ABSTRACT
Social media platforms are crucial for understanding public opinion about policy issues. In this regard, detecting stance in Twitter posts is a vital tool. In this study, we built a corpus of tweets from 2020 and 2021, annotated with stance towards COVID-19 vaccines and vaccination, and test BERTimbau as a way to automatically detect stance in such tweets. Our model reached 86% accuracy in 2020, 77% in 2021, and 79% in the combined 2020/2021 set. Our results also highlight the time-dependent nature of data distribution and, as a consequence, stance classification. Therefore, this research also contributes to the field by shedding some light on the existing methodological challenges in analyzing complex public policy debates over time.