Machine learning-based textual analysis is a viable tool for police survey research. Analyzing large numbers of police free-text responses provides more nuanced understanding of police perceptions of the public. Officers' attention to professionalism guards against de-policing, while attention to perceived unfair criticism increases it. The public’s integrity has a stronger effect on propensity to de-police than the public’s knowledge about police work.The aims of this paper are twofold. First, the introduction and justification of machine-learning-based structural topic modeling (STM) in a policing context. STM provides a flexible platform for leveraging open-ended survey questions (Roberts, Stewart, & Tingley, 2014). Despite advances in other fields, to date, the method has received little attention in the criminology (and adjacent) literature, and to our knowledge, this paper is the first policing study to use STM methods. The second aim is to establish hypothesis validity for the STM technique. This goal is accomplished by testing the association between the derived structural topics and officers' general motivation regarding proactivity (or its absence – de-policing).