Re_Trans: Combined Retrieval and Transformer Model for Source Code Summarization.

Area of Science:

• Software Engineering
• Artificial Intelligence
• Natural Language Processing

Background:

• Source code summarization (SCS) aids program comprehension and software maintenance.
• Existing retrieval-based methods offer accuracy but struggle with novel code.
• Generative methods provide broader coverage but can lack precision due to limited training data.

Purpose of the Study:

• To develop a hybrid approach for source code summarization that combines the strengths of retrieval-based and generative methods.
• To improve the accuracy and efficiency of generating natural language descriptions for source code.

Main Methods:

• Proposed Re_Trans method: a hybrid approach utilizing retrieval for initial summarization, followed by a generative transformer model if necessary.
• Incorporated Abstract Syntax Tree (AST)-augmented and code sequence-augmented information for enhanced semantic extraction.
• Developed a new SCS retrieval library using a large public dataset.

Main Results:

• Evaluated on a dataset of 2.1 million Java code-comment pairs.
• Demonstrated improved performance over state-of-the-art (SOTA) benchmarks.
• Achieved enhanced effectiveness and efficiency in source code summarization.

Conclusions:

• The Re_Trans method effectively integrates retrieval and generative techniques for superior source code summarization.
• The use of AST-augmented and code sequence-augmented information improves semantic understanding.
• The proposed approach offers a promising solution for accurate and efficient SCS.