Geotagging TweetsCOV19: Enriching a COVID-19 Twitter Discourse Knowledge Base with Geographic Information

Abstract

Various aspects of the recent COVID-19 outbreak have been extensively discussed on online social media platforms and, in particular, on Twitter. Geotagging COVID-19-related discourse data on Twitter is essential for understanding the different discourse facets and their regional relevance, including calls for social distancing, acceptance of measures implemented to contain virus spread, anti-vaccination campaigns, and misinformation. In this paper, we aim at enriching TweetsCOV19—a large COVID-19 discourse knowledge base of more than 20 million tweets—with geographic information. For this purpose, we evaluate two state-of-the-art Geotagging algorithms: (1) DeepGeo—predicting the tweet location and (2) GeoLocation—predicting the user location. We compare pre-trained models with models trained on context-specific ground truth geolocation data extracted from TweetsCOV19. Models trained on our context-specific data achieve more than 6.7% improvement in Acc@25 compared to the pre-trained models. Further, our results show that DeepGeo outperforms GeoLocation and that longer tweets are, in general, easier to geotag. Finally, we use the two geotagging methods to study the distribution of tweets per country in TweetsCOV19 and compare the geographic coverage, i.e., the number of countries and cities each algorithm can detect.