CRaDLe: Deep code retrieval based on semantic Dependency Learning

Area of Science:

• Computer Science
• Software Engineering
• Artificial Intelligence

Background:

• Code retrieval is crucial for programmers reusing open-source code snippets.
• A significant challenge is the semantic gap between natural language queries and code.
• Neural networks are increasingly used to learn semantic matching for code retrieval.

Purpose of the Study:

• To propose CRaDLe, a novel approach for code retrieval using statement-level semantic dependency learning.
• To address the semantic gap in code retrieval by incorporating structural code information.

Main Methods:

• CRaDLe distills code representations by fusing statement-level dependency and semantic information.
• It learns a unified vector representation for code and description pairs.
• The approach models the matching relationship between queries and code snippets.

Main Results:

• Comprehensive experiments on real-world datasets demonstrate CRaDLe's effectiveness.
• The proposed approach accurately retrieves relevant code snippets for given queries.
• CRaDLe significantly outperforms state-of-the-art code retrieval methods.

Conclusions:

• Statement-level dependency information is valuable for capturing code semantics in retrieval.
• CRaDLe offers a promising new direction for improving code retrieval systems.
• The method effectively bridges the semantic gap, enhancing retrieval performance.