Application of Seq2Seq Models on Code Correction.

Area of Science:

• Software Engineering
• Artificial Intelligence
• Programming Language Analysis

Background:

• Automated program repair is crucial for software quality.
• Existing seq2seq models face computational and memory challenges.
• Large datasets are often required for effective model training.

Purpose of the Study:

• To improve the efficiency and effectiveness of seq2seq models for program correction.
• To introduce a novel pyramid encoder for enhanced computational and memory efficiency.
• To explore transfer learning for classifying programming error types on small datasets.

Main Methods:

• Applied various seq2seq models to the Juliet Test Suite for C/C++ and Java.
• Introduced a pyramid encoder architecture to optimize seq2seq models.
• Utilized transfer learning with models pre-trained on the Juliet Test Suite for error type classification on the ITC benchmark.

Main Results:

• Achieved 75% repair rate for C/C++ and 56% for Java on program correction tasks.
• The pyramid encoder significantly improved computational and memory efficiency with comparable repair rates.
• Successfully performed error type classification on a small dataset (ITC benchmark) using transfer learning.

Conclusions:

• Seq2seq models with pyramid encoders offer an efficient and effective solution for automated program repair.
• Transfer learning presents a viable strategy for addressing small programming language datasets in error classification.
• The Juliet Test Suite is a valuable resource for training and evaluating program repair and classification models.