Geolocation of Tweets in Spanish with Transformer Encoders

Tweet geolocation is very important in many contexts: disaster relief, opinion polling, recommendation systems, etc. There are some recent studies showing that tweets with geolocation tags are sparse in several settings. Current state of the art geolocation algorithms for tweets are based on natural language processing methods. Most of these algorithms have been tested in English.Transformers are machine learning models based on attention mechanisms. These models have been proven successful in many natural language processing and computer vision scenarios. In this paper, we propose a transformer model for tweet geolocation. We describe several experiments for tweets in Spanish located in the Mexican region.