Explainable Subjective Stance Classification with SetFit in Political Discourse

Abstract

Stance classification in Natural Language Processing (NLP) is not just an academic exercise but a crucial tool for understanding political discourse and the attitudes underlying political statements. This research addresses the challenge of limited annotated datasets in political science by proposing a practical sentence-level dataset sourced from professional politicians for binary subjective stance classification - support or oppose - using bootstrapping in a SetFit model. The study leverages the Sentence Transformers architecture and incorporates traditional linguistic approaches to enhance explainability. We employ corpus linguistics, tailored lexicons, and lexicogrammatical rules to identify key linguistic features such as positive affect, negative affect, pro polarity, con polarity, certainty, emphatics, doubt, hedges. SHAP analysis quantifies the influence of these features on SetFit model decisions. Our findings demonstrate that iterative bootstrapping significantly enhances the efficacy of few-shot learning in subjective stance classification, and we highlight the importance of linguistic features, particularly pro/con polarity and affective expressions. The StanceSentences dataset and our hybrid analytical approach offer a benchmark for future research, emphasizing the need for nuanced, multi-layered analysis in political discourse.