Predicting IoT Privacy Concerns From Tweets Through BERTweet-Based Fusion of Psycholinguistic Big Five and Risk Features: Toward Integration With IoT Security Frameworks

The rapid expansion of the Internet of Things (IoT) into everyday environments has heightened concerns about how personal data is perceived, shared, and protected. While much research on IoT privacy has concentrated on technical safeguards, individuals’ attitudes toward privacy are equally shaped by psychological and linguistic factors. This study investigates how privacy concerns expressed in social media discourse can be predicted by integrating psycholinguistic patterns with personality and risk-related features. We introduce a deep learning framework, BERTweet, that fuses textual representations with auxiliary features derived from the Big Five personality traits and general risk propensity. Our dataset comprises 85,112 tweets (a consensus subset where GPT-4.1, Grok, and DeepSeek all agreed), drawn from five publicly available corpora. Privacy annotations were generated through a multi-step process that began with zero-shot classification and was finalised via majority voting across the three large language models. The proposed fusion architecture dynamically balances contributions from linguistic, personality, and risk dimensions. Evaluation on a held-out test set of 8,512 tweets shows that the BERTweet-Large fusion model achieved 96.8% accuracy, with strong F1 performance for sensitive (low-privacy, high-disclosure) tweets (97.6%) and non-sensitive (high-privacy, low-disclosure) tweets (95.2%). We benchmark this model against BERT, RoBERTa, and DeBERTa backbones under identical training conditions to substantiate the choice of BERTweet for tweet-level privacy prediction. Unlike approaches that rely on explicit metadata or direct tagging, our method reveals the implicit ways privacy concerns emerge in everyday language use. The findings highlight the value of combining psycholinguistic, Big Five, and risk features in deep learning systems, offering a scalable approach to building privacy-aware models that address the complexities of IoT and social media ecosystems. The study therefore contributes a psychologically grounded, empirically validated framework that links tweet-level language, inferred traits, and risk propensity to IoT privacy concern, and provides a transparent benchmark for future privacy-aware language models.