A psycholinguistic NLP framework for forensic text analysis of deception and emotion

Psycholinguistics is an interdisciplinary area of research that bridges elements of linguistics with various branches of psychology. One of its goals is to identify and explain the links that exist between our psyche and the language we speak. In this research, we are expanding upon previous research that we did using several different Natural Language Processing (NLP) techniques to identify persons of interest from a scenario that was generated by a large language model (LLM). We used a different approach to this topic, which allowed us to develop a more nuanced method of reverse engineering and breaking down the psycholinguistic features of each suspect. Through the application of n-grams paired with deception, emotion, and subjectivity over time, we were able to identify and measure cues that can be used to better identify persons of interest from a larger pool of candidates. That dataset was smaller and somewhat limited in scope. We successfully identified the guilty parties from the fictional murder case using a combination of Latent Dirichlet Allocation, word vectors, and pairwise correlations. This research was larger in scope, number of potential suspects, and in the diversity of the corpus used. We were able to determine the guilty parties identified in ground truth using our methodology in this case specifically by focusing on entity to topic correlation, deception detection, and emotion analysis.

The study analyzed 18 LLM-generated fictional police interviews using four psycholinguistic variables: deception over time (Empath library), anger/fear/neutrality/subjectivity levels (transformer models), entity-to-keyword correlation with suspicion phrases, and Natural Language Inference to detect contradictory narratives. Researchers tracked temporal patterns using n-grams and composite emotosubjectivity graphs.

Background & Context

The intersection of psycholinguistics and artificial intelligence represents a growing frontier in deception detection research. Psycholinguistics bridges linguistics with psychology to identify links between our psyche and language, offering forensic investigators potential tools for identifying persons of interest through textual analysis.

This 2025 study by Adkins, Al Bataineh, and Khanal advances an emerging methodology that applies Natural Language Processing to forensic text analysis. Building on previous research with only 10 artificially created characters in a fictional murder scenario, the researchers expanded their investigation to test whether psycholinguistic NLP techniques could scale to more complex investigative scenarios with larger suspect pools.

The research addresses a critical gap in digital forensics: how to systematically narrow down persons of interest from large populations using linguistic cues in written or transcribed communications. The text analyzed could be from emails, instant messages, or even transcribed interviews, making the methodology potentially applicable across diverse investigative contexts.

Research Design & Methodology

The researchers had a large language model generate a larger fictional scenario with a pool of 18 suspects and two conspirators. Instead of one dialog-driven dataset, they created 18 separate fictional police interviews all generated using an LLM, representing a more realistic investigative scenario where suspects are interviewed individually.

The analytical framework employed four key psycholinguistic variables to identify persons of interest:

• Deception over time: Calculated using a Python library called Empath, which identifies deception-related words through statistical comparison with word embeddings
• Emotion analysis: Anger, fear, neutrality, and subjectivity levels across time, calculated using Hugging Face pre-trained transformers
• Entity-to-keyword correlation: Python lists containing suspicion-related linguistic cues including phrases like "I think," "I suspect," and "maybe it was"
• Contradictory narratives: Natural Language Inference where one statement is encoded as Premise and another as Hypothesis, with contradictory language labeled as Contradiction

The LLM was instructed to create 18 fictional people with two guilty conspirators—one with average anxiety/fear responses, and a second more morally pliable profile comfortable with deception and lying. This design allowed researchers to test whether their methods could detect psycholinguistic patterns embedded in LLM-generated guilty suspects.

Results & Key Findings

The researchers successfully determined the guilty parties identified in ground truth using entity to topic correlation, deception detection, and emotion analysis. The methodology correctly identified the two conspirators from the pool of 18 suspects.

Entity correlation findings revealed clear patterns: Owen Bishop was the person whom most others suspected of being guilty, while Gabriela Rios was the second most suspected person. Using spaCy and textblob features, average prediction scores ranged between 46 and 51%, showing small but meaningful variance.

Temporal deception patterns emerged: Specific bigrams like "potential consequences" and "Officer Whitaker" appeared at timestamps when deception levels were notably elevated. Cross-referencing the emotosubjectivity graph showed Gabriela's subjectivity crested at key moments, suggesting non-factual or speculative content.

The research demonstrated that combining multiple NLP techniques—n-grams, emotion tracking, deception scoring, and Natural Language Inference—provided converging evidence to narrow suspect pools. Through n-grams paired with deception, emotion, and subjectivity over time, they identified cues to better identify persons of interest from larger pools.

Discussion & Significance

This study represents an important methodological advancement in computational forensic linguistics. The researchers developed a more nuanced method of reverse engineering and breaking down psycholinguistic features of each suspect, moving beyond simple keyword matching to temporal patterns of linguistic behavior across multiple dimensions simultaneously.

The framework's multi-layered approach addresses a fundamental challenge in deception detection: no single indicator reliably signals deception. By integrating deception scores, emotional trajectories, entity correlations, and contradiction detection, the methodology triangulates evidence from multiple psycholinguistic channels. The temporal dimension—tracking how deception and emotion shift across interview timestamps—adds dynamic context missing from static text analysis.

However, the researchers acknowledge important constraints. While successful in solving the fictional crime, the methodology did not get the rigorous evaluative assessment hoped for. The methodology allows exploration of latent psycholinguistic cues that can lead investigators to identify a primary suspect, but validation with real-world data remains necessary.

Limitations & Considerations

The most significant limitation is the use of entirely LLM-generated data rather than authentic human communications. The LLM-generated police interviews presented interesting challenges, as the LLM created two inadvertent problems regarding the 18 datasets. LLM-generated text may not accurately reflect the psycholinguistic patterns of actual deceptive human speech, potentially limiting generalizability.

The researchers state unequivocally they are not calculating guilt—a critical ethical distinction. The methodology identifies patterns suggesting increased investigative attention may be warranted, not definitive proof of culpability. Additionally, The LLM's selection of conspirators appears based on the fact everyone suspected them—not actionable evidence that could push an investigation forward, highlighting the circularity risk when suspects emerge primarily through peer suspicion rather than independent evidence.

The study lacks validation metrics such as false positive rates, sensitivity/specificity measures, or comparison against human investigator judgments. Without benchmarking against established deception detection methods or real forensic cases, the practical accuracy of the framework remains uncertain.

Practical Applications

For digital forensics practitioners, this research offers a systematic framework for prioritizing investigative resources. When faced with large volumes of digital communications—emails, chat logs, interview transcripts—the multi-variable approach could help investigators identify which individuals warrant deeper scrutiny. The temporal analysis capability is particularly valuable, as it can reveal moments when suspects' linguistic patterns shift, potentially corresponding to knowledge of incriminating information.

For consumers and legal professionals, this research underscores both the promise and peril of AI-assisted forensic analysis. While NLP tools can process vastly more text than human investigators, their application must be tempered by understanding that linguistic patterns suggest probabilities, not certainties. Any operational deployment would require extensive validation on real cases, explicit acknowledgment of error rates, and human oversight to prevent algorithmic bias from disproportionately flagging innocent individuals based on linguistic style rather than substantive evidence.