Detecting Institutional Dialog Acts in Police Traffic Stops

Detecting Institutional Dialog Acts in Police Traffic Stops Vinodkumar Prabhakaran Stanford Univ., CA Camilla Griffiths Stanford Univ., CA Hang Su UC Berkeley, CA Prateek Verma Stanford Univ., CA Nelson Morgan ICSI Berkeley, CA Jennifer L. Eberhardt Stanford Univ., CA Dan Jurafsky Stanford Univ., CA {vinodkpg,camillag,jleberhardt,jurafsky}@stanford.edu, {suhang3240,prateek119}@gmail.com, morgan@uprise.org Abstract We apply computational dialog methods to police body-worn camera footage to model conversations between police officers and community members in traffic stops. Rely- ing on the theory of institutional talk, we de- velop a labeling scheme for police speech dur- ing traffic stops, and a tagger to detect insti- tutional dialog acts (Reasons, Searches, Of- fering Help) from transcribed text at the turn (78% F-score) and stop (89% F-score) level. We then develop speech recognition and seg- mentation algorithms to detect these acts at the stop level from raw camera audio (81% F- score, with even higher accuracy for crucial acts like conveying the reason for the stop). We demonstrate that the dialog structures pro- duced by our tagger could reveal whether offi- cers follow law enforcement norms like intro- ducing themselves, explaining the reason for the stop, and asking permission for searches. This work may therefore inform and aid ef- forts to ensure the procedural justice of police- community interactions. 1 Introduction Improving the relationship between police officers and the communities they serve is a critical societal goal. We propose to study this relationship by ap- plying NLP techniques to conversations between of- ficers and community members in traffic stops. Traf- fic stops are one of the most common forms of po- lice contact with community members, with 10% of U.S. adults pulled over every year (Langton and Durose, 2013). Yet past research on what people ex- perience during these traffic stops has mainly been limited to self-reported behavior and post-hoc nar- ratives (Lundman and Kaufman, 2003; Engel, 2005; Brunson, 2007; Epp et al., 2014). The rapid adoption of body-worn cameras by po- lice departments in the U.S. (laws in 60% of states in the U.S. encourage the use of body cameras) and across the world has provided unprecedented insight into traffic stops.1 While footage from these cam- eras is used as evidence in contentious cases, the un- structured nature and immense volume of video data means that most of this footage is untapped. Recent work by Voigt et al. (2017) demonstrated that body-worn camera footage could be used not just as evidence in court, but as data. They devel- oped algorithms to automatically detect the degree of respect that officers communicated to drivers in close to 1,000 routine traffic stops captured on cam- era. It was the first study to use machine learning techniques to extract insights from this footage. This footage can be further used to unearth the structure of police-community interactions and gain a more comprehensive picture of the traffic stop as an every day institutional practice. For instance, knowing which requests the officer makes, whether and when they introduce themselves or explain the reason for the stop …

Content truncated to 500 words