Tremendous volumes of messages on social media platforms provide supplementary traffic information and encapsulate crowd wisdom for solving transportation problems. However, social media messages manifested in human languages are usually characterized with redundant, fuzzy and subjective features. Here, we develop a data fusion framework to identify social media messages reporting non-recurring traffic events by connecting the traffic events with traffic states inferred from taxi global positioning system (GPS) data. Temporal-spatial information of traffic anomalies caused by the traffic events are then retrieved from anomalous traffic states. The proposed framework successfully identified accidental traffic events with various scales and exhibited strong performance in event descriptions. Even though social media messages are generally posted after the occurrence of anomalous traffic states, resourceful event descriptions in the messages are helpful in explaining traffic anomalies and for deploying suitable countermeasures.