Hidden in plain sight: detecting misogyny beneath ambiguities and implicit bias in language

This thesis explores NLP methods for detecting misogyny in social media, ranging from explicit instances to implicit and ambiguous expressions that vary across languages and platforms. With a focus on Italian and English, the research investigates monolingual, cross-lingual, and multimodal approaches, leveraging transformer-based models and large language models (LLMs). It critically examines the limitations of these models in misogyny detection, particularly in handling unintended biases related to identity terms, the ambiguity of pejorative language, and the implicit nature of harmful discourse. A key contribution of this work is the conceptualization of pejorative epithet disambiguation as a necessary step for misogyny detection, framed as a word sense disambiguation task. To support this, the thesis introduces PejorativITy, a newly developed corpus for pejorative epithets in Italian. Beyond explicit misogyny, this research delves into the complexities of implicit misogyny detection and explanation, investigating how LLMs can help uncover the underlying assumptions embedded in misogynistic language. To facilitate these experiments, the thesis introduces ImplicIT-Mis, the first dataset specifically designed for implicit misogyny in Italian. The study evaluates LLMs’ ability to recognize and reconstruct implied meanings in misogynistic statements, which often require nuanced comprehension of social cues, stereotypes, irony, and backhanded compliments—elements that challenge traditional classification methods reliant on explicit hate speech markers. A unique aspect of this research is the application of argumentation theory, which helps decompose inferential processes behind misogynistic language. By incorporating reasoning-based detection tasks, the experiments reveal both the strengths and limitations of LLMs in capturing hidden social dynamics. The findings emphasize the potential of NLP models in identifying misogyny but highlight ongoing challenges in developing context-aware, multilingual models capable of adapting to the ever-evolving landscape of online discourse.