p24 / Conference
DUTH at EXIST 2025: Multilingual Sexism Detection with Soft Labels and Transformers.
CLEF 2025 Working Notes, CEUR Workshop Proceedings 4038, 1793-1800 · 2025
In plain language
The DUTH system detects multilingual sexist content while accounting for disagreement between annotators.
Research context
Preserving annotation disagreement helps examine socially sensitive language tasks beyond a single fixed label.
Approach
A multilingual transformer is fine-tuned with soft labels, oversampling, and class weighting across binary detection, sexism type, and intent tasks.
Findings
Binary detection is competitive, especially on Spanish data. More nuanced type and intent classification remains harder.
Summary based on abstract.