arXiv-2022 A Survey of Deep Learning Models for Structural Code Understanding

2022-05-01 · arXiv· Software · code-understanding

Paper: A Survey of Deep Learning Models for Structural Code Understanding

Survey: Deep Learning Models for Structural Code Understanding

Abstract

Methods and applications for code understanding continue to proliferate. This survey groups recent code-understanding work into sequence-based and graph-based models. It also covers metrics, datasets, and downstream tasks, and offers recommendations for future research on structural code understanding.

Introduction

Structural modeling of code: how to represent structural information in code effectively, and how to select structural signals that help a given downstream task.

General-purpose code representation learning: how to learn code representations that transfer beyond a single programming language.

Task-specific adaptation: how to choose architectures tailored to downstream tasks, handle task-specific data, and adapt models under few-shot learning, transfer learning, and cross-language settings.

Preliminary

Structures in code

Lexical analysis yields natural code sequences (NCS); syntactic analysis yields abstract syntax trees (AST); semantic analysis and intermediate-code generation further yield control-flow graphs (CFG) and data-flow graphs (DFG).