Integrating Transformers and Hybrid Machine Learning Models for Automated Sexism Detection in Online Discourse

The detection of sexism in online discourse presents a formidable challenge, as implicit biases and subtle linguistic nuances often evade traditional approaches. Addressing these complexities, a hybrid framework is proposed, seamlessly integrating traditional machine learning techniques with advanced deep learning architectures, particularly Bidirectional Encoder Representations from Transformers (BERT). Through domain adaptation and weak supervision, this framework overcomes limitations in scalability and manual annotation, enabling robust detection across diverse platforms. Key techniques, including Term Frequency-Inverse Document Frequency (TF-IDF) vectorization, sentiment analysis, and adaptive hyperparameter optimization, drive effective feature extraction. Evaluated on a dataset of over 10,000 annotated texts, the framework demonstrates superior performance compared to conventional methods and Recurrent Neural Networks (RNNs), achieving significant gains in detecting microaggressions and benevolent sexism. Beyond advancing the state-of-the-art in sexism detection, this work establishes a scalable, adaptable foundation for broader content moderation systems, addressing ethical challenges in digital interactions. These results underscore the power of self-attention mechanisms in capturing linguistic subtleties and offer a pathway toward scalable, adaptable, and ethically sound content moderation solutions.