De’coding concepts: Automatic characterization of populist moments on Twitter

• Main Author: Mateus, Vasco Lopo Hipólito Branco Ribeiro
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10071/30527
• topic: Populismo -- Populism; Text-as-data; Hype machine; Portuguese political actors; Processamento de linguagem natural - -- NLP Natural language processing; BERT; Texto-como-variáveis; Máquina dos media; Políticos Portugueses

In 1995, McChesney stated that the political economy of communication studies was at a breaking point. In the following two decades, new communication platforms and the heightened commercialization of media systems brought political discourse to the forefront of political debates and public interests. Portugal, which was once seen as resilient against the populist movements throughout the 2010s, has exhibited a noteworthy shift towards embracing these trends in recent election cycles. Simultaneously, the country appears to be aligning with foreign trends in its adoption of social media for political strategies. This parallel “catching-up” echoes a broader phenomenon that is also prevalent in social sciences' inertia to adopt contemporary methodologies and more advanced data techniques, despite the ever-expanding realm of information available. In our research, we developed a machine-learning model with the objective of using deep learning techniques for binary classification to create a robust method capable of discerning complex concepts, like populism and treating them as variables. The model’s accurate performance on classifying tweets from Portuguese political actors, over the last two election cycles, allowed us to establish some primary features of populist moments through a complementary fundamental analysis. The model employed not only serves as a strong foundation for future research within populism but also demonstrates notable adaptability to other possible subjects within social sciences. Through different guidelines and objectives in the labeling process, our model's versatility represents an initial stride towards streamlining this kind of research, effectively reducing the labor-intensive demands currently associated with analyzing large datasets.