Project name: Artificial Intelligence for Disaster Response (AIDR). For a more recent update, please click here.
My Crisis Computing Team and I at QCRI have been working hard on the Twitter Dashboard for Disaster Response. We first announced the project on iRevolution last year. The experimental research we’ve carried out since has been particularly insightful vis-a-vis the opportunities and challenges of building such a Dashboard. We’re now using the findings from our empirical research to inform the next phase of the project—namely building the prototype for our humanitarian colleagues to experiment with so we can iterate and improve the platform as we move forward.
Manually processing disaster tweets is becoming increasingly difficult and unrealistic. Over 20 million tweets were posted during Hurricane Sandy, for example. This is the main problem that our Twitter Dashboard aims to solve. There are two ways to manage this challenge of Big (Crisis) Data: Advanced Computing and Human Computation. The former entails the use of machine learning algorithms to automatically tag tweets while the latter involves the use of microtasking, which I often refer to as Smart Crowdsourcing. Our Twitter Dashboard seeks to combine the best of both methodologies.
On the Advanced Computing side, we’ve developed a number of classifiers that automatically identify tweets that:
- Contain informative content (in contrast to personal messages or information unhelpful for disaster response);
- Are posted by eye-witnesses (as opposed to 2nd-hand reporting);
- Include pictures, video footage, mentions from TV/radio
- Report casualties and infrastructure damage;
- Relate to people missing, seen and/or found;
- Communicate caution and advice;
- Call for help and important needs;
- Offer help and support.
These classifiers are developed using state-of-the-art machine learning tech-niques. This simply means that we take a Twitter dataset of a disaster, say Hurricane Sandy, and develop clear definitions for “Informative Content,” “Eye-witness accounts,” etc. We use this classification system to tag a random sample of tweets from the dataset (usually 100+ tweets). We then “teach” algorithms to find these different topics in the rest of the dataset. We tweak said algorithms to make them as accurate as possible; much like training a dog new tricks like go-fetch (wink).
We’ve found from this research that the classifiers are quite accurate but sensitive to the type of disaster being analyzed and also the country in which said disaster occurs. For example, a set of classifiers developed from tweets posted during Hurricane Sandy tend to be less accurate when applied to tweets posted for New Zealand’s earthquake. Each classifier is developed based on tweets posted during a specific disaster. In other words, while the classifiers can be highly accurate (i.e., tweets are correctly tagged as being damage-related, for example), they only tend to be accurate for the type of disaster they’ve been trained for, e.g., weather-related disasters (tornadoes), earth-related (earth-quakes) and water-related (floods).
So we’ve been busy trying to collect as many Twitter datasets of different disasters as possible, which has been particularly challenging and seriously time-consuming given Twitter’s highly restrictive Terms of Service, which prevents the direct sharing of Twitter datasets—even for humanitarian purposes. This means we’ve had to spend a considerable amount of time re-creating Twitter datasets for past disasters; datasets that other research groups and academics have already crawled and collected. Thank you, Twitter. Clearly, we can’t collect every single tweet for every disaster that has occurred over the past five years or we’ll never get to actually developing the Dashboard.
That said, some of the most interesting Twitter disaster datasets are of recent (and indeed future) disasters. Truth be told, tweets were still largely US-centric before 2010. But the international coverage has since increased, along with the number of new Twitter users, which almost doubled in 2012 alone (more neat stats here). This in part explains why more and more Twitter users actively tweet during disasters. There is also a demonstration effect. That is, the international media coverage of social media use during Hurricane Sandy, for example, is likely to prompt citizens in other countries to replicate this kind of pro-active social media use when disaster knocks on their doors.
So where does this leave us vis-a-vis the Twitter Dashboard for Disaster Response? Simply that a hybrid approach is necessary (see TEDx talk above). That is, the Dashboard we’re developing will have a number of pre-developed classifiers based on as many datasets as we can get our hands on (categorized by disaster type). In addition to that, the dashboard will also allow users to create their own classifiers on the fly by leveraging human computation. They’ll also be able to microtask the creation of new classifiers.
In other words, what they’ll do is this:
- Enter a search query on the dashboard, e.g., #Sandy.
- Click on “Create Classifier” for #Sandy.
- Create a label for the new classifier, e.g., “Animal Rescue”.
- Tag 50+ #Sandy tweets that convey content about animal rescue.
- Click “Run Animal Rescue Classifier” on new incoming tweets.
The new classifier will then automatically tag incoming tweets. Of course, the classifier won’t get it completely right. But the beauty here is that the user can “teach” the classifier not to make the same mistakes, which means the classifier continues to learn and improve over time. On the geo-location side of things, it is indeed true that only ~3% of all tweets are geotagged by users. But this figure can be boosted to 30% using full-text geo-coding (as was done the TwitterBeat project). Some believe this figure can be doubled (towards 75%) by applying Google Translate to the full-text geo-coding. The remaining users can be queried via Twitter for their location and that of the events they are reporting.
So that’s where we’re at with the project. Ultimately, we envision these classifiers to be like individual apps that can be used/created, dragged and dropped on an intuitive widget-like dashboard with various data visualization options. As noted in my previous post, everything we’re building will be freely accessible and open source. And of course we hope to include classifiers for other languages beyond English, such as Arabic, Spanish and French. Again, however, this is purely experimental research for the time being; we want to be crystal clear about this in order to manage expectations. There is still much work to be done.
In the meantime, please feel free to get in touch if you have disaster datasets you can contribute to these efforts (we promise not to tell Twitter). If you’ve developed classifiers that you think could be used for disaster response and you’re willing to share them, please also get in touch. If you’d like to join this project and have the required skill sets, then get in touch, we may be able to hire you! Finally, if you’re an interested end-user or want to share some thoughts and suggestions as we embark on this next phase of the project, please do also get in touch. Thank you!