Replicando abordagem para geração automática de tratadores de exceção

Abstract

In this work, we replicated an approach based on deep learning for automated exception handling, but we were unable to obtain the same performance originally reported. The model was implemented using the source code and data set provided by the author of the original work. To carry out the approach, the Google Colab environment was used, which provides a runtime to execute scripts in the Python language. Using the configuration of the free version, we obtained a reduction of 0.56% in accuracy and 0.12% in the F1 - score, resulting from an increase of 1.98% in sensitivity and a loss of 2.59% in model precision try block locator. However, the catch block generator model showed a gain of 0.1% in Accuracy and 0.2% in the BLEU score. Finally, we conclude that the probable causes of such divergences were the configuration of the model’s hyperparameters and the differences between the environments used.