Read-Me.Org

View Original

Text mining police narratives of domestic violence events to identify coercive control behaviours

By Min-Taec Kim and George Karystianis (UNSW)
AIM To construct a measure of coercive control behaviours from free-text narratives recorded by the NSW Police Force for each domestic violence event, and to assess how useful this measure might be in detecting coercive control behaviours when combined with existing measures and/ or predicting which events are followed by violence within 12 months. METHOD We developed a text-mining system to capture a (non-exhaustive) set of behaviours that could be used to exert coercive control. Our text mining system consisted of a set of rules and dictionaries, with our definition of ‘coercive control behaviours’ drawn from Stark (2007). We concentrated on behaviours that were not already well captured by fixed fields in the police system. We applied this text mining system across all police narratives of domestic violence events recorded between 1 January 2009 and 31 March 2020, and used this data to construct our measure of coercive control behaviour. We then compared the incidence of coercive control behaviours using our measure against existing incident categories (where they exist). Finally, we estimated a gradient boosted decision tree (xgboost) model three times (once with standard predictors, once with just our coercive control measure, and once with both sets of variables included) and compared the performance of each model. This allowed us to estimate the contribution of our measure for improving the prediction of future violence. RESULTS There were 852,162 behaviours extracted by the text-mining system across 526,787 domestic violence events, with 57% of these events having at least one coercive control behaviour detected and 8% having three or more distinct subcategories of coercive control behaviours. These events were associated with 223,645 unique persons of interest. Our text mining system agreed with the pre-existing police incident categories (where they exist) in 80 to 99% of cases, with the text mining approach identifying an additional 30 to 60% of events that were not identified using the police incident categories, depending on the behaviour. The inclusion of the text mining variables did not improve the performance of our predictive model. CONCLUSION Our text mining system successfully extracted coercive control behaviours from police narratives, allowing us to identify how often these events are recorded in police narratives and supplement existing police categories of DV behaviours. Use of our measure in conjunction with existing incident categories significantly expands our ability to identify coercive control behaviours, however it does not improve our ability to predict which events are followed by further domestic violence.  

Crime and Justice Bulletin No. CJB260. No. 26o
Sydney:  NSW Bureau of Crime Statistics and Research. 2023. 28p.