Event detection on social networks has emerged as a fast-growing research area in artificial intelligence, driven by the surge of real-time user-generated content. There are many challenges, such as short and noisy messages, multi-language messages, temporal constraints for emerging narrative happenstances, and the detection of weak-signal or unpredicted events. This paper offers a synthesis of recent contributions, including both surveys and original research works. We review various approaches, such as traditional machine learning (ML) methods, Deep Learning (DL), Transformer-based models, graph neural networks, contrastive learning, prompt-based learning, and few-shot. Benchmark data like Twitter, Weibo, CrisisLex, MAVEN, and PHEME are considered, along with standard evaluation measures like F1-score, precision, recall, NMI, and ARI. Common limitations such as high computational cost, non-generalizability, and poor explainability are also discussed. A unified taxonomy is proposed to organize and aggregate the rich diversity of contributions in this research area. It has the central constituents of event detection on social networks, i.e., significant challenges, methodological frameworks, datasets, evaluation metrics, and the steps of the detection pipeline: data preprocessing, representation, detection, aggregation, and evaluation. This broad structure provides a solid foundation for advancing robust, scalable, and multilingual event detection systems for the dynamic and noisy nature of social networks.