Tweet Classification for Crisis Response

Tweet classification for crisis response is a text classification task that aims at identifying
whether a tweet is related to a specific crisis event or not. Humanitarian
organisations that intend to respond to people in need in the early hours of a crisis
suffer from monitoring the massive number of tweets posted in real time. Therefore,
the main objective of tweet classification models for crisis response is to filter the
crisis-related tweets to simplify the work for these organisations. Still, crisis events
have different characteristics, which prevents current models trained on past events
from generalising in identifying tweets from new disasters, which is infeasible to be
manually labelled at the crisis onset. This thesis introduces frameworks under the
umbrella of distant supervision and domain adaptation to minimize the gap or maximize
the similarities between training and testing data from disaster events. The
contributions demonstrate the effectiveness of using automatically labelled training
data from past or emerging events in tweet classification tasks for English and Arabic
crisis tweets. To this end, we propose an automatically labelling framework that
utilises distant supervision via an external knowledge base. Then, we introduce an
approach that unifies our framework and adaptation techniques which automatically
labels incoming tweets from an emerging incident. This approach can be seen
as a robust method to classify unseen English tweets from current events. However,
it has its restrictions when applied to tweets from other languages, especially
if the language comes with limited resources, different text structures, and different
people’s behavior in posting tweets such as Arabic. Hence, we adapt our framework
with significant changes to suit Arabic user-generated posts. Our results for
both English and Arabic tweets show that our original and adaptive approaches
continuously improve the classifier’s performance compared with existing labelling
techniques in different adaptation methods.