Linguistic Features and Framing Construction in Identity - Impersonation Frauds: A Natural Language Processing - Based Study on Telecommunication Network Fraud

The research assumes that fraudulent messages exhibit distinct linguistic style and content-based cues that enable detection. Using TextMind and machine learning (TF-IDF, LDA) combined with grounded theory coding, we identified 10 high-frequency linguistic categories (e.g., function words, pronouns, cognitive mechanisms) and 7 content themes (e.g., authority impersonation, time pressure, benefit inducement).

The dual-dimensional framework achieved 82.1% detection accuracy (AUC = 0.86), outperforming single-dimension models. Each message includes linguistic frequency scores, topic ratings, and victim outcome (compliance vs. non-compliance). Researchers can use the data for predictive modeling, cross-method validation, or replication of fraud detection analyses.