A GIS services company wanted to create a national interactive map of accidents, they had a person which read the articles with certain keywords and added manually geolocation data in a digital platform + information about the accidents like the number of dead people, the vehicle type: motorbike, bus, car. The responsible person worked just 7 days a week from 9am to 5pm, and many accidents happen in the weekend and during the night.

We created an NLP solution that identifies the location text in the news, then the text is processed with geocoding software, and added to the digital platform automatically. In order to train the model we used 326 labeled news articles and achieved 97% accuracy in identifying the location text in the news article text. We managed to use so little annotated data because of an algorithm called active learning, which succeeded to show what samples from the unlabeled news articles, the model is confused the most and can not identify the location text, and need to be manually labeled.