A study of pyramid structure for code correction

A study of pyramid structure for code correction Chin, Sang Huang, Shan; Zhou, Xiao We demonstrate the implementations of pyramid encoders in both multi-layer GRU and Transformer for seq2seq tasks. We apply the models to the code correction task on Juliet Test Suite for C/C++ and Java of Software Assurance Reference Dataset (SARD), successfully repaired 90.1% of faulted code in the test dataset, and show that a pyramid structure can greatly improve memory efficiency and therefore computation efficiency. We successfully carried out error type classification task on ITC benchmark examples (with only 685 code instances) using transfer learning with models pre-trained on Juliet Test Suite, pointing out a novel way of processing small datasets.