Detecting Conversational Toxicity Using Neural Networks

Abstract

Along with providing a platform for users to express their views freely and connect to the masses, it also becomes a place for hateful behavior and cyber bullying. This negativity sometimes becomes the cause to stop people from expressing their point of view. Issues like this happen quite often, and the modulators have limited capabilities to deal with it manually. Even though much research is going on to determine whether a post or comment is toxic or not, social media platforms are still looking for efficient and faster ways to detect and remove toxic content. Automating this will not only help in identifying toxic material, but it will also eventually promote discussions online at the same time maintaining user safety. The paper addresses this issue with the aid of Wikipedia’s talk page edits dataset, found on Kaggle, which is used to train the deep learning model, which ultimately classifies the data into six categories: toxic, severe toxic, obscene, threat, insult, and identity hate. The deep learning technique used here is Long Short-Term Memory (LSTM) with the GloVe word embeddings.