, A. Miaschi, A. P. Aprosio, G. Rambelli, L. Siciliani, M. A. Stranisci, Evalita 2026: Overview of the 9th evaluation campaign of natural language processing and speech tools for italian, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for I… find
  6. A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, et al., Unsupervised cross-lingual representation learning at scale, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 8440–8451.
  7. T. Ranasinghe, M. Zampieri, Multilingual offensive language identification with cross-lingual embeddings, 2020. URL: https://arxiv.org/abs/2010.05324. arXiv:2010.05324 . ↗
  8. J. Tiedemann, M. Aulamo, D. Bakshandaeva, M. Boggia, S.-A. Grönroos, T. Nieminen, A. Raganato, Y. Scherrer, R. Vazquez, S. Virpioja, Democratizing neural machine translation with OPUS-MT, Language Resources and Evaluation (2023) 713–755. doi: 10.1007/s10579-023-09704-w . — [doi:10.1007/s10579-023-09704-w](https://doi.org/10.1007/s10579-023-09704-w) find
  9. J. Tiedemann, S. Thottingal, OPUS-MT — Building open translation services for the World, in: Proceedings of the 22nd Annual Conferenec of the European Association for Machine Translation (EAMT), Lisbon, Portugal, 2020. find
  10. T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, et al., Huggingface’s transformers: State-of-the-art natural language processing, 2020. URL: https://arxiv.org/abs/1910.03771. arXiv:1910.03771 . ↗
  11. I. Loshchilov, F. Hutter, Decoupled weight decay regularization, 2019. URL: https://arxiv.org/abs/ 1711.05101. arXiv:1711.05101 . ↗
  12. A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, et al., Pytorch: An imperative style, high-performance deep learning library, 2019. URL: https://arxiv.org/abs/1912.01703. arXiv:1912.01703 . ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mercurio-2026-llana-multipride-fine-tuning-approach/