Elamur Research⌕
☰
home / papers / 2026 / Ghavidel-Rajabi at MultiPRIDE: Identity, Toxicity, or Complexity? A Language-Specific Feature Selection Approach to Reclamatory Intent Detection

Ghavidel-Rajabi at MultiPRIDE: Identity, Toxicity, or Complexity? A Language-Specific Feature Selection Approach to Reclamatory Intent Detection

Paper · 2026 · Houman Rajabi, Fatemeh Ghavidel and Kourosh Ghahremani

Summary

The paper addresses MultiPRIDE @ EVALITA 2026 Task A. The task is to recognise when LGBTQ+ users take back slurs as self-empowering language, rather than using them to attack or neutrally. The authors argue that the cues signalling reclamation differ between languages, so they select features separately for English, Spanish and Italian. They extract a pool of 61 sociolinguistic features from unfiltered tweets using spaCy large pipelines. A Random Forest ranks these features for each language. The selected features pass through a small MLP and are fused with a monolingual BERT embedding in a dual-stream network.

Results vary sharply by language. Italian is the strongest, with Macro F1 of 0.8981, driven mainly by length and structural features. Spanish reaches 0.6960, relying on toxicity scores and hashtags. English reaches only 0.5460. The English model depends on first-person pronouns, so it confuses self-reclamation with reports of being targeted, and it almost always predicts the majority class.

The work matters for content moderation because it shows that reclamation cues are not shared across languages. It also suggests that English likely needs semantic role labeling.

Key points

  • Training data:
    • English: 1,026 samples, 88 positive (8.5%).
    • Spanish: 876 samples, 133 positive (15.1%).
    • Italian: 1,086 samples, 207 positive (19.0%).
    • Sources are the TWITA collection (Italian) and the LGBTQI+ Dataset 2020-2022 (Spanish), annotated under the FATA protocol.
  • Architecture:
    • Text encoders are the checkpoints bert-base-uncased, dccuchile/bert-base-spanish-wwm-uncased and dbmdz/bert-base-italian-uncased, each giving a 768-dimensional vector.
    • A 50-dimensional feature vector is projected by an MLP to 128 dimensions.
    • The two streams are concatenated (896 dimensions), fused down to 256, and passed to a linear classification head.
  • Top-ranked feature per language: first_person_ratio (English), hate_speech_score (Spanish) and text_length (Italian).
  • Class imbalance: oversampling during training raised the minority class to roughly one example per three majority examples.
  • English failure: recall was only 0.1017, and the model predicted "non-reclaimed" in about 90% of cases. The authors propose semantic role labeling and user metadata as fixes.

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 1Research subjects and languages the work is about.Bias, fairness & safety · 517 works

Cite

Cite as Houman Rajabi, Fatemeh Ghavidel and Kourosh Ghahremani. 2026. Ghavidel-Rajabi at MultiPRIDE: Identity, Toxicity, or Complexity? A Language-Specific Feature Selection Approach to Reclamatory Intent Detection. 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{rajabi2026ghavidel,
  title = {{Ghavidel-Rajabi at MultiPRIDE: Identity, Toxicity, or Complexity? A Language-Specific Feature Selection Approach to Reclamatory Intent Detection}},
  author = {Houman Rajabi, Fatemeh Ghavidel and Kourosh Ghahremani},
  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-1.23/}
}

References

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

  1. Z. Waseem, T. Davidson, D. Warmsley, I. Weber, Understanding abuse: A typology of abusive language detection subtasks, in: Proceedings of the First Workshop on Abusive Language Online, Association for Computational Linguistics, 2017, pp. 78–84.
  2. M. Dynel, The reappropriation of slurs: A corpus-based study of “slut” and “bitch” on reddit, Journal of Pragmatics 173 (2021) 162–177. find
  3. D. Hershcovich, S. Frank, H. Lent, M. de Lhoneux, M. Abdou, S. Brandl, E. Bugliarello, L. C. Piqueras, I. Chalkidis, R. Cui, et al., Challenges and strategies in cross-cultural nlp, in: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2…
  4. A. Lauscher, V. Ravishankar, I. Vulic, G. Glavas, From zero to hero: On the limitations of zero-shot cross-lingual transfer with multilingual transformers, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020, pp. 4491–4503. find
  5. D. Nozza, Exposing the limits of zero-shot cross-lingual hate speech detection, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), 2021, pp. 907–914.
  6. 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
  7. A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual representation learning at scale, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 8440–8451. find
  8. V. Basile, M. Lai, M. Sanguinetti, et al., Long-term social media data collection at the university of turin, in: Proceedings of the Fifth Italian Conference on Computational Linguistics (CLiC-it 2018), CEUR-WS, 2018, pp. 1–6. find
  9. J. Mata, E. Gualda, A dataset of Spanish tweets on people and communities LGBTQI+ during the COVID-19 pandemic 2020-2022, 2022. doi: 10.5281/zenodo.14878434 , zenodo Dataset. — [doi:10.5281/zenodo.14878434](https://doi.org/10.5281/zenodo.14878434) find
  10. E. Zsisku, A. Zubiaga, H. Dubossarsky, Hate speech detection and reclaimed language: Mitigating false positives and compounded discrimination, in: Proceedings of the 16th ACM Web Science Conference, ACM, 2024, pp. 241–249. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rajabi-2026-ghavidel-rajabi-multipride-identity-toxicity/