CS224N Final Project: Detecting Key Needs in Crisis

When a crisis occurs, the world springs into action to try and understand what is happening and what help is required. During these times, twitter has become a key avenue through which to disseminate information. In this paper, we classify tweets into 18 key categories (such as [volunteering] or [injury]) in order to organize the stream of knowledge coming in. We also attempt to predict when a particular category will be tweeted about by leveraging past tweets and categories to predict future tweet categories. Our most successful result, intermixing a Word2Vec model with a standard feed-forward classifier is able to classify the tweets with 59% accuracy. We find that more sophisticated models (including added depth to a feedforward model and an LSTM that uses a sequence of words to predict a tweet) prove to be less effective than a more generic classification methodology, likely because of overfitting on labels that are more common and 2) noise created by combining various types of crises, such as earthquakes and floods, together. We show that our classification model is more successful when restrained to specific types of events, with 70% accuracy on earthquake data, for example. Additionally, we determine that leveraging past history to predict future tweet patterns is moderately successful. We showcase our LSTM methodology which has achieved a 33% accuracy. Lastly, we test out our model on the more recent twitter compilations for two natural disasters: the earthquake in Italy and Louisiana Floods. We show, via qualitative examples and analysis, that our model can likely scale to natural disasters of various forms.