Abstract
Recent advances in deep learning (DL) have significantly impacted various research disciplines, including natural language processing and image processing. In software engineering (SE), DL techniques are increasingly employed to automate tasks such as code generation, refactoring, and fault localization. While existing surveys often focus on the underlying neural architectures, this study presents the first task-oriented survey. Covering twelve major SE subareas across the entire software lifecycle—from requirements engineering to developer collaboration—we analyze 601 papers to highlight major advances, available datasets, and future opportunities.
Introduction: A Task-Oriented Perspective
Deep learning has transitioned from a research curiosity to a fundamental driver of innovation in software engineering. However, most literature reviews categorize research by technique (e.g., "how to use a Transformer for code"). This paper addresses the need for a task-oriented perspective, focusing on how DL solves specific problems within the distinct phases of the software development lifecycle. By breaking down the "silos" between sub-disciplines, the authors provide a unified roadmap for the future of DL-based SE.
Methodology and Scope
The authors conducted a systematic analysis of 601 publications from premier SE and AI venues (including ICSE, ASE, NeurIPS, and AAAI) published between 2000 and 2023. These works were categorized into 12 primary software engineering tasks.
As shown in the table above, research effort is not distributed equally. Subareas like Bug finding (114 papers) and Program repair (64 papers) have received significant attention, while earlier lifecycle phases like Requirements engineering and Software refactoring represent emerging frontiers with fewer studies.
The Fuel of DL: Datasets and Benchmarks
The survey serves as a directory for critical benchmarks that enable DL research:
- Code Intelligence: CodeXGLUE and CodeSearchNet act as the "ImageNet" for code, providing multi-language data for various tasks.
- Bug/Repair: Defects4J and ManyBugs remain the gold standards for evaluating automated patching.
- Clone Detection: BigCloneBench is essential for testing a model's ability to understand semantic logic beyond surface-level syntax.
A notable finding is the "data hunger" in Requirements Engineering, where datasets are often proprietary and significantly smaller than the massive repositories available for code search.
Key Insights: Beyond Vanilla Models
The shift from manually defined metrics (like cyclomatic complexity) to learned features is the most significant trend. Key takeaways include:
- Structural Awareness: Code is not just text; it is hierarchical. The most effective models incorporate Abstract Syntax Trees (ASTs) or Graph Neural Networks (GNNs) to capture control and data flow.
- Task Duality: Many SE tasks are mirrors of each other (e.g., Code Generation and Code Summarization). "Dual learning" frameworks allow models to improve by learning from these reciprocal relationships.
- Maintenance Dominance: The bulk of current research focuses on the maintenance phase, particularly bug detection and repair. This leaves a massive opportunity for DL applications in the design and requirements phases.
Results and the LLM Revolution
DL models have achieved unprecedented performance in tasks like Code Clone Detection, particularly in identifying "Type-4" clones (semantically similar but syntactically different). Furthermore, the rise of Large Language Models (LLMs) like ChatGPT and GitHub Copilot has moved automated code generation into the daily workflow of millions of developers. However, these models still face challenges with "hallucinations"—generating code that appears correct but contains logical flaws.
Challenges and Future Directions
Despite the progress, several hurdles remain:
- Interpretability: Developers need to understand why a model flags a bug. Moving away from "black box" models toward interpretable AI is crucial for trust.
- Generalization: Performance often degrades when models are applied to domain-specific or less popular programming languages.
- AI-in-the-loop: The future lies in collaborative systems where AI acts as a partner that can explain its reasoning and adapt based on real-time human feedback.
Summary updated