GraphSearchNet: Enhancing GNNs via Capturing Global Dependencies for Semantic Code Search

Code search aims to retrieve accurate code snippets based on a natural language query improve software productivity and quality. With the massive amount of available programs such as (on GitHub or Stack Overflow), identifying localizing precise is critical for developers. In addition, Deep learning has recently been widely applied different code-related scenarios, e.g., vulnerability detection, source summarization. However, automated deep still challenging since it requires high-level semantic mapping between queries. Most existing learning-based approaches rely sequential text i.e., feeding program flat sequence tokens learn semantics while structural information not fully considered. Furthermore, adopted Graph Neural Networks (GNNs) have proved their effectiveness in semantics, however, they also suffer problem capturing global dependencies constructed graph, which limits model capacity. To address these challenges, this paper, we design novel neural network framework, named GraphSearchNet, enable an effective by jointly rich both Specifically, propose construct graphs queries with bidirectional GGNN (BiGGNN) capture local enhance BiGGNN utilizing multi-head attention module supplement that missed The extensive experiments Java Python programming from public benchmark CodeSearchNet confirm GraphSearchNet outperforms current state-of-the-art works significant margin.