Human perspectives in textual classification models

Traditional text classification frameworks typically rely on the “Gold Standard” paradigm, which assumes a single ground truth for every observation. However, in sensitive domains such as the detection of racism, misogyny, or hate speech, the interpretation of a text is inherently mediated by individual sensitivity, cultural background, and social perception. In these contexts, imposing a unique label may lead to the loss of relevant socio-linguistic information, as definitions of “offensiveness” often vary despite existing institutional guidelines. This research explores classification methods that incorporate the viewpoints of multiple annotators to better capture the complexity of contested linguistic phenomena. While the ultimate goal is to integrate individual annotations into a unified multi-perspective classification—through multidimensional aggregation strategies that preserve the richness of annotator diversity, such as combining individual probability estimates into distributional summaries or composite scores—the present study focuses on the modeling of individual annotators. Specifically, we investigate how each annotator’s labeling behavior can be captured through dedicated classification models that account not only for the overall judgment but also for the specific textual portions each annotator considers relevant to their annotation. The proposed methodology was tested on a corpus of social media comments labeled by a diverse pool of participants. Preliminary results suggest that modeling individual annotator perspectives—including the text segments they deem relevant—constitutes a necessary foundation for any subsequent multidimensional aggregation strategy, and provides a more ethical framework for analyzing sensitive content. This approach advocates for a shift in textual statistics toward “pluralistic labeling,” acknowledging that for many social issues, the statistical truth lies in the distribution of interpretations rather than in a single consensus.