Mayank Kejriwal from the University of Southern California, discusses the role innovative technology can play when it comes to natural disasters and other crises
According to the United Nations, recent years have witnessed hundreds of millions of people affected by natural disasters such as earthquakes and hurricanes, with tens of millions displaced due to conflict alone. Unsurprisingly, there is a funding shortfall in holistically addressing such crises, especially in developing countries. Low cost, innovative technology can play an important role during such crises; for example, by giving frontline workers situational awareness during, and in the immediate aftermath of, events such as the Nepal earthquake of 2015 and the ongoing COVID-19 pandemic.
Predictive analytics
By processing and analysing tens of thousands of messages (including SMS and fragments from social media channels such as Twitter) in real-time, a situational awareness system could use predictive analytics to help aid agencies and local governments decide where to deploy critical resources such as medicine, food, clean water and shelter for maximal effectiveness. Similarly, by deriving useful signals from the grassroots-level data broadcast by ordinary citizens and frontline workers via microblogs and texts in a crisis zone, rescuers are equipped with the requisite knowledge needed to devise sophisticated strategies both for imminent rescue attempts and for longer-term assistance.
COVID-19
With appropriate privacy safeguards, such data could be used in the current COVID-19 crisis to understand which communities are most deeply affected, and where a shortfall of supplies and virus outbreaks are likely to occur in upcoming time periods. In summary, accurate prediction makes preparation possible, especially when resources are constrained, and with appropriate preparation, preventive measures can be used to mitigate the devastating impacts of a crisis.
Building such a situational awareness system is challenging since it must necessarily be adaptive, capable of deriving signals from large quantities of noisy data, and amenable to use by non-technical frontline workers and first responders. Our group has been actively involved in using artificial intelligence (AI) research to build just such a system over a period of more than three years. We were a leading participant, for example, in the recently concluded THOR (which stands for the Text-enabled Humanitarian Operations in Real Time) project that was funded by DARPA (Defence Advanced Research Projects Agency) in the United States. THOR was a highly collaborative project, involving both academics and industry from across the country, and leading to impressive advances in machine translation and other research in automatic language understanding in just a few short years.
AI clustering
Our group made advances in an important area of AI called clustering, which is the science of automatically grouping many data points (e.g., text messages) to reveal structure in the data. Our algorithm was automatically able to detect urgent situations such as food, evacuation or medical needs emerging on the ground in real-time, and our system was tested on datasets derived from challenging disasters such as the 2015 Gorkha earthquake in Nepal, which killed nearly 9,000 people.
The key idea is as follows. During a humanitarian crisis, THOR sorts through (and does predictive analytics on) incoming data from social media and text in real-time and derives the necessary signals for enabling rapid deployment of support. Clearly, this would be an impossible task for humans to complete without significant help (since thousands of messages are streaming in by the minute), allowing AI technologies such as language understanding to offer a solution. Key signals such as the location and the nature of the need (for example, a food crisis versus an impending collapse of a building) are extracted using algorithms that can then direct resources to casualties and responders. Importantly, these messages can be clustered according to their urgency, with help then targeted to areas that most need it.
Critical decision making
We attempt to foster trust in this system by showing relevant data in a visual, interpretable format, and leaving critical decisions in the hands of human operators. We developed advanced, interactive visualization tools, using state-of-the-art machine learning technology, to support such contextual interpretations. Our tools use open-source software that only needs a web browser, such as Chrome or Firefox, to be deployed, requiring minimal setup and technical knowledge. For example, the SAVIZ tool (pictured) developed by our group, can be run on a Web browser and allow decision-makers to get a quick and broad overview by scanning messages automatically grouped by situational needs (e.g., food and medicine).
In the long-term, our research will be used to inform humanitarian efforts, by making it more efficient and effective, bringing aid to those who need it most, when they need it most. Our systems are also designed to be highly adaptive and to be deployable in novel crises. Our active testing of the system on the ongoing COVID-19 pandemic, for example, is demonstrating great promise. Specifically, we have been deriving situational signals from global COVID-19 tweets that have been collected by a group in our institute since January 2020 to better understand the socioeconomic consequences of, and public reaction to, the crisis.
Please note: This is a commercial profile
References
- Kejriwal, M., Gilley, D., Szekely, P., & Crisman, J. (2018, April). Thor: Text-enabled analytics for humanitarian operations. In Companion Proceedings of The Web Conference 2018 (pp. 147-150).
- Kejriwal, M., Peng, J., Zhang, H., & Szekely, P. (2018, October). Structured Event Entity Resolution in Humanitarian Domains. In International Semantic Web Conference (pp. 233-249). Springer, Cham.
- Kejriwal M., & Zhou P. (2019) SAVIZ: Interactive Exploration and Visualization of Situation Labeling Classifiers over Crisis Social Media Data, International Conference on Advances in Social Networks Analysis and Mining, Vancouver, Aug 27-30, pp705-708.
