Task-adaptive debiasing with SCM for sentiment analysis

Sentiment analysis models often mirror social regularities found in data, which can lead to unfair or unreliable predictions. We address this challenge with a task adaptive debiasing method guided by the Stereotype Content Model, using the Warmth and Competence dimensions as signals. For each training example we compute SCM scores from a validated lexicon and set instance specific adversarial weights, so that examples with stronger stereotypical cues receive more debiasing pressure while neutral cases are largely preserved. We evaluate this approach on diverse text domains including product reviews, short movie phrases, and social media posts. Alongside standard accuracy, we monitor distributional parity using common fairness diagnostics such as the Demographic Parity gap and Kolmogorov–Smirnov $$p$$ values. The results show a clear pattern. When SCM cues behave as nuisance signals that do not carry the label, adaptive debiasing reduces disparities with little impact on accuracy. When those cues are closely tied to the label, debiasing tends to harm both accuracy and parity. Compared with fixed schedules, the adaptive scheme is more conservative and yields a better balance between fairness and accuracy when bias is separable, while limiting harm when bias and label meaning are tightly coupled. This study introduces an SCM guided, instance specific weighting scheme for adversarial debiasing, offers practical guidance on when to apply it, and provides evidence across multiple domains that supports careful, context aware use of debiasing in sentiment analysis.