Improving Problem-Oriented Policing with Natural Language Processing

The policing approach known as Problem oriented policing (POP) was outlined by Herman Goldstein in 1979. Despite POP being shown as an effective method to reduce crime it is difficult to implement because of the high analytical burden that accompanies it. This analytical burden is centred on understanding the mechanism by which a crime took place. One of the factors that contributes to this high burden is that a lot of the required information is stored in free- text data, which has traditionally not been in a format suitable for aggregate analysis. However, advances in machine learning, in particular natural language processing, are lowering the barriers for extracting information from free-text data.

This thesis explores the potential for pre-trained language models (PTMs) to efficiently unlock the information in police crime free-text data. PTMs are a new class of machine learning model that are ‘pre-trained’ to recognise the meaning of language. This allows the PTM to interrogate large quantities of free-text data. Thanks to this pre-training, PTMs can be adapted to specific natural language processing tasks with much less effort. Efficiently unlocking the information in the police free-text crime data should reduce the analytical burden for POP. In turn, the lower analytical burden should facilitate the wider adoption of POP. The thesis concludes that the evidence suggests PTMs are potentially an efficient method for extracting useful information from police free text data.