Fusing Semantic and Structural Features for Code Error Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Software Engineering

Background:

• Large Language Models (LLMs) based on the Transformer architecture excel at processing sequential data, showing potential for automated code error detection.
• However, current LLMs exhibit limitations in effectively handling structural code dependencies, hindering their performance in code analysis.

Purpose of the Study:

• To introduce a novel hybrid model that integrates semantic understanding from RoBERTa with the structural learning capabilities of Graph Neural Networks (GNNs).
• To enhance the accuracy and robustness of automated code error detection, specifically targeting common programming faults like runtime, index, and import/module errors.

Main Methods:

• Development of a hybrid model combining RoBERTa for semantic analysis and GNNs for structural dependency learning.
• Implementation of a fusion technique to effectively integrate the outputs of both components.
• Experimental evaluation comparing the hybrid model against baseline models on code error detection tasks.

Main Results:

• The proposed hybrid model demonstrates superior performance compared to existing models in detecting programming faults.
• Experimental evaluation shows significant improvements in accuracy and robustness.
• The model achieved a 1.75% increase in test accuracy over competitive baselines.

Conclusions:

• The integration of semantic and structural learning through a hybrid RoBERTa-GNN model effectively addresses the limitations of traditional LLMs in code error detection.
• The developed fusion technique is crucial for the model's enhanced performance.
• This approach offers a more robust and accurate solution for identifying common programming errors, advancing automated software quality assurance.