Comparing Different Transformer Models’ Performance for Identifying Toxic Language Online

Abstract: There is a growing use of the internet and alongside that, there has been an increase in the use of toxic language towards other people that can be harmful to those that it targets. The usefulness of artificial intelligence has exploded in recent years with the development of natural language processing, especially with the use of transformers. One of the first ones was BERT, and that has spawned many variations including ones that aim to be more lightweight than the original ones. The goal of this project was to train three different kinds of transformer models, RoBERTa, ALBERT, and DistilBERT, and find out which one was best at identifying toxic language online. The models were trained on a handful of existing datasets that had labelled data as abusive, hateful, harassing, and other kinds of toxic language. These datasets were combined to create a dataset that was used to train and test all of the models. When tested against data collected in the datasets, there was very little difference in the overall performance of the models. The biggest difference was how long it took to train them with ALBERT taking approximately 2 hours, RoBERTa, around 1 hour and DistilBERT just over half an hour. To understand how well the models worked in a real-world scenario, the models were evaluated by labelling text as toxic or non-toxic on three different subreddits. Here, a larger difference in performance showed up. DistilBERT labelled significantly fewer instances as toxic compared to the other models. A sample of the classified data was manually annotated, and it showed that the RoBERTa and DistilBERT models still performed similarly to each other. A second evaluation was done on the data from Reddit and a threshold of 80% certainty was required for the classification to be considered toxic. This led to an average of 28% of instances being classified as toxic by RoBERTa, whereas ALBERT and DistilBERT classified an average of 14% and 11% as toxic respectively. When the results from the RoBERTa and DistilBERT models were manually annotated, a significant improvement could be seen in the performance of the models. This led to the conclusion that the DistilBERT model was the most suitable model for training and classifying toxic language of the lightweight models tested in this work.