This repository contains the LaTeX source, figures, and tables for the research paper on enhancing CodeBERT representations using domain-adaptive masked language modeling (MLM) for source code comment classification. The study uses the IRSE/FIRE dataset and explores the impact of combining original C code data with Python-derived silver-standard examples generated using x-ai/grok-4-fast.

Title: Improving Source Code Comment Relevance with Domain-Adaptive CodeBERT

Authors:

• L. Simiyon Vinscent Samuel
• Sundareswaran Rajaguru
• Vijayaraaghavan K. S.
• Dr. A. Vijayalakshmi (Supervisor)

Affiliation:
Department of Computer Science and Engineering, Sri Sivasubramaniya Nadar College of Engineering, Chennai, India

• Datasets:

  • Original IRSE/FIRE C code dataset (11,452 pairs, 8,326 cleaned training samples)
  • Rule-based synthetic comments (500 samples)
  • Silver-standard Python examples (1,640 samples) generated from humaneval.jsonl via x-ai/grok-4-fast

• Training Approach:

  • CodeBERT base MLM continued on combined corpus
  • 15% mask probability, batch size 16, learning rate 5e-5, up to 7 epochs
  • Fine-tuning with a classification head for relevance prediction

• Baselines:

  • TF–IDF features with Logistic Regression, Linear SVM, Multinomial NB, Random Forest
  • 10-fold cross-validation for hyperparameter tuning

• Evaluation Metrics: Accuracy, Precision, Recall, Weighted F1

• Configurations:

  1. Train Original → Test Original
  2. Train Combined → Test Combined
  3. Original → Combined
  4. Combined → Original

• Findings:

  • Domain-adaptive MLM significantly improves CodeBERT representations.
  • Cross-evaluation shows robust transfer even with Python-derived silver data.
  • Silver-standard augmentation increases vocabulary and diversity but introduces minor noise.