Conventional clustering-based method for event detection on social networks

Anonymous

Abstract: Social networks are becoming the preferred channel to report and discuss events happening around the world. The information stream such channels contain can be used to detect and describe the ongoing events to take informed decisions in numerous domains. A typical framework for event detection is to first cluster the stream of tweets, and then analyze the clusters to decide which deal with real-world events. In this context, content representation models and clustering approaches are critical. Classical approaches are usually based on TF-IDF for the representation of the text content and on dynamic clustering for the clustering part. In this paper, we propose to compare TF-IDF with recent text representation models and we propose an event detection method based on conventional clustering. We show that, contrary to previous results, language models based on Transformer architectures are competitive with TF-IDF. We also show that our approach outperforms the most used approach of the literature.

Paper Type: long