Enhancing Social Media Safety with Automated Toxicity Classification using Semantic Topic Models

In social media platforms, there is a growing interest in developing automated solutions for toxicity classification using machine learning algorithms. This paper presents a novel approach to automated toxicity classification using semantic topic models, achieving high accuracy and transparency. The report emphasizes the importance of detecting and adjusting for biases to ensure fair categorization. Moreover, the report highlights the challenges of natural language processing tasks, importance of understanding and addressing potential biases in machine learning models. Datasets are produced by the Origami project at Google, labeled by human labelers with adjectives such as "toxic," "very poisonous," "accusation," "danger," "obscene," and "identity hatred." From the analysis, it shows that proposed method achieved better results on toxicity classification in terms of accuracy (0.85), precision (0.87), recall (0.84) and f1-score (0.85) respectively. For non-toxicity classification, the proposed method achieves accuracy (0.92), precision (0.89), recall (0.88), f1-score (0.88) and AUC-ROC (0.96) respectively.