The Bloomsbury Colleges

Using Machine Learning to Understand Coercive Control

Principal Supervisor: Professor Karen Devries, London School of Hygiene and Tropical Medicine (LSHTM)

Co-Supervisor: Dr David Weston, Birkbeck College, University of London

Project Description

Background

Globally, WHO estimates that one in three women will experience violence from an intimate partner in her lifetime. Alongside physical and sexual violence, emotional violence and coercive control are common features of these experiences.  While the prevalence and types of acts of physical and sexual violence experienced by women are relatively well-characterised, emotional violence and coercive control remain less well understood.

The UK Crown Prosecution Service defines coercive control as “a pattern of behaviour that is used to make a person subordinate and/or dependent.” Current academic literature conceptualises three major ‘facets of the construct of coercive control in a relationship – 1) intentionality and motivation within the perpetrator to obtain control over the target; whether conscious or unconscious; 2) perception of the behavior as negative by the target, and 3) the ability of the perpetrator to make a credible threat as perceived by the target’2. These dynamics manifest in a large range of different behavioral acts, which make coercive control difficult to identify. Victims often remain in coercive and controlling relationships for years before being able to label them as abusive, and are often at risk for heightened abusive and controlling behaviours post-separation.

According to OfCom, in 2023, 67% of UK adults used WhatsApp as their main mode of text communication. Coercive, controlling and emotionally abusive communication between intimate partners is therefore highly likely to be captured in messages sent via WhatsApp and other digital platforms.  Using modern machine learning methods, it is possible to characterise various features of textual communication as to overt qualities such as tone, including aggressive and sarcastic tones, and covert disruptive behaviour such as trolling. Indeed a pilot study funded by the home office has shown great promise for identifying coercive control 4 from textual communication albeit with a very small curated sample.

If machine learning analysis is able to identify fine-grained patterns of textual communication indicative of coercion and emotional abuse, there are numerous potential applications. A better characterisation of the communication patterns indicative of abuse would broaden academic understanding of these forms of abuse. These tools could be used by women themselves early on in relationships to understand whether communication can be classified as coercive or abusive. It could be offered to women by therapists, health practitioners, and legal practitioners to support diagnosis and appropriate provision of support.

Aims and objectives

The aim of this project is to explore whether machine learning can improve understanding and characterisation of communication patterns in coercive control against women. Specific  objectives will be developed by the student. Provisional objectives are to: 1) describe features of text-based communication typical of coercive control; 2) compare these to features of non-coercive communication; 3) understand whether it is possible to develop a model that can classify text communication as coercive or non-coercive.

Proposed methodology

The approach comprises three main phases:

  • Data Collection and Curation
  • Manual Labelling
  • Model Development and Evaluation

Data Collection and Curation

In  partnership with organisations who work with survivors of violence, we will  invite survivors of violence who have experienced coercive control to refine project objectives in collaboration with the student and supervisors. A data collection strategy will be developed, and we envisage inviting about 50 women survivors to share anonymised versions of their WhatsApp histories. We will sample women without histories of coercive control, and aim to recruit approximately 50 women from each group. Text from these histories, stripped of all personally identifiable information but retaining timestamp metadata, will be the main data used in the analysis.

Manual Labelling

The raw text histories will be manually annotated to identify regions of coercive/non-coercive communication. These annotations will be validated by independent annotators and will serve as our ground-truth for subsequent experiments.

Model Development and Evaluation

to the student will investigate the extent to which regions of coercive control communication can be identified using Natural Language Processing (NLP) tools, beginning with replicating existing experiments in coercive control and further refine these approaches by using methods from adjacent problems such as identifying intimate partner violence from tweets. In particular, the student could explore whether linguistic or emotional frequency patterns serve as reliable features.

Ethical considerations

Devries has extensive experience working on violence against women and children, and will ensure that procedures are safe, ethical and respectful, and do not put participants at risk of further harm3. We will invite a survivor of violence to participate as a thesis advisor and will also invite survivors of violence to review our manual labelling framework developed during the PhD. The student will receive training in safe approaches to involve survivors.

Significance

The work will be among the first application of machine learning to characterise and identify patterns of communication in coercive control, and is therefore likely to be highly significant for those working in the fields of violence and women’s health. If the approach is successful there are several practical applications possible.

The project is likely to be of substantial interest to violence survivors. One of the main issues that consistently emerges for survivors is wanting to better understand, and wanting more public awareness, of the coercive and controlling aspects of their abuse.

Timescale

The PhD will be organised over 3 years, starting from October 2026. Year 1 will involve the literature review. Year 2 will involve data collection and manual labelling, and exploratory data analysis. Year 3 will involve model development and evaluation. The PhD will be by publication.

Outcomes

The student will produce three academic papers: one literature review on the use of machine learning to characterise coercive control and emotional abuse; one descriptive paper which presents findings from the annotation and exploratory data analysis phase, and one based on evaluating NLP models for detecting coercive control.

Plans for dissemination

The student will widely disseminate the work, directly to survivors via partner organisations, at key academic conferences on violence (SVRI, ISPCAN) and applied computer science conferences, and via popular media.

Subject Areas/Keywords

Subject Area: behaviour, violence, computer science, epidemiology

Keywords: Natural Language Processing, machine learning, artificial intelligence, coercive control, violence

Key References:

Further details about the project may be obtained from:

Principal Supervisor: Professor Karen Devries, karen.devries@lshtm.ac.uk

Co-Supervisor: Dr David Weston, dj.weston@bbk.ac.uk

Further information about PhDs at LSHTM is available from:

Applying for a Research degree | How to apply | LSHTM

How to Apply

The application process has two steps. To be considered for the funding, applicants must meet all eligibility criteria and complete both steps outlined below by the scholarship deadline stated for the project they are applying for.

  • Step 1
    Submit an application for research degree study via the LSHTM application portal. Applicants should apply via the Faculty of Infectious & Tropical Diseases (ITD).
    • Students should submit a research proposal based on the advertisement for their project.
    • Incomplete applications will not be considered for this studentship.
  • Step 2

Once you have submitted an application to study you should receive an automated email from the Scholarships Team (scholarships@lshtm.ac.uk) providing you with the link to our online scholarships application portal.

  • This will provide you with a temporary password to use the first time you login (via e-vision), which you should then update.
  • Please search for ‘Bloomsbury PhD Studentship’ if you wish to apply for this funding, and then answer the questions online to indicate your interest in one of the two funded PhD projects available.  Once you are happy with your responses you can press submit and should receive a confirmation of receipt email at your contact email address.
  • The scholarships portal will not be able to accept applications after the relevant project deadline (see below).
  • The Scholarships team will be in touch with an outcome in due course.

Applications for this project will only be reviewed and processed after the deadline. All applications that are submitted before the deadline will be considered equally, regardless of submission date.   

Applicants must meet the School’s minimum English language proficiency requirements if shortlisted for this funding by Monday 15 June 2026. Failure to do so may result in any scholarship offer being withdrawn and offered to a reserve candidate instead.

By submitting an application for this funding applicants agree to its Terms & Conditions.

Closing date for applications is:

16:00 (GMT) on 1 March 2026