Enhancing Misogyny Detection in Bilingual Texts Using FastText and Explainable AI

Abstract

Gendered disinformation not only harms women's rights and democratic norms, but it also threatens national security by worsening societal divisions through authoritarian regimes' intentional weaponization of social media. Despite the severity of the situation, efforts to engage digital platforms for improved protection against gendered disinformation are frequently ignored, showing the difficult task of countering online misogyny in the face of commercial interests. This study addresses the challenge of detecting misogynous content in bilingual (En-glish and Italian) online communications. Leveraging FastText word embeddings and explainable AI techniques, we propose a model that enhances interpretability and accuracy in identifying misogynous speech. Our approach integrates language-specific transformer-based models with machine learning algorithms optimized through hyperparameter tuning and regularization. The model's effectiveness is demonstrated by robust performance metrics, achieving an F1-score of up to 0.95. The effectiveness of our methodology is demonstrated through robust performance metrics. With the application of LIME, we also provide insights into model predictions, facilitating a better understanding of model decisions.