Hierarchical Discriminative Classification for Text-Based Geolocation

Abstract

Text-based document geolocation is commonly rooted in language-based information retrieval techniques over geodesic grids. These methods ignore the natural hierarchy of cells in such grids and fall afoul of independence assumptions. We demonstrate the effectiveness of using logistic regression models on a hierarchy of nodes in the grid, which improves upon the state of the art accuracy by several percent and reduces mean error distances by hundreds of kilometers on data from Twitter, Wikipedia, and Flickr. We also show that logistic regression performs feature selection effectively, assigning high weights to geocentric terms.

Huntsville, hospitals, and hockey teams: Names can reveal your location

“…Each region is then considered as a label to train the classifiers. The approach of using k-d tree is also used in Rahimi et al (2015); Han et al (2012) and Wing and Baldridge (2014). See Table 3 for an example of the following methods.…”

“…Some use KL divergence between the distribution of a users words and the words used in each region (Wing and Baldridge, 2011;Roller et al, 2012), regional topic distributions (Eisenstein et al, 2010;Ahmed et al, 2013;Hong et al, 2012), or feature selection/weighting to find words indicative of location (Priedhorsky et al, 2014;Han et al, 2012Han et al, , 2014Wing and Baldridge, 2014).…”

Geographical Entity Annotated Corpus of Japanese Microblogs

“…As mentioned in the previous section, there are several studies for the task. The majority of these studies models the location inference as a multi-class classification problem on the grid over geo-spatial areas or cities on the gazetteer such as GeoNames and DBpedia [12], [16].…”

Evaluating hypotheses in geolocation on a very large sample of Twitter

“…In text-based geolocation, researchers have used KL divergence between the distribution of a users words and the words used in geographic regions (Wing and Baldridge, 2011;Roller et al, 2012), regional topic distributions (Eisenstein et al, 2010;Ahmed et al, 2013;Hong et al, 2012), or feature selection/weighting to find words indicative of location (Priedhorsky et al, 2014;Han et al, 2012aHan et al, , 2014Wing and Baldridge, 2014). Han et al (2012b) showed that information gain ratio is a useful metric for measuring how location-indicative words are.…”