PejorativITy: Disambiguating Pejorative Epithets to Improve Misogyny Detection in Italian Tweets
Arianna Muti, Federico Ruggeri, Cagri Toraman, Lorenzo Musetti, Samuel Algherini, Silvia Ronchi, Gianmarco Saretto, Caterina Zapparoli, Alberto Barrón-Cedeño
TL;DR
The paper tackles misogyny detection in Italian tweets where pejorative epithets can shift meaning by context. It introduces PejorativITy, a 1,200-tweet corpus annotated for word-level pejorativity and sentence-level misogyny, and a two-pronged approach that injects disambiguation information into misogyny classifiers via concatenation or substitution. Fine-tuning a BERT-based model for pejorative disambiguation and then using that output to enrich misogyny detection yields consistent improvements on the authors’ corpus and on AMI benchmarks, with larger gains when gold annotations are available. Contextual embeddings analysis and open-source LLM prompts further illuminate how models encode pejorativity and where current models fall short, highlighting both the promise and the limitations of word-sense disambiguation and prompting for this task. The work contributes a new resources, a practical pipeline for integrating word sense information into misogyny tasks, and empirical evidence that disambiguation can reduce false positives in misogyny detection, potentially informing safer and more accurate social media moderation tools.
Abstract
Misogyny is often expressed through figurative language. Some neutral words can assume a negative connotation when functioning as pejorative epithets. Disambiguating the meaning of such terms might help the detection of misogyny. In order to address such task, we present PejorativITy, a novel corpus of 1,200 manually annotated Italian tweets for pejorative language at the word level and misogyny at the sentence level. We evaluate the impact of injecting information about disambiguated words into a model targeting misogyny detection. In particular, we explore two different approaches for injection: concatenation of pejorative information and substitution of ambiguous words with univocal terms. Our experimental results, both on our corpus and on two popular benchmarks on Italian tweets, show that both approaches lead to a major classification improvement, indicating that word sense disambiguation is a promising preliminary step for misogyny detection. Furthermore, we investigate LLMs' understanding of pejorative epithets by means of contextual word embeddings analysis and prompting.