Location metadata extraction from Geosocial data of Road Accident using Deep Learning models

Road accident detection and prevention is one of the most challenging problems in the research field revolving around a multitude of problems that need to be addressed. In this context, the research includes creating an automated system to extract geosocial data from multiple sources like news articles and leverage the power of deep learning techniques to classify the collected data as road accidents and extract the useful metadata such as time and date of road accident, location of accident, etc. The news articles are classified into three classes namely non-road accident news, particular road accident, and generic road accident using Word2Vec-based and LSTM-based models. For the combination of LSTM model with K(5)-Fold CV, the road accident news classification model obtains the highest accuracy of about 96%. For location metadata extraction, BERT (Bidirectional Encoder Representations from Transformers)-based baseline models (DBBCS, DTRS, DRBS, DRLS, BLUWwmFS, BLCWwmFS, DBUDS, DBCDS) have also been compared based on various performance metrics such as exact match (EM) and F1-scores. The BLCWwmFS (bert-large-cased-whole-word-masking-finetuned-squad) is found to be the best performing model with an EM and F1-scores of 0.645 and 0.673 respectively for location metadata extraction.