Abstract: Peaks in a time series of social media posts can be used to identify events. Using peaks in the number of posts and keyword bursts has become the go-to method for event detection from social media. However, those methods suffer from the random peaks in posts attributed to the regular daily use of social media. This paper proposes a novel approach to remedy that problem by introducing a semantic decay filter (SDF). The filter's role is to eliminate the random peaks and preserve the peak related to an event. The filter combines two relevant features, namely the number of posts and the decay in the number of similar tweets in an event-related peak. We tested the filter on three different data sets corresponding to three events: the STEM school shooting, London bridge attacks, and Virginia beach attacks. We show that, for all the events, the filter can eliminate random peaks and preserve the event-related peaks.