Significance of Linguistic Indicators for Location Prediction

Twitter is one of the most widely and frequently use microblog. Thousands of tweets are being posted daily on twitter. With the tweet, peoples are share their current location using check-ins point of interest that is known as geo-tagged tweet. However, most of the people not share their locations with the tweets. Location prediction is beneficial for various fields such as targeted advertising, rescue operations during a disaster and recommendation, etc. On the other hand, location prediction from the text of the tweet is a difficult task as it encounters many problems such as spelling mistakes, nonstandard English, etc. In this study, we propose a new set of linguistic features to predict locations from the texts of the tweet more effectively. Two twitter datasets, five popular ML models and five evaluation metrics are used in experiments to explore the impact of the proposed features. The experimental results reveal that proposed set of features outperform as compared to baseline features. In addition, AUC also demonstrates the significance of the proposed features in the form of 59.94 % and 69.65% using MSM2013 and Ritter dataset respectively. The findings indicate that proposed set of features are the most significant indicators and has proved their effectiveness and strong relationship with location prediction.