A Zoom into Ecuadorian Politics: Manifesto Text Classification using NLP
Fernanda Barzallo, María Emilia Moscoso, Margorie Pérez, María Baldeon-Calisto, Danny Navarrete, Daniel Riofrío, Pablo Medina, Susana K. Lai-Yuen · 2023
Political science research on party manifestos helps to understand a candidate’s strategies and proposed actions during electoral campaigns. To achieve this, political scientists classify a manifesto’s sentences and quasi-sentences into seven main domains established by the Comparative Manifesto Project. However, manually coding is a time-consuming and labor-intensive task that can lead to biases. Automatically classifying manifestos has shown to produce good and reproducible annotations. In Ecuador, research on automatic manifesto analysis has been limited. Moreover, there is no large labeled Ecuadorian corpus available for training. Therefore, in this work we develop a Transformer network for automatically analyzing Ecuadorian manifestos using a cross-domain training approach. We implement a fractional factorial experimental design to determine which Transformer model, type of pre-processing operations, and Spanish text data should be used to maximize the accuracy of the classification model. The results show that the DistilBERT architecture trained with Mexico’s and Argentina´s manifestos increase the classification accuracy. Without using an Ecuadorian corpus for training, the implemented DistilBERT achieves a 44% accuracy on the Ecuadorian test set, which has a comparable performance to other models in literature trained with a Spanish corpus.