Conversational Stance Detection for Rumour and Misinformation Analysis: Datasets, Conversational Modeling, Evaluation Protocols, and Emerging LLM Challenges

Stance detection has become a relevant task for analysing rumours and misinformation in online conversations, because it captures how users react to disputed claims through support, denial, questioning, or neutral commentary. Over the last decade, the field has evolved from feature-engineered classifiers to neural, transformer-based, and graph-aware conversational models. This progress, however, has produced a fragmented literature in which datasets, annotation schemes, evaluation protocols, and application settings are not always directly comparable. This survey reviews conversational stance detection with a focus on rumour and misinformation analysis. It examines benchmark resources such as PHEME and RumourEval, discusses the evolution of modelling approaches, and analyses evaluation practices related to cross-event testing, class imbalance, data leakage, and reproducibility. Particular attention is devoted to conversational structure, context modelling, and the methodological implications of neutral and ambiguous stance categories. Beyond message-level prediction, the survey discusses how stance signals are used for higher-level conversational analysis, including rumour verification, uncertainty monitoring, belief consolidation, and early warning scenarios. It also identifies persistent challenges related to cross-domain generalization, interpretability, multilingual coverage, benchmark saturation, and robustness of downstream measurements. The survey also discusses the transition from benchmark-oriented stance classification toward trustworthy conversational measurement and uncertainty-aware misinformation analysis. By organizing and critically discussing the literature, the survey provides a structured reference for researchers working on conversational misinformation analysis, social computing, and trustworthy natural language processing.