Location Extraction from Twitter Messages Using a Bidirectional Long Short-Term Memory Neural Network with Conditional Random Field Model

Abstract

In the context of disaster management, location information is crucial in disaster scenarios to infer the incident location and facilitate disaster relief. In recent years the advent of social media has brought not only great opportunity to enhance disaster management in a crowdsourced perspective, but also a major challenge to interpret the noisy information. A conventional approach to location extraction from texts is Named Entity Recognition (NER), however it shows unsatisfactory performance on informal and colloquial texts such as social media messages, especially for the uncommon place names. To address this issue, we proposed a Bidirectional Long Short-Term Memory (LSTM) Neural Network with Conditional Random Field (CRF) layer to identify geo-entities especially the rarely known local places in social media messages, and the use of orthographic, semantic and syntactic features was explored to achieve best performance. The proposed model was tested on a dataset collected from Twitter, showing promising performance in detecting location information when compared with off-the-shelf NER tools.