The increasing production of service description documents by enterprises and service providers has prompted the need for automated service workflow generation. This paper explores a method that combines large language models (LLMs) and graph neural network (GNNs) to address this challenge. Automatically extracting service workflows from documents also provides a convenient and efficient solution for situations where process mining algorithms cannot be utilized due to the absence of logs. The proposed method begins by employing LLMs to extract activity nodes and conditional nodes from service workflow description documents. These nodes are then organized into an initial graph using chain connections and auxiliary connections provided by the LLMs. Subsequently, an inductive GNN is utilized to analyze node embeddings within the service workflow, enabling the learning of semantic representations and connection rules. This facilitates the discovery of potential connections among nodes, leading to the generation of efficient service workflows. The experimental results demonstrate that the proposed method achieves an F1-Score of 0.818 in predicting node relationships. Furthermore, ablation experiments have been conducted to validate the effectiveness of incorporating both LLMs and GNNs.