Enhancing source code classification effectiveness via prompt learning incorporating knowledge features.

Area of Science:

• Computer Science
• Artificial Intelligence
• Software Engineering

Background:

• Pre-trained language models like CodeBERT are used for source code tasks.
• Existing methods use the '[CLS]' token, requiring extra layers and increasing costs.
• These methods may not fully utilize source code and text knowledge, limiting performance.

Purpose of the Study:

• To introduce CodeClassPrompt, a novel text classification technique.
• To leverage prompt learning for extracting rich knowledge from pre-trained models.
• To reduce computational costs and enhance classification accuracy in source code tasks.

Main Methods:

• CodeClassPrompt utilizes prompt learning to extract knowledge from input sequences.
• It eliminates the need for additional neural network layers.
• An attention mechanism synthesizes multi-layered knowledge into task-specific features.

Main Results:

• CodeClassPrompt achieves competitive performance across four distinct source code tasks.
• The technique significantly reduces computational overhead compared to traditional methods.
• Enhanced feature representation and classification accuracy were observed.

Conclusions:

• CodeClassPrompt offers an efficient and effective approach for source code-related text classification.
• Prompt learning is a viable strategy for leveraging pre-trained models in software engineering.
• The method presents a promising direction for reducing computational expenses in AI for code.