Resolving ambiguity in code refinement via conidfine: A conversationally-Aware framework with disambiguation and...

Area of Science:

• Software Engineering
• Natural Language Processing
• Artificial Intelligence

Background:

• Code refinement is crucial in software development.
• Ambiguous review comments hinder developer understanding and task completion.
• Conversational data between developers and reviewers can resolve ambiguity.

Purpose of the Study:

• To propose a novel method for addressing ambiguous code review suggestions using conversational data.
• To enable models to autonomously detect ambiguity and extract relevant conversational segments for clarification.

Main Methods:

• Developed an Ambiguous Discriminator using multi-task learning for ambiguity classification and confusion point generation.
• Introduced a Type-Driven Multi-Strategy Retrieval Framework to extract actionable information based on ambiguity type.
• Constructed a semantic auxiliary instruction library for precise alignment between review suggestions and conversation context.

Main Results:

• Significantly enhanced the performance of state-of-the-art models (TransReview, T5-Review, CodeT5, CodeReviewer, ChatGPT) on code refinement tasks.
• Demonstrated the effectiveness of conversational information in improving model performance on fine-grained ambiguity resolution.
• Human evaluations confirmed the accuracy of ambiguity detection and the correctness of generated confusion points.

Conclusions:

• This research is the first to address ambiguous review suggestions in code refinement.
• The proposed method effectively resolves ambiguity by leveraging conversational context.
• The findings provide a foundation for future research in improving code review clarity and effectiveness.