Towards Learning Generalizable Code Embeddings Using Task-agnostic Graph Convolutional Networks

Code embeddings have seen increasing applications in software engineering (SE) research and practice recently. Despite the advances embedding techniques applied SE research, one of main challenges is their generalizability. A recent study finds that code may not be readily leveraged for downstream tasks are particularly trained for. Therefore, this article, we propose GraphCodeVec , which represents source as graphs leverages Graph Convolutional Networks to learn more generalizable a task-agnostic manner. The edges graph representation automatically constructed from paths abstract syntax trees, nodes tokens code. To evaluate effectiveness consider three benchmark (i.e., comment generation, authorship identification, clones detection) used prior benchmarking add new classification, logging statements prediction, defect prediction), resulting total six considered our evaluation. For each task, apply learned by four baseline approaches compare respective performance. We find outperforms all baselines five out tasks, its performance relatively stable across different datasets. In addition, perform ablation experiments understand impacts training context extracted trees) model Networks) on generated embeddings. results show both can benefit producing high-quality while improvement robust Our findings suggest future using graph-based deep learning methods capture structural information tasks.