How Social Media Text Analysis Can Inform Disaster Management

The emerging field of crisis informatics (e.g., Palen et al. (2010)) is driven by the insight that, in the digital age, the ability to efficiently access and process huge amounts of unstructured data is crucial to situational awareness, knowledge building, and decision-making of organizations responsible for saving lives and property of people affected by a crisis. Disaster events like hurricane Katrina, 9/11, the Haiti earthquake, or the Central-European Flooding 2013 have demonstrated that there is urgent need to understand how information is shared during a crisis and to improve strategies and technologies for turning information into relevant insights and timely actions. Within crisis informatics, social media offer an interesting new opportunity for improvement of disaster management by providing fast, interactive communication channels and enabling participation of the public (Starbird and Palen 2011). However, social media data are big data in terms of volume, velocity, variety and veracity, and, accordingly, the demands and challenges with respect to the development of appropriate information technologies are especially high.

The paper presents possibilities for social media analysis that arise within a disaster management software that has been developed as part of the Slandail project (Slandail 2014), funded by the European community. Slandail deals with data in different modalities (texts and images) and languages (English, German and Italian) as well as with the integration of cross-lingual and cross-cultural aspects of crisis communications and has a special focus on issues related to the legal and ethical correctness of data use. End-users from Ireland, Germany, and Italy have been involved in the development of the system from design to testing.

Computational methods from NLP offer a wide variety of possibilities for systematically and efficiently searching, filtering, sorting and analyzing huge amounts of data and thereby can enable end-users from disaster management to face the problem of information overload posed by social media. In this section, we describe how interests on the side of disaster management have guided our choice of the methods used and our way to apply them in context of our software.

In this section, we demonstrate main functionalities of our software by means of example. As test data sets we used Facebook and Twitter data that had been created during the Central European flooding in June 2013 in Germany and Austria. For the Facebook flood corpus, we collected data from public pages or groups containing the words ‘Hochwasser’ or ‘Fluthilfe’ in their names via the public API (about 36k messages). For the Twitter flood corpus (about 354k tweets), we retrieved the current version of the research corpus of the QuOIMA project (QuOIMA 2011), that had been collected from the API filtering by disaster-related hash tags as well as by names of manually chosen public accounts connected to disaster management and flood aid (ibid.).

The example use case we present will be built around the topic extraction functionality. The dashboard in Fig. 1 gives an overview of topics and topic proportions for the Facebook flood corpus for the entire period of the event.

By clicking on a name of a topic, it is possible to change to the analysis mode and to close-read or further analyze the messages belonging to this topic. The analysis view is shown in Fig. 2.

In the analysis mode, one could get an overview of the content of messages in a certain topic by showing typical topic words, for instance. Figure 3 includes typical words for the volunteering topic. By touching one of the words with the mouse, its co-occurring terms will get connected to it by edges to form a graph.

The dashboard in Fig. 4 changes temporal perspective and analyses the development of topics over the time of the event for the Twitter flood corpus. Again, clicking on the dots on the graph lines gives access to the messages showing the respective topic at the respective day for inspection or further analysis.

The dashboard in Fig. 5 gives an example of filtering for relevance of messages by number of their retweets. The peak around 20th June is connected to heavy rainfalls and thunderstorms that made alarm levels and subjective worries of the people rise anew but finally did not cause mayor new floodings. By only showing messages retweeted more than 6 times, a disaster manger searching for practically relevant information can filter out precaution and pure worries as ‘noise’.

To get an idea of what is happening at the 20th June to cause these emotional reactions, one could either change to the close-reading modus for the relevant topics or extract a situational overview by the help of differential analysis to show keywords that were significantly more frequent at this day (target corpus) than they had been before (reference corpus), see Fig. 6.

The third possible temporal perspective focuses on changes in topic prominence at one day or interval in comparison with the day or interval before. Figure 7 illustrates a possible outcome of analysis for Facebook.

On basis of this insight into hot topics, one could, again, ask further questions. For instance, one could be interested in the most popular organizations involved in donations in kind, or want to find out locations where many volunteering activities are organized. Figures 8 and 9 reveal the results of the respective analyses.

While our examples were developing from the point of view of the topic aspect (‘what?’), each other aspect can equally well serve as a starting point for filtering and further analysis. For instance, the location aspect (‘where?’) is often of high interest for disaster managers. As before, the temporal perspective can either reveal an overview (as in Fig. 9), significant changes or the temporal development of the aspect. In Fig. 10, geographical hot spots are identified, while Fig. 11 shows the geographical unfolding of the flood event over its lifetime.

In the work presented in this paper we were bridging between the fields of informatics and disaster management in order to design and create a social media text analysis software suitable for information gathering and knowledge acquisition in context of a crisis. Our first focus was on showing which methods can be chosen and how they can be applied in a system as to meet the special interests and requirements on the end-user side. Following this, various example use cases illustrated our general approach and demonstrated the capacity of our software to extract information useful for disaster management from huge collections of social media data. An approach along these lines can help to meet the challenges and make use of the opportunities that digitalization and the rise of social media have brought to disaster management.