SafeWebUH at SemEval-2023 Task 11: Learning Annotator Disagreement in Derogatory Text: Comparison of Direct Training vs Aggregation
This work addresses the challenge of handling subjectivity in annotation for derogatory text detection, which is important for improving content moderation systems, but it is incremental as it builds on existing BERT methods and focuses on a specific task.
The paper tackled the problem of modeling annotator disagreement in derogatory text detection by fine-tuning a BERT model on SemEval-2023 datasets, finding that individual annotator modeling and aggregation reduced the Cross-Entropy score by an average of 0.21 compared to direct training on soft labels, with annotator metadata contributing an additional average reduction of 0.029.
Subjectivity and difference of opinion are key social phenomena, and it is crucial to take these into account in the annotation and detection process of derogatory textual content. In this paper, we use four datasets provided by SemEval-2023 Task 11 and fine-tune a BERT model to capture the disagreement in the annotation. We find individual annotator modeling and aggregation lowers the Cross-Entropy score by an average of 0.21, compared to the direct training on the soft labels. Our findings further demonstrate that annotator metadata contributes to the average 0.029 reduction in the Cross-Entropy score.