SocialPET: Socially Informed Pattern Exploiting Training for Few-Shot Stance Detection in Social Media

Social media platforms offer a goldmine to collect data for analyzing opinions and attitudes expressed by large numbers of users (Alturayeif, Luqman and Ahmed, 2023a), which because of the large volume requires development of automated tools to support stance detection from textual content (AlDayel and Magdy, 2021; Küçük and Can, 2020). Stance detection is the process of automatically determining a user's viewpoint or position as favor or against regarding a particular subject of interest, often known as the target (Alturayeif, Luqman and Ahmed, 2023b; Khiabani and Zubiaga, 2023). In particular, there is a notable interest within the Natural Language Processing (NLP) community for examining the identification of attitudes expressed towards political figures on Twitter (Mohammad, Kiritchenko, Sobhani, Zhu and Cherry, 2016; Sobhani, Inkpen and Zhu, 2017). Much of the previous research in stance detection has generally assumed that there is sufficient training data to develop a model that determines the stance towards a particular target. In a realistic scenario, however, one may have access to limited training data when new targets emerge for which sufficient data could not be labeled.