Police Text Analysis: Topic Modeling and Spatial Relative Density Estimation

Sarah Huestis-Mitchell, Xiuyuan Cheng, Yao Xie

Abstract

We analyze a large corpus of police incident narrative documents in understanding the spatial distribution of the topics. The motivation for doing this is that police narratives in each incident report contains very fine-grained information that is richer than the category that is manually assigned by the police. Our approach is to split the corpus into topics using two different unsupervised machine learning algorithms - Latent Dirichlet Allocation and Non-negative Matrix Factorization. We validate the performance of each learned topic model using model coherence. Then, using a k-nearest neighbors density ratio estimation (kNN-DRE) approach that we propose, we estimate the spatial density ratio per topic and use this for data discovery and analysis of each topic, allowing for insights into the described incidents at scale. We provide a qualitative assessment of each topic and highlight some key benefits for using our kNN-DRE model for estimating spatial trends.

Introduction

Traditional crime analysis often relies on broad categories assigned by dispatchers during 911 calls. However, these labels frequently mask the nuanced details contained within the unstructured "incident narratives" written by responding officers. Furthermore, standard crime heatmaps based on absolute density (Kernel Density Estimation) are often biased by population centers, highlighting downtown areas simply due to higher activity levels rather than specific emerging trends. This paper introduces a framework that extracts hidden themes from police narratives and maps them using a novel relative density estimation technique to identify localized hotspots that are otherwise invisible.

Methodology

The proposed pipeline consists of two primary stages: topic modeling and spatial analysis.

Topic Modeling and Validation

The researchers applied Latent Dirichlet Allocation (LDA) and Non-negative Matrix Factorization (NMF) to a dataset of approximately 475,000 incident narratives from the Atlanta Police Department (2013–2017). Text was preprocessed using a bag-of-words model with bigrams and TF-IDF filtering. To determine the optimal model and number of topics, coherence measures (UMass, UCI, and NPMI) were evaluated.

Coherence measures (UMass, UCI, NPMI) comparing LDA and NMF performance across varying numbers of topics.

As shown in the coherence plots, NMF generally outperformed LDA as the number of topics increased, providing more interpretable and stable clusters for this specific corpus of short, noisy police text.

kNN-Relative Density Estimation (kNN-RDE)

To address population bias in spatial mapping, the authors developed kNN-RDE. Instead of absolute density, this method calculates the ratio of a specific topic's local density to the overall density of all incidents. By using a k-nearest neighbors approach to define local radii, the model identifies areas where a specific crime type is "over-represented" relative to the city-wide average.

Results and Discussion

The models successfully distilled five primary themes from the narratives, ranging from traffic violations to residential burglary.

Comparison of top keywords and weights for the five topics identified by NMF and LDA.

Spatial Insights

The kNN-RDE approach revealed significant geographic nuances. While standard heatmaps showed massive "blobs" over downtown Atlanta, the relative density model "subtracted" this background noise. For instance, it revealed that vehicle-related crimes were four times more prevalent in the affluent Buckhead area than the city average. In lower-population zones like Ben Hill, the model identified clear hotspots for residential burglary that would have been overlooked by absolute density models due to low total call volume.

Dispatcher vs. Model

A key finding was the model's ability to capture nuances conflated by human dispatchers. For example, specific types of vehicle damage or localized traffic violation clusters at specific intersections were clearly identified by the topic model but lost in broader dispatcher categories.

Conclusion

This research demonstrates that combining NLP with relative spatial statistics provides a more granular and less biased view of urban crime patterns. Future work will focus on the temporal dynamics of these hotspots and addressing potential biases inherent in officer narrative quality.

Summary updated

Abstract

We analyze a large corpus of police incident narrative documents in understanding the spatial distribution of the topics. The motivation for doing this is that police narratives in each incident report contains very fine-grained information that is richer than the category that is manually assigned by the police. Our approach is to split the corpus into topics using two different unsupervised machine learning algorithms - Latent Dirichlet Allocation and Non-negative Matrix Factorization. We validate the performance of each learned topic model using model coherence. Then, using a k-nearest neighbors density ratio estimation (kNN-DRE) approach that we propose, we estimate the spatial density ratio per topic and use this for data discovery and analysis of each topic, allowing for insights into the described incidents at scale. We provide a qualitative assessment of each topic and highlight some key benefits for using our kNN-DRE model for estimating spatial trends.
BibTeX
@misc{huestismitchell2022police,
  author = {Sarah Huestis-Mitchell and Xiuyuan Cheng and Yao Xie},
  title = {{Police Text Analysis: Topic Modeling and Spatial Relative Density Estimation}},
  year = {2022},
  eprint = {2202.04176},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.04176},
}
Figure