RBMD: RoBERTa-Based Module Detection in Multi-Programming Language Software Systems

The dynamic nature of web and software systems requires modularization methods that can adapt to frequent updates and diverse structures while maintaining scalability and efficiency. In this research, we propose a RoBERTa-based Module Detection framework that leverages transformer models to classify and manage software modules using the semantic content of source code, comments, and related textual data. Unlike traditional dependency graph-based methods, which face computational bottlenecks, our content-driven approach offers a streamlined, scalable solution for software systems. This framework represents a significant step in automating modularization, eliminating the need for extensive documentation, and leveraging semantic content to manage complexity in modern software development. Experimental results demonstrate the method's efficiency: for Mozilla Version 3.7, the model achieved 92.55 % accuracy and 92.47 % F1-score after four epochs, while for Version 134.0, it reached 98.13% accuracy and 98.02 % F1-score with rapid convergence and minimal training loss. Additionally, the model achieved an outstanding accuracy and F1-score of 99.70 % for Chromium.