Towards Location Approximation of Disaster Related Tweets: Predicting Latitude-Longitude Level Coordinates of Tweets Using Latent Semantic Analysis

Various studies have presented ways of approximating the location of tweets through performing text mining techniques on the contents of tweets. Methodologies of this type create word density based models for each computationally or geopolitically defined region. Incoming tweets are then compared against the text models of the defined regions to find which region the new tweet most probably comes from. These studies assume that each region can be distinguished by the distribution of tokens in each respective regional word model. This study proposes a new approach to approximating the location of tweets under a similarly valid assumption?that semantically similar tweets occur in geographically close locations. This is done through modeling text content of tweets in a semantic space and geotagging tweets down to the granularity of latitude and longitude coordinates using a new algorithm for location approximation. Through using a common method for semantic modeling such as Latent Semantic Analysis, this study provides a baseline for semantic similarity based location approximation from which more state of the art and tweet specific semantic modeling methods can develop from. Moreover, this pioneering method for latitudelongitude granularity location approximation paves the way for more innovative research that deviates from the commonly implemented location approximation to defined regions.