Abstract: Commonsense Plausibility Estimation evaluates whether textual content aligns with human daily experience and general knowledge. Compared with early methods that relied on rules and knowledge bases, methods based on pre-trained models implicitly encode commonsense knowledge through self-supervised learning and achieve breakthroughs in the generalization and flexibility. This paper systematically reviews these mainstream methods from two perspectives: the knowledge source(the knowledge acquisition path during inference) and the fine-tuning dependency(the sensitivity of methods to labeled data and computational resources during training). Additionally, 48 benchmark datasets are reviewed based on task types and commonsense knowledge categories. Finally, three breakthroughs are highlighted: complex reasoning optimization for zero-shot inference, fusion mechanism of heterogeneous knowledge, and representation and utilization of multi-modal commonsense knowledge.