In safety management, process models are widely adopted to guide the implementation of safety-critical protocols. Extracting process models from text can help to enhance tasks such as compliance verification and risk assessment in safety management. However, existing process model extraction (PME) methods do not fully exploit syntactic dependency information in text structures, leading to limited modeling of contextual semantic relations. In this paper, we propose a dependency tree-based multi-task learning method for process models extraction (DT-MPME). It can effectively capture structural information within dependency tree and extract multi-granular features. Besides, an attention-weighted feature fusion layer is designed for efficiently integrating global, local, and hierarchical text features. The experimental results show that the proposed method outperforms existing mainstream approaches in terms of classification accuracy on two widely used datasets.