Witness statements for busy coppers
Applying machine learning to unstructured statements...
“We feel sure this innovation could be introduced to policing over time and we would be pleased to support the team as they move forward.”
-- Russell Gould, Clue
“WIT was a very impressive piece of technology, and demonstrated clearly how we can make unstructured text more useful to us across the work we do.”
-- Alex Blatchford, MPS
WIT is a web application that allows witnesses to fill in their own statements.
WIT uses natural language processing (NLP) to identify entities in statements. Entities can be used to elicit additional information from the witness; for example, if the witnesses mentions a "car", WIT can ask for its colour and model. Furthermore, WIT represents in a graph the entities that have been discovered across multiple statements related to a single incident. This allows officers to quickly identify key people, locations, and objects involved in an incident.
Our main motivation for developing this project, in the context of Hack the Police 4 (HTP4), was to help the force save time and resources, and to allow officers to gather evidence in cases where it is not possible to record witnesses' statements in the traditional way. For instance, if an incident occurs during rush hour, witnesses may be unable to stay and wait while officers administer interviews one by one. Using WIT, an officer can generate a personalised code for each witness, who can then use their phone to access the WIT application, submit their code, and provide their statement. Officers can focus on witness care and simultaneously supervise multiple witnesses while they prepare their statements. Videos, picture, or audio recorded on the phone could also be submitted.
We spoke with several officers in HTP4 who pointed us in the right direction to ensure that statements gathered through WIT would be up to standard. WIT's forms follow the Five Part Statement Structure, including the "ADVOKATE" set of questions. Furthermore, we designed the forms so that they include the fields in the MG11 form of the Prosecutor's Manual. This affords the possibility of automatically generating the relevant MG11 documents (and MG15 for audiovisual evidence) from reports submitted via WIT.
We were also motivated by a desire to find applications for areas of Artificial Intelligence that we are passionate about. In particular, Anshul was interested in using NLP for entity recognition, and David was eager to use knowledge graphs. This led us to applying the spaCy NLP library to process witness statements, which opened the door for a wide range of applications.
In addition to eliciting additional information and constructing entity graphs, we envision the possibility of using WIT to find inconsistencies in statements, or across multiple statements. Entity graphs could also be exported to case management software such as Clue, a HTP4 sponsor who provided us with training datasets. Another way to take forward the project, inspired by a suggestion received during our presentation of WIT at New Scotland Yard, would be to use the application to automatically identify, from initial statements, witnesses who observed an incident in its entirety, who can therefore provide highly detailed statements.