From Narratives to Insight: Improving Police Administrative Data with Machine Learning

The UK Government’s recent policing reform agenda places new emphasis on the role of data and technology in modern policing (Home Office, 2026). As part of this, Police.AI, the proposed National Centre for AI in Policing (NPCC, 2026) aims to deploy artificial intelligence to improve internal processes and reduce administrative burden, allowing officers to focus more time on frontline responsibilities. 

Yet one of the long-standing obstacles to this vision concerns the quality of policing data itself. Differences in systems, training, and internal polices across the UK’s 43 forces, coupled with time-constraints and operational pressures have resulted in inconsistent data recording practices characterised by missing fields. These institutional patterns of digital behaviour have been significant enough that police recorded crime lost its accreditation as an official statistic with the OSR over ten years ago (Whitehead, 2024). 

Improving the quality and usability of operational data is therefore central to the broader reform agenda. The planned National Police Service (NPS), for example, is expected to strengthen the use of data to identify emerging trends and support evidence-based decision making. Achieving this ambition requires not only improved data collection methods, but finding new ways to extract value from information within existing police systems and archives. In many cases, operational systems already contain rich contextual information, yet this remains difficult to extract and analyse consistently using conventional analytical methods.

Differences in systems, training, and internal polices across the UK’s 43 forces, coupled with time-constraints and operational pressures have resulted in inconsistent data recording practices characterised by missing fields.

One largely untapped source of information lies in the short narrative descriptions routinely recorded by officers when documenting incidents. These free-text summaries capture contextual details about events that are not always reflected in structured administrative fields. Recent advances in natural language processing make it possible to analyse this type of narrative data at scale, allowing structured information to be extracted directly from operational text.

Extracting Structured Information from Narrative Records

To explore how narrative text could improve administrative data, we developed a supervised machine learning model capable of identifying and extracting structured information from short free-text descriptions contained within operational records (Cook et al., 2025). The approach used a modern language model (DistilBERT) trained on a labelled dataset of incident narratives.

The model learns patterns in narrative descriptions that indicate the presence of specific structured attributes. Once trained, it can infer the most likely value referenced within a short officer-written description. This allows structured variables to be recovered directly from operational text, providing a practical way to supplement or correct incomplete records.

Performance and Scalability

After training, the model correctly extracted the target attribute from narrative descriptions in more than four-out-of-five cases. As expected, performance was closely related to representation in the training data: values more frequently encountered during training were detected more reliably, while rarer cases were more difficult to identify. The most encountered errors were between categories with a high semantic similarity. Future work could potentially reduce this by adjusting feature weighting to improve the signal-to-noise ratio, encouraging semantically similar categories to separate more clearly within the model’s vector space.

Training the model required less than two hours using standard CPU hardware and processing the full dataset (15,000 records) took under three minutes. In practical terms, this represents a substantial increase in efficiency compared to manual review and correction.

Balancing Automation and Human Oversight

Machine learning models produce probabilistic predictions, meaning they estimate how confident they are in each output. This allows practitioners to control the balance between automation and reliability.

By setting a confidence threshold, only predictions above a chosen level of certainty are accepted automatically. Lower-confidence cases can instead be flagged for human review (Butcher et al, 2024).

The policing reform agenda places renewed emphasis on the role of data and technology in improving operational effectiveness. 

In our experiments, adjusting this threshold produced a useful middle ground: around 60% of records could be processed automatically while maintaining error rates close to one-in-twenty. This type of hybrid workflow allows institutions to benefit from large efficiency gains while retaining expert oversight where needed.

Strengthening Administrative Data Using Narrative Text

Operational text contains structured insight that policing institutions are often unable to analyse systematically using traditional statistical approaches. Techniques such as these provide a way to recover missing information from existing records and improve the quality of administrative datasets without large-scale changes to data collection systems.

Used carefully, this type of technology could help revitalise under-used data sources, strengthen analytical capability, and support more informed decision-making across the policing system.

Conclusion

The policing reform agenda places renewed emphasis on the role of data and technology in improving operational effectiveness. However, the value of these ambitions will ultimately depend on the quality and usability of the data that underpin them.

Operational systems already contain large volumes of narrative text that capture important contextual details about incidents. While traditionally difficult to analyse at scale, recent advances in natural language processing make it possible to extract structured insight from these records efficiently.

Used carefully, these approaches offer a practical way to improve the completeness and usability of existing administrative datasets. Combined with appropriate safeguards and human oversight, including clear governance around confidence thresholds and safeguards against automation bias, they provide a route to strengthening the evidence base that supports policing decisions without imposing additional reporting burdens on frontline officers.