Abstract: Span extraction reading comprehension is an important task in machine reading comprehension. In practical applications, many of the answers to reading comprehension problems are composed of multiple discontinuous spans of text. In contrast to existing sequence tagging method, we propose a multi-task joint learning method combining dynamic span number prediction and sequence labeling. Dynamically predicting the number of spans can capture the type of multi-span question, and the sequence tagging can effectively locate multiple answer spans. At the same time, we construct an adversarial training and two-stage training approach, which can further enhance the generalization ability of the model. Experimental results on the Challenge of AI in Law 2021 (CAIL2021) and QUOREF datasets show that our method achieves better results than the baseline models.