Language Influences on Tweeter Geolocation

We investigate the influence of language on the accuracy of geolocating Twitter users. Our analysis, using a large corpus of tweets written in thirteen languages, provides a new understanding of the reasons behind reported performance disparities between languages. The results show that data imbalance has a greater impact on accuracy than geographical coverage. A comparison between micro and macro averaging demonstrates that existing evaluation approaches are less appropriate than previously thought. Our results suggest both averaging approaches should be used to effectively evaluate geolocation.