Word embeddings and deep learning for location prediction: tracking Coronavirus from British and American tweets

Abstract

With the propagation of the Coronavirus pandemic, current trends on determining its individual and societal impacts become increasingly important. Recent researches grant special attention to the Coronavirus social networks infodemic to study such impacts. For this aim, we think that applying a geolocation process is crucial before proceeding to the infodemic management. In fact, the spread of reported events and actualities on social networks makes the identification of infected areas or locations of the information owners more challenging especially at a state level. In this paper, we focus on linguistic features to encode regional variations from short and noisy texts such as tweets to track this disease. We pay particular attention to contextual information for a better encoding of these features. We refer to some neural network-based models to capture relationships between words according to their contexts. Being examples of these models, we evaluate some word embedding ones to determine the most effective features' combination that has more spatial evidence. Then, we ensure a sequential modeling of words for a better understanding of contextual information using recurrent neural networks. Without defining restricted sets of local words in relation to the Coronavirus disease, our framework called DeepGeoloc demonstrates its ability to geolocate both tweets and twitterers. It also makes it possible to capture geosemantics of nonlocal words and to delimit the sparse use of local ones particularly in retweets and reported events. Compared to some baselines, DeepGeoloc achieved competitive results. It also proves its scalability to handle large amounts of data and to geolocate new tweets even those describing new topics in relation to this disease.

Keywords: Apache; Bi-LSTM; Coronavirus spread; Geolocation; Neural networks; Tweets; Word embeddings.