Individuazione degli stereotipi negli adolescenti: sviluppo di un corpus annotato ed esperimenti di classificazione con LLMs

This thesis explores the application of advanced machine learning techniques to detect and analyze stereotypes in textual data, with a focus on the adolescent demographic. Utilizing the newly created STERHEOSCHOOL corpus, along with the established StereoHoax and FB-Stereotypes corpora, this work aims to develop, refine, and compare methods for automated stereotype detection. The research begins with the comprehensive construction and annotation of the STERHEOSCHOOL corpus, designed specifically to capture the nuances of stereotype expression among teenagers in educational settings. This corpus is annotated with multiple layers, including stereotype categorization, stance identification, and forms of discredit, employing a robust annotation schema inspired by current computational linguistics and social sciences research. A neural classification system based on the GilBERTo model is developed and meticulously trained on the STERHEOSCHOOL dataset. Initial classification efforts focus solely on this corpus to benchmark its capabilities and establish a foundational understanding of the model’s effectiveness in stereotype detection. Subsequent phases extend the classification to incorporate the StereoHoax and FB-Stereotypes datasets, enabling a comprehensive analysis across different contexts. This work demonstrates the feasibility and efficacy of using sophisticated machine learning models for social linguistics analysis, providing a scalable approach to addressing bias and stereotypes in digital communication. The results highlight the potential of machine learning in understanding and mitigating stereotypes, thus contributing to more informed and inclusive content strategies.