Improving the CodeGeeX Model Based on the Relative Convolutional Multi-Head Attention Method

In the NL2Code field, the process of accurate natural language text and code generation based on the text content is crucial. Therefore, in order to improve the efficiency and accuracy of the CodeGeeX code generation model, we optimized the text recognition of the model, mainly by speeding up the retrieval of the input text, and then quickly identify and judge based on the keywords in the text. Such improvements aim to improve the speed and accuracy of keyword recognition, reduce subsequent computing efficiency, and finally improve the overall speed and quality of code generation. Finally, the experimental results show that the proposed method improves the identification accuracy compared with the existing methods.