Civil unrest is public manifestations, where people demonstrate their position for different causes. Sometimes, violent events or riots are unleashed in this kind of events, and these can be revealed from tweets posted by involved people. This study describes a methodology to detect riots within the time of a protest to identify potential adverse developments from tweets. Using two own datasets related to a violent and non-violent protest in Peru, we applied temporal clustering to obtain events and identify a tweet headline per cluster. We then extracted relevant terms for the scoring and ranking process using a different domain and contrast corpus built from different sources. Finally, we filtered the relevant events for the violence domain by using a contrast evaluation between the two datasets. The obtained results highlight the adequacy of the proposed approach.