Exploiting Geolocation, User and Temporal Information for Monitoring Natural Hazards on Twitter

Víctor Fresno, Arkaitz Zubiaga, Heng Ji, Raquel Martínez

Procesamiento del Lenguaje Natural. 2015.

During emergency situation events it is important to acquire as much information about the event as possible, and social media sites like Twitter offer important real-time user contributed data. Typical Information Filtering techniques are keyword-based approaches or focused on co-occurrence with keywords. However, these approaches can miss relevant local information if messages do not contain an initially considered event-related keyword. Considering geolocation, user and temporal information within a pseudo-relevance feedback approach we can find event-related terminology but not co-occurring with initially considered keywords. Thus, taking into account the temporal aspect we can modify a query expansion function like Kullback-Leibler divergence in order to improve the Information Filtering process. Our proposed approaches have been evaluated in two Twitter datasets associated with real-world events, obtaining encouraging results.