Improving Policing with Natural Language Processing

Anthony Dixon, Daniel BirksACL 2021

TL;DR

This paper proposes using Natural Language Processing (NLP) to enhance Problem-Oriented Policing (POP) by unlocking insights from unstructured police free-text data. By automating the analysis of crime reports, NLP can reduce the analytical burden on police, enabling more effective crime prevention and a shift towards a less reactive policing model.

What Problem Does This Paper Address?

Traditional policing often relies on a reactive approach, focusing on responding to crimes after they occur, leading to costly criminal justice system interventions. Problem-Oriented Policing (POP) offers a more proactive strategy by identifying and addressing the underlying conditions that allow crimes to happen. However, implementing POP effectively requires a deep understanding of crime events, which is often buried in unstructured free-text data (e.g., officer notes, witness statements). Police agencies typically lack the resources and expertise to analyze this data at scale, creating a significant impediment to widespread POP adoption. This paper aims to bridge this gap by demonstrating how NLP can automate the extraction of valuable insights from this data, thereby supporting and expanding POP initiatives.

Methods

The paper advocates for the application of several NLP techniques to support POP:

  • Classification: Used to identify and categorize contextual factors within crime reports (e.g., presence of alcohol, type of entry). This enhances structured data and aids in grouping similar crimes for POP analysis.
  • Named Entity Recognition (NER): Applied to extract specific details from crime reports, such as weapon types or victim-offender relationships. This helps in more precise problem grouping and understanding crime methods.
  • Clustering: An unsupervised technique to group similar crimes based on their descriptions, moving beyond pre-existing administrative categories. This can reveal emerging modus operandi and inform targeted POP responses.
  • Data Richness Analysis: While not a specific NLP technique, the paper emphasizes the importance of understanding the quality and detail of the free-text data itself, as NLP outputs are dependent on this input.
  • Pre-trained Language Models (e.g., BERT): The paper suggests leveraging existing large language models, which can reduce the need for extensive feature engineering and training, making NLP more accessible to police agencies.

Datasets

Not explicitly detailed as specific datasets were used for training or evaluation. The paper discusses the nature of police free-text data, including misspellings, specialized vocabulary, and unedited text, highlighting the need for model adaptations. The authors mention using a "white-listing approach" that redacts tokens not found in a list of commonly used words as a pre-processing step for data sharing.

Key Insights and Ideas

  • NLP as a POP Enabler: NLP has the potential to significantly lower the analytical burden associated with POP, making it more accessible and scalable for police agencies.
  • Unlocking Tacit Knowledge: Police free-text data contains valuable tacit knowledge held by officers that is difficult to derive from structured data. NLP can unlock this information.
  • Shift from Reactive to Proactive Policing: By enabling better POP, NLP can contribute to a societal shift towards crime prevention rather than solely reactive responses, leading to fewer individuals entering the criminal justice system.
  • Ethical Considerations are Paramount: The paper strongly emphasizes that the use of NLP in policing must be approached with careful consideration of potential biases. These biases can lead to resource allocation inequities and perpetuate existing societal problems.
  • Data Sharing Challenges: Sharing sensitive police data for NLP research presents significant logistical and ethical hurdles, suggesting that NLP solutions may need to "travel to the data" within police IT systems.

Key Results

The paper does not present quantitative results from specific NLP model evaluations. Instead, it focuses on the potential of NLP to address the challenges of POP implementation. The authors argue that by applying NLP techniques like classification, NER, and clustering to police free-text data, police agencies can gain deeper insights into crime events, leading to more effective POP strategies. The Durham burglary example illustrates how POP, when supported by enhanced analysis (implicitly through NLP), can reduce crime and improve community safety.

Limitations and Future Work

  • Bias Mitigation: A significant limitation acknowledged is the risk of introducing or perpetuating biases through NLP. The paper stresses the need for ethical considerations and bias mitigation at all stages of NLP implementation.
  • Data Coverage and Richness: The effectiveness of NLP is limited by the completeness and detail of the police data itself. Biases in reporting and data recording can be replicated by NLP.
  • Technical Expertise: Police agencies often lack the necessary technical expertise to implement and manage NLP systems, and private providers may lack transparency.
  • Low-Resource Environments: The paper highlights the need for NLP solutions that can operate effectively in low-resource environments, without requiring high-performance computing.
  • Future Work: The authors suggest future research should focus on examining how existing NLP models can be utilized in low-resource environments to enhance the characterization of crime events for POP purposes. They also emphasize the responsibility of the academic community to investigate potential problems and develop appropriate safeguards before NLP systems are deployed operationally.

Summary updated

BibTeX
@inproceedings{dixon2021improving,
  author = {Anthony Dixon and Daniel Birks},
  title = {{Improving Policing with Natural Language Processing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure