Which Features are Learned by CodeBert: An Empirical Study of the BERT-based Source Code Representation Learning
This work identifies a critical flaw in current code representation methods, which is incremental for improving code understanding in software engineering.
The study investigated the limitations of BERT-based models in source code representation learning, finding that they rely heavily on programmer-defined names rather than understanding code logic, as demonstrated through empirical experiments.
The Bidirectional Encoder Representations from Transformers (BERT) were proposed in the natural language process (NLP) and shows promising results. Recently researchers applied the BERT to source-code representation learning and reported some good news on several downstream tasks. However, in this paper, we illustrated that current methods cannot effectively understand the logic of source codes. The representation of source code heavily relies on the programmer-defined variable and function names. We design and implement a set of experiments to demonstrate our conjecture and provide some insights for future works.