Rethinking Stance Detection in NLP: A Review of Progress and Open Challenges

Abstract

The rapid growth of social media and online platforms has made stance detection an important task in Natural Language Processing (NLP). It is widely used to understand public opinions on political, social, and controversial topics. In this study, we present a focused, up-to-date review of stance detection research published between 2021 and 2025. This period is significant due to the widespread use of large pre-trained language models and new learning approaches. We conduct a systematic literature review of peer-reviewed journals and conference papers using a clear search strategy, defined inclusion criteria, and a quality assessment process. A total of 170 articles were initially identified; 100 were selected for quality assessment, and 70 studies were included in the final analysis. Of these, 42 studies were explicitly cited and discussed, while the remaining studies contribute to aggregated analysis and overall trend evaluation. The selected studies are analyzed based on several dimensions, including modeling approaches, target dependency, datasets, evaluation metrics, generalization ability, and language coverage. This structured analysis allows a clear comparison of recent research trends. The study shows that transformer-based models are the most widely used approaches for stance detection, accounting for approximately 60–65% of the reviewed studies. There is also increasing interest in prompt-based large language models (10–15%), graph-based methods (around 10%), and multimodal frameworks (10–15%). Most benchmark datasets are still predominantly focused on English, representing over 70% of the datasets, but recent work has introduced multilingual and low-resource datasets. Macro-F1 is the most commonly used evaluation metric due to class imbalance, with performance improving from approximately 0.60–0.70 to 0.75–0.85 in recent transformer-based models. The review also identifies key challenges, including implicit stance expression, target ambiguity, limited cross-domain generalization, low-resource settings, dataset bias, and limited model explainability. This study summarizes recent advances in stance detection, identifies research gaps, and outlines future directions for developing more robust, generalizable, and ethically responsible stance detection systems.