💻 computer science

An Intelligent Decision Support System for Detecting Coercive Control: A Privacy-Preserving Approach

This paper presents a privacy-preserving framework that generates clinically accurate synthetic data to train a DistilBERT model for detecting coercive control, effectively overcoming the "Forensic Data Gap" by achieving high performance on real-world legal and social media benchmarks without compromising user privacy.

Original authors: Dhruv Patel, Abhishek Kaushik, Anju Johnson

Published 2026-08-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dhruv Patel, Abhishek Kaushik, Anju Johnson

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a crime, but the crime scene is a secret conversation between two people. The problem is, the law has very strict rules: you can't just grab the real chat logs because that would violate the people's privacy. This leaves the detective with a huge empty space where the evidence should be—a "Forensic Data Gap." To solve crimes like this, scientists usually need to study thousands of real examples to teach their computers what "bad behavior" looks like. But when the bad behavior is subtle, like a parent slowly making a child feel crazy or a partner slowly controlling them through silence and guilt (known as "coercive control"), it's incredibly hard to find real examples without breaking privacy laws.

This is where a branch of science called Natural Language Processing (NLP) comes in. Think of NLP as a way to teach computers to read and understand human language, not just by counting how many times a bad word appears, but by understanding the story, the timing, and the mood. Usually, computers are great at spotting obvious hate speech (like shouting "I hate you!"), but they are terrible at spotting the quiet, sneaky kind of abuse that happens over weeks or months. This paper tries to fill that empty evidence gap by building a "virtual crime scene." Instead of stealing real private chats, the researchers used a powerful AI to write thousands of fake conversations that are so realistic, they look and feel exactly like the real thing, but without hurting anyone's privacy.

The Big Idea: Building a Virtual Crime Scene

The researchers, Dhruv Patel, Abhishek Kaushik, and Anju Johnson, faced a tricky problem: how do you study psychological abuse without looking at real victims' private messages? They decided to build a "synthetic" dataset. Imagine a video game designer who needs to test a new level. Instead of waiting for real players to get stuck, they create a perfect, simulated version of the level to test the game mechanics. That is exactly what these scientists did, but for human behavior.

They used a smart AI (specifically a model called GPT-4) to write fake conversations between an "abuser" and a "victim." But they didn't just let the AI chat randomly. They gave it a strict rulebook based on real medical textbooks (the DSM-5) that describe a specific type of personality disorder called Narcissistic Personality Disorder (NPD). They told the AI to act out specific behaviors, like "grandiosity" (acting like you are better than everyone else), "lack of empathy" (ignoring when someone is sad), and "exploitativeness" (using people for your own gain).

The most creative part of their recipe was adding "time" to the mix. In real life, abusers often use silence as a weapon. They might wait 14 hours to reply to a text to make the other person anxious, or they might reply instantly to a sad message just to dismiss it. The researchers taught the AI to insert special "time tokens" into the fake chats, like <DELAY: 14 HOURS>. This forced the AI to simulate the slow, manipulative pacing of real abuse, rather than just writing angry words. They created 275 of these fake conversation pairs, totaling over 2,400 messages, to serve as their training ground.

Teaching the Computer to Spot the Invisible

Once they had their fake evidence, they needed to teach a computer how to spot it. They trained a smaller, faster AI model called DistilBERT on these fake conversations. Their goal was to see if this computer could learn to recognize the pattern of abuse, even if the words themselves sounded polite.

The results were quite interesting. When they tested the computer against older, simpler methods (like programs that just count bad words), the new AI won easily. The old methods failed miserably because they couldn't see the "subtext." For example, if an abuser says, "I'm only doing this because I care about your mental health," a simple word-counter sees the word "care" and thinks it's a nice message. But the new AI, having studied the fake patterns, recognized the controlling tone behind the nice words.

