The Automatic Detection of Abusive Language in Dota 2 Chat Messages

Abstract

This study addresses the pervasive issue of abusive language in online video game communication channels, focusing on Dota 2 chat messages. The aim was to employ diverse traditional machine learning algorithms and advanced deep learning architectures to identify and classify toxic and abusive language effectively. Leveraging TF-IDF, GloVe word embeddings, and self-trained embeddings, the research compared various classical machine learning models such as Naïve Bayes, Logistic Regression, and Support Vector Machine with convolutional and recurrent neural network models. The results revealed a consistent trend where deep learning models, particularly those employing GRUs and LSTMs, outperformed classical machine learning models. Experiments also demonstrated that self-trained embeddings generally outperformed GloVe embeddings in the domain of online video game chat messages.