Abstract:In order to improve the efficiency of software development, more options are provided for the develop-ment of software system. At present, many open-source software systems are often written in multiple programming languages, but there are usually associations and invocation relationships between code units in different programming languages, and the resulting security vulnerabilities are more common in the actual environment. The existing vulnerability detection technology mainly focuses on the feature of a single programming language, and it is difficult to effectively detect the vulnerabilities of mixed pro-gramming language software projects. Based on the idea of deep learning model fusion, DL-HLVD(Deep Learning-based Hybrid Language Source Code Vulnerability Detection) is proposed. DL-HLVD first uses the BERT layer to convert the code text into a low-dimensional vector, then captures the con-text features as the input of the Bidirectional Gated Recurrent Unit (BGRU), and finally uses the Condi-tional Random Field (CRF) to capture the dependencies between adjacent labels. The deep learning mod-el is used to recognize named entities for functions of different types of programming languages in mixed language software, and then reconstructs them with program slicing results to reduce the loss of syntax and semantic information in the process of code representation. The comprehensive experimental results on the SARD and CrossVul datasets show that the comprehensive recall rate of DL-HLVD on the two types of vulnerability datasets is 95.0%, and the F1 value reaches 93.6%, which is improved in all indica-tors compared with the latest deep learning methods VulDeePecker, SySeVR, and Project Achilles. The results show that DL-HLVD can improve the comprehensive performance of source code vulnerability detection in mixed language scenarios.