Advancing stance detection and fine-grained content analysis for socially relevant domains

Social media platforms play a pivotal role in shaping public discourse on critical societal issues such as climate change and public health. However, the same platforms that enable civic engagement also facilitate the rapid spread of misinformation, polarization, and harmful narratives. In this context, stance detection, the task of determining an author’s position toward a specific target, emerges as a key computational tool for understanding online opinion dynamics. Beyond stance classification, fine-grained analysis of online discourse is essential for assessing content relevance, credibility, emotional framing, and ethical considerations. Together, these tasks provide a nuanced understanding of how individuals engage with societal issues, enabling informed interventions and responsible content moderation. This thesis advances these areas through a series of incremental contributions and focuses on the three principal contributions. The first contribution advances in-target stance detection for climate change, a critical societal challenge. We develop novel datasets and progressively refined multi-task learning frameworks. These models leverage auxiliary signals such as sentiment, fine-grained emotion, toxicity, and multimodal cues like emojis to capture nuanced stances, including implicit, sarcastic, and emotionally framed content. Each successive model addresses limitations of the previous one, producing a cohesive progression of improvements that enhance performance and generalizability across the climate change domain and benchmark stance datasets. The second contribution focuses on interpretable zero-shot stance detection, enabling models to generalize to unseen targets where labeled data are unavailable, addressing the limitations of traditional in-target stance models. We propose rationale-based two-stage pipelines and ranking-based frameworks, which combine human-understandable rationales with large language model reasoning. These approaches achieve robust zero-shot performance while providing transparent and inherent interpretability, addressing both technical and societal needs for trustworthy AI in sensitive domains. The third contribution shifts toward fine-grained and ethical content analysis in socially critical domains, including COVID-19 and climate change. We introduce novel tasks of relevance detection and information categorization, and cognitively inspired multi-task frameworks that integrate ethical and empathetic reasoning with supervised learning. Additionally, a two-stage LLM-based moderation framework enables the classification of adaptation, resilience, and denial climate content while promoting credible and ethically grounded climate online discourse. Together, these contributions establish a unified trajectory from domain-specific stance detection to interpretable, generalizable, and ethically aware content analysis, laying the foundation for AI for social good that supports informed, nuanced, and trustworthy engagement with online public discourse.