Subevents detection through topic modeling in social media posts

Abstract

Event detection has been a significant topic for a long time, since the onset development of pervasive systems. The ability to gather data from various sensors, in a diverse number of formats, is a challenge due to the continuous growth of data volume. Users of social media act as human sensors, providing data and information in real time about entities and events. Most of the research about event detection – using human or non-human sensors – concentrates only on identifying events. These models assume an event to be a single entity and ignoring that it can be composed of other new events over time. The detection of subevents enriches the understanding of the main event, contextualizing it and creating a powerful knowledge about the scenario. To capture the parts of an event and the information changing over time, we created a scalable and modular topic modeling based algorithm. It identifies subevents and creates labels to represent them more accurately. We evaluate the proposed sub-event detection approach using two large-scale Twitter corpus. The first one is related to Brazil’s political protests scenario. The second analyzes the Zika Virus epidemic in the world. Our approach detected several subevents, most of them are related to real subevents. Due to the nature of social networks, with a minimum delay between an event occurrence and its dissemination, these results can open an opportunity for temporal tracking of emergence and outbreak scenarios.