Are We Ready for this Disaster? Towards Location Mention Recognition from Crisis Tweets

Proceedings of the 28th International Conference on Computational Linguistics, pages 6252–6263 Barcelona, Spain (Online), December 8-13, 2020 6252 Are We Ready for this Disaster? Towards Location Mention Recognition from Crisis Tweets Reem Suwaileh∗ Muhammad Imran† Tamer Elsayed∗ Hassan Sajjad† ∗Computer Science and Engineering Department, Qatar University {rs081123, telsayed}@qu.edu.qa †Qatar Computing Research Institute, Hamad Bin Khalifa University {mimran, hsajjad}@hbku.edu.qa Abstract The widespread usage of Twitter during emergencies has provided a new opportunity and timely resource to crisis responders for various disaster management tasks. Geolocation information of pertinent tweets is crucial for gaining situational awareness and delivering aid. However, the majority of tweets do not come with geoinformation. In this work, we focus on the task of location mention recognition from crisis-related tweets. Specifically, we investigate the influence of different types of labeled training data on the performance of a BERT-based classification model. We explore several training settings such as combing in- and out-domain data from news articles and general-purpose and crisis-related tweets. Furthermore, we investigate the effect of geospatial proximity while training on near or far-away events from the target event. Using five different datasets, our extensive experiments provide answers to several critical research questions that are useful for the research community to foster research in this important direction. For example, results show that, for training a location mention recognition model, Twitter-based data is preferred over general-purpose data; and crisis-related data is preferred over general- purpose Twitter data. Furthermore, training on data from geographically-nearby disaster events to the target event boosts the performance compared to training on distant events. 1 Introduction Twitter has shown to be an effective medium for gaining situational awareness and performing urgent needs assessment of the affected population during sudden onset disasters (Vieweg, 2012; Hughes and Palen, 2009; Purohit et al., 2018). Furthermore, the platform often breaks events and thus considered a low-latency source for timely access to information when other traditional sources are not available. De- spite these advantages, one major issue that hinders the usability of Twitter data is the lack of geolocation information. Only 1-3% of tweets has GPS-coordinates (Huang and Carley, 2019). Response authorities and humanitarian organizations heavily rely on geolocation information for both situational awareness and response tasks. While extensive research has been conducted on processing tweets for humanitarian aid, limited focus has been given to infer and extract geolocation information from them. In this work, we focus on extracting toponyms, i.e., place or location names from tweets. We refer to this as a Location Mention Recognition (LMR) task. Two main factors that influence the robustness of a LMR system are: (i) the dataset used to train the classifier, and (ii) the learning model. In this work, we explore how the choice of a training dataset influences the performance of a LMR system in the domain of humanitarian crises where the cost and time of acquiring training data should be minimized. For this purpose, one well-established approach is to use a standard Named Entity Recognition (NER) system trained on a general-purpose NER …

Content truncated to 500 words