Abstract
This study challenges the conventional "gold standard" in sexism detection by incorporating annotator attitudes—specifically Right-Wing Authoritarianism (RWA) and Social Dominance Orientation (SDO)—into classification models. By moving beyond simple demographics and utilizing psychological profiling, the researchers demonstrate that Large Language Models (LLMs) can achieve up to a 31% performance improvement in predicting individual labels. The findings underscore the importance of perspectivism in subjective Natural Language Processing (NLP) tasks like identifying gender-based violence (GBV).
Introduction
In traditional NLP, researchers typically aggregate multiple annotator labels into a single "gold standard" ground truth. However, for subjective tasks like identifying GBV, sexism, or misogyny, this aggregation often erases the diverse perspectives of marginalized groups and ignores the ideological biases of the annotators themselves. While previous studies attempted to use demographic data (age, gender, ethnicity) to predict annotator behavior, demographics are often poor predictors of individual judgment. This research addresses the gap by examining social attitudes—psychological constructs that shape perceptions of social hierarchies—and exploring how these attitudes can be modeled to improve automated content moderation.
Methodology
The researchers adopted a "strongly perspectivist" approach, preserving individual labels rather than collapsing them into a majority vote. The methodology involved three main pillars:
- Psychological Profiling: 43 annotators were recruited and administered three validated Social Psychology surveys: Very Short Authoritarianism (VSA), Short Social Dominance Orientation (SSDO), and the Brief Hostile Neosexism Scale (BHNS).
- Statistical Modeling: A mixed-effects regression model quantified the relationship between these survey scores and the likelihood of an annotator labeling a text as "sexist."
- LLM Prompting & Fine-tuning: Several LLMs (FLAN-T5, Llama 2/3, and Mistral) were tested using various prompting strategies, including "annotator-infused" prompts that included the psychological profile of the target annotator.
Datasets
The study re-annotated two prominent English-language datasets: EDOS (Explainable Detection of Sexism) and DOM (Detection of Online Misogyny). The authors collected up to 23 labels per item for 1,600 instances, creating a multi-perspective dataset that pairs text with detailed attitudinal profiles of the people who rated it.
Analysis of Annotator Attitudes
The statistical analysis revealed significant correlations between psychological traits and labeling behavior, confirming that attitudes are more robust predictors than demographics alone:
- The RWA Paradox: Surprisingly, annotators with higher VSA (Right-Wing Authoritarianism) scores were more likely to label text as sexist. This may be linked to "benevolent sexism," where individuals seek to protect women who adhere to traditional roles.
- Internalized Leniency: Conversely, those with high SSDO and BHNS (Neosexism) scores were significantly less likely to label items as sexist, likely viewing offensive text as natural or inevitable due to their belief in social hierarchies.
Experimental Results
- Performance Gains: In-Context Learning (ICL) using annotator information improved performance by 17% to 26% over a majority-class baseline. Fine-tuning provided an even greater boost of 31%.
- Model Sensitivity: LLMs proved highly sensitive to the "persona" provided in prompts; they successfully shifted classification boundaries to match the predicted bias of the described annotator.
- Prompt Structure: Providing a well-structured, brief summary of attitudes (e.g., "This worker is moderate RWA...") was more effective than providing raw survey data, which tended to degrade performance.
Conclusion and Future Work
While the study demonstrates the power of attitudinal profiling, it is limited by a relatively small sample of 43 annotators and a focus on English-language internet forum data. Future research will expand into multilingual contexts and explore complex intersectional identities to further refine the taxonomy of gender-based violence in automated systems.
Summary updated