Temporal Modelling of Geospatial Words in Twitter

Temporal Modelling of Geospatial Words in Twitter Bo Han1, Antonio Jose Jimeno Yepes2, Andrew MacKinlay2,3, Lianhua Chi2 Hugo AI1 IBM Research Australia2 University of Melbourne3 bhan@hugo.ai, {antonio.jimeno, admackin, lianhuac}@au1.ibm.com Abstract Twitter text-based geotagging often uses geospatial words to determine locations. While much work has been done in word geospatiality analysis, there has been little work on temporal variations in the geospa- tial spread of word usage. In this paper, we investigate geospatial words relative to their temporal locality patterns by fitting periodical models over time. The model jointly captures inherent geospatial local- ity and periodical factor for a word. The resultant factorisation enables better un- derstanding of word temporal trends and improves geotagging accuracy by only us- ing inherent geospatial local words. 1 Introduction Automatically inferring geographical location of social media data has become increasingly pop- ular, as geospatial information plays a vital role in applications such as advertising, influenza de- tection and disaster management. Due to a lack of abundant reliable geospatial information (Cheng et al., 2010), various text-based geotag- ging methods have been proposed (Cheng et al., 2010; Eisenstein et al., 2010; Wing and Baldridge, 2014).1 The main idea is to leverage geospatial words such as dialects and location names em- bedded in Twitter text to infer geographic loca- tions. For instance, yinz is primarily used in Pitts- burgh, #auspol is a popular hashtag in Australia, and @TransLink is frequently mentioned by Van- couver users. Social media data often comes as a stream, and its contents and topics change over time. This im- plies geospatial words in social media can be fur- 1Indeed there is a way that a user can turn on their location sharing options which offers accurate location information, however, the ratio of such users are pretty low (Cheng et al., 2010). ther distinguished based on their temporal patterns of locality (i.e. how location-indicative a word is). Some time-invariant geospatial words (e.g. CalTech) are consistently associated with a loca- tion, while other words like #ALTA216 are only transiently associated with a location (e.g. during the conference). In this paper, we investigate the word local- ity pattern over time. First, we bin streaming tweet data into a series of sliding time windows and calculate the locality estimators in each win- dow. The time-indexed locality estimators are then fed into periodical models which jointly capture the inherent geospatial locality and periodical fac- tor. We show that by removing the periodical fac- tor, we can obtain improved geotagging accuracy. Furthermore, we demonstrate the utility of fitted model parameters in explaining some intuitive ob- servations for reoccurring words. 2 Background Text-based geotagging is often formulated as a classification task (Cheng et al., 2010) which in- volves predicting a location from a set of pre- defined geographical partitions. It exploits words in tweets on the grounds that some of them carry geographical information. The accuracy is of- ten measured by the agreement of the predicted locations with true oracle locations (Wing and Baldridge, 2011). Earlier work in geotagging exploits language …

Content truncated to 500 words