Our investigation delves into the intricate catalysts triggering toxicity within online discourse, revealing how seemingly innocuous comments can unexpectedly provoke hostile reactions. Given the profound influence of social media viewpoints on individuals, mitigating toxicity emerges as a critical imperative. To address this challenge, we present a sophisticated multi-label classification framework integrating TF-IDF and Word2Vec methodologies for robust vectorization. This framework amalgamates fundamental textual data with intricate metrics derived from prior research, facilitating nuanced monitoring of sentiment shifts, topic dynamics, and conversational context. Leveraging a diverse array of algorithms, including Logistic Regression, AdaBoost, Naive Bayes, Gradient Boosting, as well as Neural Network architectures like LSTM and Bi-LSTM, our model showcases exceptional efficacy in identifying four distinct types of toxicity: ‘toxic’, ‘obscene’, ‘insult’, and ‘non-toxic’. Importantly, our study underscores the necessity of accounting for contextual subtleties and sentiment fluctuations in online interactions, advocating for the widespread adoption of advanced natural language processing techniques to foster constructive discourse and enhance digital engagement. Furthermore, our research underscores the dynamic nature of online conversations, emphasizing the need for adaptable frameworks capable of capturing evolving patterns of toxicity.