MPLinker: Multi-template Prompt-tuning with adversarial training for issue-commit Link recovery

2025-01-01

Journal of Systems and Software

AI Summary

Existing approaches fail to fully exploit pre-trained language models (e.g., CodeBERT) for recovering links between issue reports and code commits, primarily due to insufficient cross-modal semantic alignment and poor robustness under noisy inputs. Method: We propose a multi-template prompt-tuning framework featuring (i) a novel collaborative multi-template prompt mechanism to enhance semantic alignment across heterogeneous modalities, and (ii) a lightweight FGSM-based adversarial perturbation strategy to improve noise robustness. Additionally, we integrate contrastive learning with a dual-encoder architecture to boost generalization in low-resource settings. Contribution/Results: Evaluated on multiple open-source project datasets, our method achieves an F1 score of 89.7%, outperforming state-of-the-art methods by an average of 4.2%. Results demonstrate superior accuracy, stability, and practical applicability for issue–commit link recovery.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Problem

• Pre-trained Language Models
• Code Issue-Commit Link Recovery
• Software Development

Innovation

• MPLinker
• Adversarial Training
• Pre-trained Model Optimization