Extracting Geographic Addresses from Social Media using Deep Recurrent Neural Networks

Abstract

The importance of geographical, addresses in people's daily lives cannot be underestimated. People usually use the Internet to search for unfamiliar areas and then use map services to mark locations. Using social media to extract information, particularly geographical addresses, is rapidly increasing worldwide. Social media represents the right choice as a source in identifying the location that people need to find. In this paper, a deep neural network using a Bidirectional Long Short-Term Memory with CRF (BI-LSTM-CRF) model is applied for address extraction. In addition, a Bidirectional Encoder Representations from Transformers (BERT) model is implemented to extract the geographical addresses from Facebook posts. Further, we reveal how to use the BIEO tagging method to apply the sequence labeling technique to Arabic postal address extraction. An Arabic corpus from social media is annotated to evaluate our proposed model. The results show that Arabic postal addresses can be extracted through BI-LSTM-CRF and BERT models with a high F-measure.