Imagine Twitter as a location-based service that helps businesses and other organizations to act according to the location data from Twitter messages. Maybe it sounds strange or impossible, but it is a reality. Detecting locations from Twitter messages is in the field of data mining, including the analysis of the texts in a huge set of tweets that may include location names.

A problem occurs when several distinct places have the same name, which produces location ambiguity. Therefore, techniques for extracting the location entities from the Twitter textual content are needed. This issue asks for location disambiguation to remove uncertainty or confusion in naming geographical entities. There are some scoring algorithms for GeoNames search to mine the data with machine learning algorithms, and to derive better heuristics from it. This is needed to improve automated location identification, that is, geoparsing. Hence, what is the best method for this purpose?

The authors extend the information extraction model known as named entity recognition (NER), used for detecting names within data mining of huge text sets, with classification of detected locations. This approach provides the basis for better location disambiguation from Twitter messages. The authors used samples from mobile phone users’ communications via Twitter. It should be noted that the mobile phones of those using social media such as Twitter could also provide locations (geotags). Although this work is based on sample datasets from mobile phone users, the authors are focused on the locations (geotags) mentioned within Twitter text messages. Thus, the scope of the study was the text analysis of the Twitter messages containing location names.

The work includes deep learning scenarios with machine learning algorithms that learn through the classification of location names designed by the authors for this study; the solution is trained by a set of labeled location data from the developed gazetteer. In order to achieve the best results, the authors successfully sampled Twitter messages; they measured locations identified by NER software and additional classification schemas because the NER software can differentiate distinct locations that could bear the same or similar location names. They annotated location entities with location tags using a gazetteer from GeoNames that provides additional data, making it more attractive for business, marketing, or not-for-profit organizations. Thus, to increase the accuracy of the labeler, the labels of nearby geographical entities’ names should be used for structured prediction. Because the conditional random fields model can take context into account, the authors used this model to predict the true locations of the detected entities. Further, they made the identification of location entities clearer by introducing a model that could provide a social media monitoring system with visualization of identified places on a map.

I found this work to be a very valuable contribution to the literature that also provides the starting point for further research in the field of deep learning utilization.