Exploring the roles of social media data to identify the locations and severity of road traffic accidents

People tend to use social media to share information about nearby events which includes traffic accidents. Traffic accident reporting over the phone can initiate medical aid, however it often fails to correctly specify severity, location, and assessment of the overall situation. Social media information (i.e., tweets, posts, etc.) can be mined to extract supportive information to be used to improve reporting accuracy and reduce response time of first responders. In this paper, we developed a framework that can continuously analyze and extract relevant accident reports and tested it using the data from four cities in the U.S. and Nigeria. In this framework, we collected tweets from Twitter API, identified whether they are accident-related or not, clustered a group of tweets talking about the same accident, and performed a severity analysis based on the summary of the tweets. We then geolocated the accidents for which the location is mentioned (i.e. direct geo-coding) or provided an approximate location for accidents by estimating user location-based twitter feed (i.e. indirect geo-coding). We also used semantic role labeling approach for severity detection and present the accuracy with respect to annotated data. The results of empirical testing revealed that city-level locations were identified for 71-97% of the accidents and geo-coordinates were obtained for 33-83% of the accidents, varying across the study sites and geolocation methods. Our framework demonstrates that on average 9-11% cases social media precedes on publishing accident related information than that of actual police reports. We will also discuss our approach of using Distributed, Big Data frameworks to process large number of Tweets generated in a streaming manner.