Natural Language Processing: Improving Data Integrity of Police Recorded Crime

    By Darren Cook, Research Fellow in Natural Language Processing at City, University of London

    Did you know that police recorded crime data for England and Wales are not accredited by the UK’s Office for Statistics Regulation (OSR)? This decision, made by the OSR after an audit in 2014, was due to concerns about the reliability of the underlying data.

    Various factors affect the quality of police-recorded data. Differences in IT systems, personnel decision-making, and a lack of knowledge-sharing all contribute to reduced quality and consistency. Poor data integrity leads to a lack of standardisation across police forces and an increase in inaccurate or missing entries. I recently spoke about this issue at the Behavioural and Social Sciences in Security (BASS) conference at the University of St. Andrews, Scotland.

    Correcting missing values is no small feat. In a dataset of 18,000 police recorded domestic violence incidents, we found over 4,500 (25%) missing entries for a single variable. Let’s assume it takes 30 seconds to find the correct value for this variable – that’s 38 hours of effort – almost a full working week. Given that there could be as many as twenty additional variables, it would take over four months to populate all the missing values in our dataset. Expanding such effort across multiple police forces and for multiple types of crime highlights the inefficiency of human-effort in this endeavour.

    In my talk, I outlined an automated solution to this problem using Natural Language Processing (NLP) and supervised machine learning (ML). NLP describes the processes and techniques used by machines to understand human language, and supervised ML describes how machines learn to predict an outcome based on previously seen examples. In this case, we sought to predict the relationship between the victim and offender – an important piece of demographic information vital to ensuring victim safety.

    The proposed system would use a text-based crime ‘note’ completed by a police officer to classify the victim offender relationship as either ‘Ex-Partner”, “Partner”, or “Family” – in keeping with the distinction made by Women’s Aid. Crime notes are an often-overlooked source of information in police data, yet we found they consistently referenced the victim-offender relationship. The goal of our system, therefore, was to extract the salient information from the free-form crime notes and populate the corresponding missing value in our structured data fields.

    Existing solutions based on keywords and syntax parsing are used by multiple UK police forces. While effective, they require manual effort to create, update, and maintain the dictionaries, and they don’t generalise well. Our supervised ML system, however, can be automatically updated and monitored to maintain accuracy.

    When tested, our system achieved 80% accuracy, correctly labelling the relationship type in four out of five cases. In comparison, humans performed this task with approximately 82% accuracy – an arguably negligible difference. Moreover, once trained, our system could classify the entire test set (over 1,000 crime notes) in just sixteen seconds.

    However, we noted some limitations, the biggest of which was a high linguistic overlap in crime notes between ‘Ex-Partner’ and ‘Partner’ that caused several misclassifications. We believe more advanced language models (i.e., word embeddings) will improve discrimination between these relationships.

    We also discovered a potential prediction bias against minorities. Although victim ethnicity wasn’t included in our training setup, we observed reduced accuracy for Black or Asian victims. The source and extent of this bias are subjects of ongoing research.

    Our findings highlight the promise of automated solutions but serve as a cautionary tale against assuming these systems can be applied carte blanche without careful consideration of their limitations. Several outstanding questions remain. Is a system with 80% accuracy good enough? Is it better to leave missing values rather than predict incorrect ones? Incorrectly identifying a perpetrator as a current partner rather than an ex-partner could significantly impact the victim’s safety. Additionally, a model biased against certain ethnicities risks overlooking the specific needs of minority groups.

    The conference sparked lively and engaging conversation about many of these issues, as well as the role that automation can be play within the social sciences more broadly. A research article describing these results in full is the focus of ongoing work, and the presentation slides are available below as a download.

    For further information please contact Darren at darren.cook@city.ac.uk or via LinkedIn @darrencook1986

    Dr Darren Cook, An application of Natural Language Processing (NLP) to free-form Police crime notes – 1 download

    Photo by Markus Spiske on Unsplash

    Blog, Presentations