Geolocation Prediction in Social Media Data by Finding Location Indicative Words

Proceedings of COLING 2012: Technical Papers, pages 1045–1062, COLING 2012, Mumbai, December 2012. Geolocation Prediction in Social Media Data by Finding Location Indicative Words HAN Bo1,2 Paul COOK1 Timothy BALDW IN 1,2 (1) University of Melbourne (2) NICTA hanb@student.unimelb.edu.au, paulcook@unimelb.edu.au, tb@ldwin.net Abstract Geolocation prediction is vital to geospatial applications like localised search and local event detection. Predominately, social media geolocation models are based on full text data, including common words with no geospatial dimension (e.g. today) and noisy strings (tmrw), potentially hampering prediction and leading to slower/more memory-intensive models. In this paper, we focus on finding location indicative words (LIWs) via feature selection, and establishing whether the reduced feature set boosts geolocation accuracy. Our results show that an information gain ratio- based approach surpasses other methods at LIW selection, outperforming state-of-the-art geolocation prediction methods by 10.6% in accuracy and reducing the mean and median of prediction error distance by 45km and 209km, respectively, on a public dataset. We further formulate notions of prediction confidence, and demonstrate that performance is even higher in cases where our model is more confident, striking a trade-off between accuracy and coverage. Finally, the identified LIWs reveal regional language differences, which could be potentially useful for lexicographers. Keywords: Social Media, Geolocation, Feature Selection. 1045 1 Introduction With the ever-growing popularity of social media, massive volumes of user-generated data are produced everyday, e.g. in the form of Twitter messages (tweets) and Facebook updates.1 This data provides many new opportunities and challenges for natural language processing. One such challenge is geolocation prediction: predicting the geolocation of a message or user based on their social media posts. In this paper, we focus on user-level geolocation based on the aggregated body of tweets from a user, and estimate the user’s location at the city level. As is well established in previous work (Cheng et al., 2010; Wing and Baldridge, 2011; Kinsella et al., 2011), it is reasonable to assume that user posts in social media reflect their geospatial locum, because lexical priors differ from region to region. For example, a user in London is much more likely to talk about Piccadilly and British than a user in New York or Beijing. That is not to say that those terms are uniquely associated with London, of course: British could be used by a user outside of the UK to discuss something relating to the UK. However, the use of a range of such terms with high relative frequency is strongly indicative of the fact that a user is located in London. Our objective in this work is to automatically identify “location indicative words” (LIWs), that is words that implicitly or explicitly encode an association with a particular location. To this end, we refine the geolocation task in Section 3, and explore the impact of LIWs on user geolocation. Our contributions are as follows: (1) we apply feature selection methods for automatically learning LIWs, and show that both accuracy and efficiency in geolocation are vastly improved using the resultant feature set, achieving …

Content truncated to 500 words