Dealing with Annotator Disagreement in Hate Speech Classification (arXiv)

The identification of hate speech is an important issue, particularly on social media, where damaging information may spread rapidly. To lessen its effects and stop its spread, machine learning must be used to automatically detect and respond to hate speech. Finding a good training dataset is the first step in creating a successful hate speech detection algorithm. The majority of natural language processing jobs need labeled data, however classifying hate speech is challenging since hate speech is varied and sometimes subjective, which can result in conflicting judgments. This study looks at methods for dealing with annotator disagreement, a problem that is frequently disregarded. Specifically, using a refined BERT model, we assess several methods for handling annotator disagreement over the classification of hate speech in Turkish tweets.

https://arxiv.org/abs/2502.08266