This paper proposes a method to generate test scripts automatically from test procedures, in which test procedures are described in natural language. The proposed method addresses the problem of regression testing and the problem of closed-source LLMs. The proposed method includes two phases, namely traditional test script generation phase and LLM training phase. To demonstrate the effectiveness of the proposed method, the experiment is conducted on a variety of Web applications. The experiment shows that the trained CodeT5-base could achieve an average of 95% for AST metric and 86% for Cosine metric on the validation set. In terms of computational cost, using CodeT5-base requires about 1.3s to make a prediction for a step in a test procedure. These results show the potential effectiveness of the proposed method.