Improving police recorded crime data for domestic violence and abuse through natural language processing.

Introduction: Domestic Violence and Abuse (DVA) is a growing public health and safeguarding concern in the UK, compounded by long-standing data quality issues in police records. Incomplete or inaccurate recording of key variables undermines the ability of police, health services, and partner agencies to assess risk, allocate resources, and design effective interventions. Methods: We evaluated two machine learning models (Random Forest and DistilBERT) for classifying the type of victim/offender relationship (ex-partner, current partner, and family) from approximately 19,000 DVA incidents recorded by a UK police force. Models were benchmarked against a static rule-based classifier and assessed using precision, recall, and F1-score. To reduce false positives in the most challenging relationship categories, we implemented a selective classification strategy that abstained from low-confidence predictions. Results: Both machine learning models outperformed the baseline across all metrics, with average absolute gains of 11% in precision and 16% in recall. Ex-partner cases were classified most accurately, while current partner cases were classified with the least accuracy. Selective classification substantially improved precision for underperforming categories, albeit at the expense of reduced coverage. Discussion: These findings demonstrate that computational tools can enhance the completeness and reliability of police DVA data, provided their use balances predictive accuracy, interpretability, and safeguarding risks.