Investigating annotator bias in comment quality and incivility classification by formal education

Machine learning models used in algorithmic content moderation in online discussions play an increasingly important role in detecting high-quality (e.g., constructive or engaging) and uncivil (e.g., toxic, hateful, or abusive) comments. Trained on human decisions, these models risk reproducing annotator bias, particularly when the perception of such nuanced categories is influenced by social backgrounds and training data lacks sufficient representational diversity. This study investigates the potential impact of annotators' formal education on the classification of high-quality and uncivil user comments in online discussions. Using a dataset of 13,677 German user comments annotated by crowd workers with low, medium, and high formal education, we investigate divergences in classification outcomes when classifiers are trained on data annotated by different educational groups. In addition, we assessed statistical associations between fine-grained comment features and the annotations provided by each subgroup. Our results indicate that classifiers for high-quality and uncivil comments yield significantly different results when trained on data labeled by annotators with varying educational backgrounds. Furthermore, comment features used in traditional operationalizations of deliberative quality (e.g., solution proposals, additional knowledge) and incivility (e.g., vulgarity, accusation of lying) are more strongly associated with annotations by crowd annotators with medium and high formal education. We argue that such group-specific differences should be considered when developing machine learning models in both practice and research to reflect a more inclusive understanding of valuable and norm-violating user contributions to online discussions.