Analysing Opinion Dynamics during Polarised Referendum Campaigns: What Twitter Data Reveals about the 2015 Greek Bailout Referendum

The Greek 2015 referendum on the bail out was an exceptional event on many fronts. Apart from the result itself -subsequently ignored by the governing elite- the campaign was unprecedentedly short. With only nine days from the announcement of the referendum to the day of the vote, it constituted the shortest referendum campaign in modern history. While the democratic credentials of such a concentrated campaign were rightly critizised on numerous normative grounds, the short campaign is a blessing from a data analysis perspective. Its nine-day duration allows us to gather a universe of social media (Twitter) data related to the event, specifically more than half-a-million Tweets. This paper innovates in three ways: First, we tackle the problem of classifying the huge corpus of tweets in terms of referendum position advocated by individual tweets. We use a combination of expert coding and crowd-sourced annotation of over 5,000 randomly sampled Tweets. This provides our training and test data for building a machine learning classifier, which we apply to the remaining corpus. Last, but by no means least, we use steps one and two above to analyse the opinion dynamics during this highly polarised referendum and highlight the importance of specific events in shifting opinion. We conclude by critically evaluating the prospects of using such social media data for forecasting the results of hotly contested voting events.