Learning Moral Diversity: Modelling Individual Perspectives in Moral Classification of Texts

Traditional Natural Language Processing (NLP) approaches to moral value classification typically aggregate multiple annotators' labels into a single "ground truth." This study challenges this practice, arguing that moral judgment is inherently subjective and that disagreement between annotators is a meaningful signal rather than mere noise. The researchers investigate whether explicitly modeling individual annotator perspectives can improve classification accuracy and provide deeper insights into moral judgment.

The authors extend a pretrained BERT model by adding an "Annotator Layer" designed to capture individual labeling patterns. They test two variants: a Bias-only (BO-AL) variant, which learns a simple bias matrix for each annotator, and a Linear Transformation (LT-AL) variant, which uses more complex linear layers to represent annotator-specific behavior. The models were trained on the Moral Foundations Twitter Corpus (MFTC), which contains tweets labeled by 23 different annotators. The researchers evaluated performance by comparing their models against a standard fine-tuned BERT baseline that ignores annotator identity.

Modeling individual perspectives consistently outperformed the standard aggregation approach. The Linear Transformation variant (LT-AL) achieved a 10.2% improvement in macro F1 scores across all five moral foundations compared to the baseline. The results suggest that annotator biases are learnable features that, when accounted for, allow the model to better navigate the subjectivity of moral classification. Furthermore, the authors demonstrate that models trained on aggregated labels may mask significant inconsistencies and provide a misleading impression of performance.

This research highlights the limitations of the "ground truth" paradigm in subjective NLP tasks. By treating disagreement as a source of information rather than noise, the authors provide a framework for building more nuanced models that respect the diversity of human opinion. This approach is particularly relevant for social media analysis, where the ambiguity of short text and the diversity of user backgrounds make consensus difficult to achieve.