Identifying Euroscepticism Using a Text-As-Data Approach: An Experimental Study Employing Parliamentary Speeches

Abstract

This study proposes a comprehensive measure of Euroscepticism through textual analysis of parliamentary speeches. This innovative approach offers direct insights into politicians' attitudes towards European integration, enabling quantification of the intensity of Eurosceptic sentiments. We leverage advanced Natural Language Processing (NLP) techniques, including supervised classification and transfer learning. The objectives of this research are three-fold: A) to introduce a novel method for measuring Euroscepticism using advanced NLP techniques and machine learning models, b) to compare the predictive performance of these various techniques for the task at hand, and c) to mitigate, to the extent possible, the inherent difficulties and limitations of crowdsourced data annotation when the process requires expert knowledge. Results indicate that in the absence of a humanly curated annotated dataset, GPT outperforms other approaches (F1 score 0.74). On the contrary, when such a dataset exists, Few-Shot classification proves to be the optimal solution (F1 0.87). These findings pave the way for a more accurate assessment of Euroscepticism and potentially other political phenomena, through the innovative use of NLP techniques.