A Comparison of Word Embeddings for Comment Toxicity Detection: Detection Power of Computer

The rapid expansion of online communities and social media platforms, the issue of toxic comments has become increasingly prevalent. Toxic comments not only contribute to a negative user experience but also have the potential to propagate hate speech, harassment, and cyberbullies. Therefore, there is a pressing need for robust and efficient comment toxicity detection systems that can automatically identify and flag toxic comments. The proposed methodology involves preprocessing the data-set by removing punctuation, applying stemming and lemmatization techniques to standardize the text. The data is then transformed into a matrix representation using tokenization, padding, and word embedding techniques. In the word embedding phase, three popular algorithms, namely Word2Vec, GloVe, and FastText, are compared to determine the most suitable approach for capturing the semantic meaning of words. The architectural design comprises several layers, encompassing an embedding layer, bidirectional LSTM layer, dropout layers, and dense layers. The bidirectional LSTM layer plays a crucial role in assimilating contextual information from both preceding and subsequent words within the comment. Training involves utilizing the training set, followed by evaluation. This project addresses the crucial need for comment toxicity detection and provides an effective methodology for identifying toxic comments in online platforms. By automatically detecting and flagging toxic comments, this system can contribute to fostering a healthier and more respectful online environment. Index Terms—comment toxicity detection, social media plat-forms, toxic comments, hate speech, harassment, cyberbullies, stemming, lemmatization, word embedding, Word2Vec, GloVe, FastText, bidirectional LSTM.