Self-Supervised Learning Models for Source Code Representation and Software Assessment

Abstract

Software quality assessment is essential for maintaining reliable and scalable software systems as codebases become larger and more complex. Traditional prediction methods depend on manually engineered metrics and labeled defect datasets, limiting scalability and semantic understanding. This study investigates self-supervised learning for source code representation to support automated software quality assessment with reduced dependence on manual annotation. The proposed framework learns contextual code embeddings from large software repositories and integrates them with predictive models to evaluate defect density, code complexity, and maintainability indicators. Experimental results show that learned representations capture meaningful structural and semantic relationships within source code and improve prediction performance compared with traditional metric-based approaches. The findings suggest that self-supervised representation learning can strengthen automated software engineering analysis, support early detection of quality risks, and provide practical benefits for developers and maintenance teams working with large-scale software systems.