Gender violence (GV) in cyberspace has developed in a complex and evolving manner. This, in turn, has been transferred to digital services such as social networks, where perpetrators, anonymously take advantage of the most vulnerable people through harassment and the use of abusive communications. This growing cyber threat creates a gap in the study, based on its digital trail, which allows us to understand the linguistic communication pattern based on computational analysis to develop comprehensive actions that mitigate its impact. This problem can be addressed from several angles such as legal frameworks, social awareness mechanisms, and technological solutions. The latter has developed mechanisms, such as content filtering, which has now improved with AI. These have proven effective in detecting and blocking harmful content, including abusive language and threats, on digital platforms. The purpose of their development is to improve security in cyberspace by automatically identifying and eliminating inappropriate material. This research proposes to improve classification in content filtering by developing an AI-powered classification model with a refined dataset to address online gender-based violence, given that abuse is not only based on expressions with insulting connotations but also on discriminatory expressions. Using topic modeling, key themes were identified, and negative sentiment subcategories were created. This data was processed with a Large Language Model (LLM), specifically RoBERTa, which was fine-tuned to classify texts into high, medium, and low aggression levels. The results demonstrate the high accuracy of the proposed model in distinguishing between these aggression levels, underlining its effectiveness in mitigating online Gender-Based Violence (GBV).