Automatic Type Detection of 311 Service Requests Based on Customer Provided Descriptions

Abstract

The 311 phone number is for reporting non-emergency service requests (SRs) to authorities. This service is available through web, e-mail, and text message as well. Through this service, citizens would describe the issue and its location and the authorities would determine its category and the responsible unit and track the problem until it is resolved. The number of 311 SRs would amount to hundreds of thousands every year in some cities and determining the category of SRs manually is time-consuming, burdensome, and prone to human error. Additionally, these categories are not standardized across the states. In this paper, we standardize these categories across two cities and study the recurrent neural network’s ability in automatically determining the category of SRs based on the transcript of customer-provided descriptions. According to our results, the automatic categorization of these descriptions is not only faster and less cumbersome, but also more accurate than manual categorization. A close look at the mistakes made by the machine in labeling SRs revealed that in many cases either the SR’s description was insufficient to infer its category or the category identified by the machine was correct but the ground truth label assigned to that SR was incorrect.