How Can We Help You? Automatic Categorisation of Urban Problem Reports

In almost all German cities with more than 100.000 inhabitants a citizens can report problems from potholes to fly tipping online. Most systems require citizens to categorise their reports manually, which is supposed to make processing easier. In this paper, we use an LLM-based text classifier to automatically assign citizen reports to categories. This should be easy if categories are well-defined and easy to separate. We develop and test our approach on n = 4912 problem descriptions from Stadtverbesserer, the reporting system of a large German city, Oldenburg. Using the open source Llama 3.1 models, we obtained an accuracy of 79%, which was mainly due to substantial overlap between two pairs of categories covering trash and green spaces. We suggest that future work should focus on automatic routing of reports to agencies instead of intermediate categories.