Classification of Toxicity in Social Media Comments Using the Binary Relevance – Logistic Regression and BERT Model

Social media provides a platform for people to get together and has transformed our lives in many ways. It allows people from all walks of life to unite and create a welcoming and positive network. Increased social media usage leads to an increased number of toxic comments. Exiting research stated that to classify and reduce the toxicity in comments in the past years with novel approaches. In this work, a comparison study is done to analyze the different vectorization and modeling techniques. The main purpose is to classify and categorize toxic comments from social media content. The data was analyzed and understood, using the Spacy emoji extension for SpaCy to extract emojis from the comments, and the emosent-py package was used to identify the sentiment score of the emoji. The resultant feature vector was concatenated with textual data. The work is compared different techniques for solving multi-label classification problems and has found the Binary Relevance – Logistic Regression model to be the most suitable for textual data. Binary Relevance, vectorization methods, and Chain Classifies are used to determine the best suitable technique to solve the multi-label classification problem. The text representation and the emoji feature vector were used as input to the Binary Relevance – Logistic Regression model, which returned 61% accuracy. After including the BERT model accuracy can be increased to 98.3%. However, the problem of understanding the context of the emoji was still present, as many potential false positives were identified. Since the context of emoji usage is uncertain, the prediction is more of a hit-or-miss.