The computer trained on the fake data managed to detect coercive control with an F1-score of 0.75. In the world of computer science, this is a strong signal that the model is actually learning the behavior, not just guessing. Even more impressively, when they tested this computer on real data it had never seen before—like actual court transcripts from the UK and real Reddit threads—it still performed well, scoring 0.72 on the legal texts and 0.65 on the Reddit data. This suggests that the "fake" training data was good enough to teach the computer how to spot the "real" thing.

What the Computer Can't Do (And Why That's Okay)

The paper is very clear about what this tool is not. It is not a magic crystal ball that can diagnose a person with a mental illness. It is also not perfect. The computer made mistakes, mostly by being too cautious. It flagged some normal, heated arguments as abuse (a "false positive"). The researchers explain that this is actually a feature, not a bug, for a specific reason.

Imagine a smoke detector in a house. You want it to be so sensitive that it goes off even if you just burn a piece of toast, because you'd rather have a false alarm than miss a real fire. In the same way, this AI is designed to be a "triage tool" for human investigators. It's meant to flag suspicious conversations so a real human expert can look at them. The computer says, "Hey, this conversation has some weird patterns, please check it," and the human decides if it's actually abuse. The paper emphasizes that the computer cannot replace human judgment because it doesn't understand the complex social history between two people; it only sees the text and the timing.

The Verdict: A New Tool for a Privacy-Preserving Future

The main takeaway from this study is that we can fill the "Forensic Data Gap" without violating anyone's privacy. By using strict medical rules to generate fake conversations, the researchers created a dataset that is so realistic that a computer trained on it can successfully spot real-world abuse patterns. They proved that you don't need to steal private secrets to build tools that protect people.

However, the authors are careful to say this is a "suggestive" step forward, not a final solution. The fake data worked well in their tests, but more research is needed to see if it works across different cultures and languages. They also noted that the computer still struggles with some subtle forms of manipulation, like when an abuser hides their control behind "fake concern."

In the end, this paper offers a hopeful new path: a way to build intelligent tools that help detectives and social workers spot psychological abuse, all while keeping the real victims' private lives completely safe. It's a bit like training a guard dog using a realistic-looking robot dog; the guard dog learns to spot the threat, but no real animals were ever put in danger to teach it.

Technical Summary: An Intelligent Decision Support System for Detecting Coercive Control

Problem Statement
The empirical study of digital coercive control and psychological manipulation is currently hindered by a "Forensic Data Gap." While privacy regulations like the GDPR rightfully protect user data, they prevent researchers from accessing real-world logs of abusive interactions necessary to train detection models. Furthermore, existing toxicity datasets and lexically driven algorithms (e.g., VADER, TF-IDF) fail to capture the nuances of implicit abuse. These systems rely on overt hostility, profanity, or direct threats, missing the "semantic gap" where coercive control operates through subtext, omission, gaslighting, and the strategic manipulation of time (digital chronemics) rather than explicit aggression. Consequently, current automated moderation tools frequently dismiss reports of sustained narcissistic abuse because individual messages appear benign or supportive in isolation.

Methodology
To address the data scarcity and the limitations of shallow classification, the authors developed a three-stage, privacy-preserving framework:

