Linguistic Markers of Deception in Insurance Claim Narratives: A Forensic Linguistic and Computational Analysis of Fraudulent and Genuine Claims

Abstract

Detecting deception in written narratives remains a major challenge in forensic linguistics and fraud investigation. While computational methods have improved automated text classification, identifying reliable linguistic indicators of deceptive communication continues to require theoretically informed analysis. This study investigates linguistic markers of deception in fraudulent and genuine insurance claim narratives by integrating forensic linguistic theory with computational text analysis. The research draws upon Vrij et al.'s (2000) Reality Monitoring framework, Hancock et al.'s (2008) deception detection model, and corpus-based discourse analysis proposed by Coulthard and Johnson (2007). A corpus of 300 insurance claim statements obtained from a United Kingdom insurance company was analyzed, comprising 150 verified genuine claims and 150 claims retrospectively identified as fraudulent. The narratives were examined for a range of linguistic features, including self-disclosure rate, contextual embedding, narrative reconstruction, cognitive operations involving spatial, temporal, and affective references, hedging frequency, and lexical diversity. These linguistic variables were subsequently used to train and evaluate a supervised machine learning classifier for deception detection. Quantitative results demonstrated that the classifier achieved an overall accuracy of 81% on held-out test data, indicating that linguistic features provide substantial predictive value for distinguishing deceptive from truthful narratives. Fraudulent claims were characterized by significantly lower levels of self-disclosure, reduced contextual embedding, and fewer instances of coherent narrative reconstruction, while exhibiting higher frequencies of hedging expressions and greater linguistic uncertainty. Genuine claims, in contrast, contained richer contextual detail, more temporally organized event descriptions, and stronger experiential grounding consistent with reality monitoring theory. The findings support the view that deception is reflected in systematic linguistic and discourse-level patterns rather than isolated lexical cues. This study contributes to forensic linguistics by demonstrating the effectiveness of combining discourse analysis with computational modeling for deception detection. It also offers practical implications for insurance fraud investigation, suggesting that linguistically informed automated screening systems can complement traditional investigative procedures while highlighting the need for cautious interpretation of probabilistic linguistic evidence.