A Multi-Model Consensus approach for Narcissistic Personality Detection in social media

Narcissistic traits are increasingly expressed through online behavior, yet traditional assessment methods often struggle to capture these traits in naturalistic digital contexts. To address this gap, a novel computational framework ATTUNE is introduced for detecting narcissistic personality traits from Twitter data. The framework integrates attention-based language models, zero-shot learning, and psycholinguistic feature extraction (e.g., via Empath) to classify tweets into different categories of narcissism. A hybrid labeling strategy was employed, combining rule-based heuristics, fuzzy matching, and multi-model consensus to enhance scalability and contextual sensitivity. This approach employed data preprocessing, hybrid annotation, and feature-enriched classification, and was able to achieve a Label Completeness score of 64.84% and Model Consistency of 86.84%. These results surpassed earlier LIWC-based approaches, which reported 59% completeness. By reducing reliance on self-report measures and leveraging naturalistic language use, this work offers a scalable and context-aware method for computational assessment of personality traits in digital environments