AIWizards at MULTIPRIDE

A Hierarchical Approach to Slur Reclamation Detection

Luca Tedeschini, Matteo Fasulo
MULTIPRIDE at EVALITA 2026
Paper arXiv Code Task

Abstract

This project addresses MULTIPRIDE, a binary classification task focused on detecting whether LGBTQ+ related terms are used with reclamatory intent. The system targets cases where potential slurs or derogatory terms are used non-discriminatorily, for example as self-identification or as an expression of community belonging.

Our contribution focuses on Task B, where user bios are available alongside tweet text. We propose a dual encoder architecture that combines a custom user encoder and text encoder through gated feature fusion, producing a representation for final reclamation classification.

Method

The architecture combines contextual signals from user bios with the tweet text that contains the potentially reclaimed term. Both encoders produce hidden representations that are weighted through a gating mechanism before final binary classification.

User Encoder

Encodes profile information from user bios to capture contextual signals that may help distinguish reclamatory usage from discriminatory usage.

Text Encoder

Models tweet content using task labels from the official MULTIPRIDE data, focusing on the language surrounding potentially reclaimed terms.

Gated Fusion

Combines user and text hidden states through a gating mechanism before the final binary classification layer.

Dual encoder model architecture for MULTIPRIDE Task B

Task Setting

MULTIPRIDE includes a text-only Task A and a contextual Task B where user bios are available in addition to tweet text. This work focuses on Task B for Italian and Spanish, using bio context to enrich the representation available to the classifier.

Because the official dataset is not publicly released by the organizers, the repository includes a mock dataset so the code can run and the pipeline can be inspected.

Ethical Note

The user encoder is intended only to improve contextual modeling for the shared task. It is not designed or endorsed as a tool for identifying, inferring, or exposing a user's LGBTQ+ identity.

BibTeX

@inproceedings{tedeschini2026aiwizardsmultipride,
  title     = {AIWizards at MULTIPRIDE: A Hierarchical Approach to Slur Reclamation Detection},
  author    = {Tedeschini, Luca and Fasulo, Matteo},
  booktitle = {Proceedings of the 9th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026)},
  series    = {CEUR Workshop Proceedings},
  volume    = {4195},
  address   = {Bari, Italy},
  publisher = {CEUR-WS.org},
  year      = {2026},
  url       = {https://ceur-ws.org/Vol-4195/20.pdf}
}