Constraint-Based Procedural Generation:
- Data Synthesis: The authors generated a synthetic dataset of 275 conversational dyads (over 2,400 message turns) using GPT-4 (gpt-4-0613).
- Clinical Anchoring: Prompts were strictly constrained by DSM-5 criteria for Narcissistic Personality Disorder (NPD) and the Narcissistic Personality Inventory (NPI). The generation simulated specific behavioral clusters: Grandiosity, Lack of Empathy, Exploitativeness, and Entitlement.
- Temporal Encoding: A key innovation was the embedding of explicit temporal tokens (e.g., <DELAY: 14 HOURS>) to simulate the weaponization of silence, ghosting, and asynchronous pacing. This allowed the model to learn the "when" of abuse, not just the "what."
- Control Sets: To prevent the classifier from learning generator-specific artifacts (e.g., distinguishing AI text from human text), the authors created a "Synthetic Normative Control Set" (using the same GPT-4 parameters for non-abusive dialogues) and a "Cross-Generator Holdout Corpus" (using Llama-3-70B).
Multi-Dimensional Validation:
- Expert Assessment: Two clinical psychology experts blindly evaluated 15% of the synthetic dyads, achieving a strong inter-rater reliability (Cohen's $\kappa = 0.84$ ).
- Statistical Fidelity: Lexical diversity (Type-Token Ratio) and semantic overlap (SBERT cosine similarity) were compared between synthetic and real-world baselines. The synthetic data showed low discriminability from real human dialogues in terms of distribution, confirming high fidelity.
Model Training and Evaluation:
- Architecture: A DistilBERT sequence model was fine-tuned exclusively on the synthetic data. The model was designed to process multi-turn histories via a sliding-window approach to capture longitudinal patterns.
- Ablation Studies: The authors tested the model with and without temporal tokens to isolate the contribution of "digital chronemics" versus semantic subtext.
- Transfer Learning: The model was evaluated on external, real-world benchmarks without retraining (Train-on-Synthetic, Test-on-Real or TSTR):
  - Legal Benchmark: 40 UK Court of Appeal transcripts (Section 76 Serious Crime Act 2015).
  - Domain Transfer: 1,428 Reddit threads from the Contextual Abuse Dataset (CAD).
  - Cross-Generator: 75 dyads generated by Llama-3-70B.

Key Results

Performance on Synthetic Data: The fine-tuned DistilBERT model achieved an F1-score of 0.75 and an AUC of 0.82 on the primary 5-fold cross-validation. This significantly outperformed lexical baselines (VADER F1=0.17, SVM F1=0.45) and zero-shot LLMs (Llama-3 F1=0.57).
Role of Temporal Encoding: Removing explicit delay tokens reduced the F1-score from 0.75 to 0.71, demonstrating that while semantic subtext is primary, temporal pacing provides a statistically significant structural signal for detection.
Generalization:
- Legal Evidence: The model achieved an F1-score of 0.72 on real UK legal transcripts, proving its ability to generalize to formal, court-admissible evidence.
- Reddit Data: It achieved an F1-score of 0.65 on the CAD dataset, indicating it does not rely solely on generic lexical hostility.
- Cross-Generator: The model maintained an F1-score of 0.76 on the Llama-3 holdout, confirming it detects behavioral patterns rather than specific generator artifacts.
Error Analysis: The model exhibited high Recall (0.88) but moderate Precision (0.65). This indicates a tendency to flag highly emotional normative conflicts as potential abuse (False Positives), a trade-off the authors argue is acceptable for a triage tool where missing abuse (False Negatives) is the critical failure mode.

Significance and Claims
The paper claims to offer a methodological solution to the Forensic Data Gap, demonstrating that clinically constrained synthetic data can safely bypass privacy bottlenecks while preserving the task-relevant linguistic and behavioral characteristics of real-world abuse.

Paradigm Shift: The study argues that NLP must move beyond isolated sentence classification toward computational psychometrics that model longitudinal behavioral clusters and temporal dynamics.
Expert System Role: The authors position their system not as an autonomous diagnostic tool for personality disorders, but as an Intelligent Decision Support System (IDSS) for digital forensic examiners. It serves as a "human-in-the-loop" triage assistant to prioritize evidence for human review.
Privacy and Ethics: By generating data anchored in clinical criteria rather than harvesting victim logs, the framework allows for the development of protective interventions without compromising user privacy or re-traumatizing victims.
Limitations: The authors modestly acknowledge that the system is not yet ready for autonomous deployment. They highlight the need for human oversight to interpret context, the current limitation to English-language NPD markers, and the necessity for future work on multilingual and culturally diverse datasets.

In conclusion, the study provides preliminary evidence that sequence-aware transformers, trained on privacy-preserving synthetic data, can effectively detect the subtle, context-dependent patterns of coercive control that current lexical safety mechanisms miss.

Technical Summary: An Intelligent Decision Support System for Detecting Coercive Control

More like this