Cloud Software Code Generation via Knowledge Graphs and Multi-Modal Learning

As cloud computing continues to experience rapid growth, the demand for cloud-native applications is escalating, leading to more complex and diverse developmentrequirements. In this context, automated code generation plays a pivotal role inaccelerating cloud-native application development. However, existing studies oftenoverlook the full potential of multi-program representations, typically focusing ona single representation derived from well-structured programs. This paper pro-poses a novel approach that harnesses a code knowledge graph to enhance codegeneration for cloud computing applications. Our method not only identifies con-nections among related programs based on textual and structural similarities butalso employs multiple representations to enrich program depiction, thus facilitat-ing the generation of cloud-optimized code. We introduce the Aligns features ofdifferent representations Before Fusion and Code Generation (ABCFG) strategy,which utilizes shared-retentive networks (shared-RetNet) and an AST-based Trans-former as encoders. By employing contrastive loss, features are initially aligned andthen seamlessly integrated with natural language representations through cross-attention mechanisms and a multi-layer perceptron (MLP), effectively capturing thediverse modalities required for robust cloud-based software development. To assessour approach, we conducted experiments on two open-source datasets. The resultsdemonstrate that our model, without the necessity for extensive pre-training,outperforms state-of-the-art large language models in BLEU-4, CodeBLEU, andROUGE-L metrics, paving the way for more efficient and intelligent cloud-nativesoftware development.