Enhancing misogyny detection through context-aware semantic enrichment

Detecting online misogyny remains a critical challenge, particularly when abusive language is implicit or context-dependent. Traditional keyword-based models often fail to capture these subtleties. This research investigates the impact of sentence-level semantic enrichment on misogyny detection performance across various model types. The input text was enriched through three methods—MiniLM, FastText, and ConceptNet—each generating semantically meaningful paraphrases. These enriched inputs are evaluated using traditional classifiers, deep learning, and transformer models. The results show that semantic enrichment, especially with ConceptNet, significantly boosts detection accuracy across all different models. ELECTRA combined with ConceptNet achieved the highest performance, with an F1 score of 92% and an accuracy of 97%, while the BiLSTM models achieved a strong balance of efficiency and precision. Our findings highlight the value of semantically enriched representations in improving the detection of nuanced misogynistic language on online platforms.