A multi-element approach to location inference of Twitter: A case for emergency response

Abstract

Since its inception, Twitter has played a major role in real-world events-especially in the aftermath of disasters and catastrophic incidents, and has been increasingly becoming the first point of contact for users wishing to provide or seek information about such situations. The use of Twitter in emergency response and disaster management opens up avenues of research concerning different aspects of Twitter data quality, usefulness and credibility. A real challenge that has attracted substantial attention in the Twitter research community exists in the location inference of twitter data. Considering that less than 2% of tweets are geotagged, finding location inference methods that can go beyond the geotagging capability is undoubtedly the priority research area. This is especially true in terms of emergency response, where spatial aspects of information play an important role. This paper introduces a multi-elemental location inference method that puts the geotagging aside and tries to predict the location of tweets by exploiting the other inherently attached data elements. In this regard, textual content, users' profile location and place labelling, as the main location-related elements, are taken into account. Location-name classes in three granularity levels are defined and employed to look up the location references from the location-associated elements. The inferred location of the finest granular level is assigned to a tweet, based on a novel location assignment rule. The location assigned by the location inference process is considered to be the inferred location of a tweet, and is compared with the geotagged coordinates as the ground truth of the study. The results show that this method is able to successfully infer the location of 87% of the tweets at the average distance error of 12.2 km and the median distance error of 4.5 km, which is a significant improvement compared with that of the current methods that can predict the location with much larger distance errors or at a city-level resolution at best.