Abstract
The results suggest that text mining can automatically extract information from police-recorded domestic violence events that can support further public health research into domestic violence, such as examining the relationship of abuse types with victim injuries and of gender and abuse types with risk escalation for victims of domestic violence. Potential also exists for this extracted information to be linked to information on the mental health status.
Introduction
Domestic violence (DV) is a critical global public health issue, yet data collection remains a significant challenge. While police databases like the New South Wales Police Force’s (NSWPF) "WebCOPS" capture structured data such as dates and locations, the most vital details—the specific nature of the abuse and the resulting injuries—are often buried within unstructured "event narratives." These free-text descriptions, written by officers at the scene, are difficult to analyze at scale due to typos, police jargon, and complex clinical language. Traditionally, researchers had to manually review these records, a task that is impossible for datasets containing hundreds of thousands of entries. This study presents a rule-based text mining system designed to transform these messy narratives into structured, actionable data.
Methodology
The researchers developed a knowledge-driven, rule-based text mining approach using the GATE (General Architecture for Text Engineering) framework. Unlike machine learning models that can act as "black boxes," this approach relies on transparent, human-defined linguistic patterns to ensure high precision in a sensitive domain.
The system was built using three primary components:
- Custom Dictionaries: The team created 22 specialized dictionaries covering anatomy, weapons, verbs of force (e.g., "choking," "kicking"), and relationship types, including common misspellings found in reports.
- Syntactical Rules: 64 rules were developed to identify linguistic structures, such as a verb of action followed by a specific abuse type.
- Normalization: To ensure consistency, various mentions were mapped to a "canonical" form (e.g., "punch," "punched," and "punching" were all categorized as "punching").
Data and Evaluation
The study utilized a massive corpus of 492,393 domestic violence event narratives from the NSWPF recorded between 2005 and 2016. A training set of 200 narratives was used to refine the rules, while an evaluation set of 100 unseen narratives was used to test accuracy against human "gold standard" annotations.
The system demonstrated high reliability, achieving 90.2% precision for abuse types and 85.0% precision for victim injuries. Recall rates were similarly strong, at 89.6% for abuse types and 86.3% for injuries.
Key Insights and Findings
When applied to the full dataset of nearly half a million reports, the system revealed patterns that are often "hidden" in traditional databases:
- The "Invisible" Data: Over 71% of reports contained specific abuse types not captured in structured data fields, highlighting the massive loss of information in standard reporting.
- Prevalence of Non-Physical Abuse: "Emotional/verbal abuse" was the most common type identified (33.46%), underscoring that psychological harm is a central component of DV.
- Complexity of Violence: Over 33% of reports described three to five different types of abuse occurring within a single event.
- Injury Mapping: The most frequent physical abuse was "punching" (24.58%), and the most common injuries were "bruising" (29.03%) and "cuts/abrasions" (28.93%).
Conclusion and Future Directions
While the system is highly effective, it does face limitations, particularly with "implicit" injuries that require human inference (e.g., "dried blood on lips"). However, the ability to process nearly 500,000 records with high precision opens new doors for public health research. Future work aims to link this police data with health department records to explore the nexus between mental health and domestic violence, potentially identifying "victim phenotypes" that predict the escalation of violence and allow for earlier intervention.
Summary updated