How Does the Representation of an Annotator’s Moral Value Profile Influence an LLM’s Offensiveness Judgements?

Abstract

Offensive-language detection is often evaluated as a single-label classification task, although research on subjective annotation shows that people can make systematically different judgements about the same text. This thesis focuses on predicting annotator-specific offensiveness judgements rather than establishing an objective offensiveness label. Moral values are examined as one possible personalization signal because they provide an individual-level description of normative concerns that may be relevant when people interpret potential violations of social and moral norms. The study asks whether different representations of the same measured moral-value profile change an LLM’s approximation of an annotator’s observed binary judgement.

The experiments use D3CODE, which combines repeated individual offensiveness annotations with annotator-level scores on six moral foundations. Five deterministic renderings of each profile—natural language, categorical, numeric, ordinal, and top-k—are compared with a no-profile, unpersonalized reference. Qwen2.5-72B-Instruct is used as the primary model; Qwen2.5-7B-Instruct and Llama-3.3-70B-Instruct provide within-family and cross-family robustness comparisons. The models are evaluated in zero-shot prompting and in three same-annotator few-shot settings with k = 3 examples selected randomly, by semantic retrieval, or by diversity. Evaluation combines Accuracy, Macro-F1, paired McNemar comparisons, and an instance-level cross-representation decision-disagreement diagnostic.

The zero-shot condition exhibits the largest spread in aggregate performance across profile representations and cross-representation disagreement above 20% for each selected model. Same-annotator few-shot context is associated with lower disagreement, although profile representations continue to change individual predictions. Retrieval-based prompting produces the strongest Macro-F1 across the selected models and the lowest disagreement in most sub-categories for the primary model. No representation performs consistently best across every tested model, setting, and metric. Instead, the results show that profile representation and demonstration selection are consequential parts of the experimental treatment and should be reported, controlled, and evaluated in personalized, value-conditioned judgement prediction.