The conceptual ambiguity among terms like ’hate speech’, ’toxic speech’, and ’dangerous speech’ creates a significant bottleneck for both research and automated moderation. Traditional NLP models, often focused on lexical cues, struggle to differentiate these nuanced forms of linguistic violence, especially when the harm is implicit. This paper addresses this gap with a twofold objective. First, we conduct a conceptual review and propose a unified ontology that differentiates these concepts—including verbal aggression and cyberbullying—based on their core attributes, such as their target, intent, and associated rhetorical hallmarks. Second, we propose a computational methodology designed to operationalize this ontology. Our framework uses a multi-stage NLP pipeline that leverages semantic analysis, specifically Semantic Role Labeling and Named Entity Recognition, to deconstruct speech acts into their core components (e.g., target and action). This component-based approach allows for a granular classification that can robustly distinguish between seemingly similar phenomena, such as a general insult and a targeted identity-based attack. This methodology is particularly promising for low-resource languages, such as Portuguese, as it relies on core semantic tasks for which multilingual models are available, rather than requiring massive, task-specific labeled datasets.