National Leaders’ Twitter Speech to Infer Political Leaning and Election Results in 2015 Venezuelan Parliamentary Elections

The large adoption of Twitter during electioneering has created a valuable opportunity to monitor political deliberation nationwide. Recent work has analyzed online attention to forecast elections results addressing some limitations of opinion polling. However, the reproducibility of such methods remains a challenge given that most of them rely on the number of political parties or candidates mentions. In this study, we propose a method to infer citizens' political alignment in order to predict elections results. To this end, first we collect 750K tweets posted during 2015 Venezuelan Parliamentary election either inside the Venezuela's bounding box or by its political leaders. Second, we build a dictionary characterizing the political leader's speech applying automated content analysis to our corpus. We show that the automatically generated dictionary is an useful tool to improve the accuracy on political election results prediction tasks. Third, using a data set of 1,000 manually-annotated individuals, we show that a support vector machine (SVM) classifier trained on our political dictionary predicts the user political alignment with 87% of accuracy. Finally, using our political dictionary, we design a simple metric to quantify the political lining reflected in a given tweet. We find that the tweets categorized with this metric reflect election results, with 98.72% of accuracy (1.28% mean squared error).