What do pre-trained code models know about code?

TL;DR

This paper investigates what pre-trained code models understand about source code by using diagnostic probing tasks to analyze their knowledge of code properties across different models.

Contribution

It introduces four probing tasks to evaluate the understanding of code in various pre-trained models and compares their capabilities across different layers and properties.

Findings

GraphCodeBERT performs most consistently overall.

BERT performs surprisingly well on some code tasks.

Probing reveals model deficiencies and layer-specific information.

Abstract

Pre-trained models of code built on the transformer architecture have performed well on software engineering (SE) tasks such as predictive code generation, code summarization, among others. However, whether the vector representations from these pre-trained models comprehensively encode characteristics of source code well enough to be applicable to a broad spectrum of downstream tasks remains an open question. One way to investigate this is with diagnostic tasks called probes. In this paper, we construct four probing tasks (probing for surface-level, syntactic, structural, and semantic information) for pre-trained code models. We show how probes can be used to identify whether models are deficient in (understanding) certain code properties, characterize different model layers, and get insight into the model sample-efficiency. We probe four models that vary in their expected knowledge of code properties: BERT (pre-trained on English), CodeBERT and CodeBERTa (pre-trained on source code, and natural language documentation), and GraphCodeBERT (pre-trained on source code with dataflow). While GraphCodeBERT performs more consistently overall, we find that BERT performs surprisingly well on some code tasks, which calls for further investigation.