code_review

code_review is the replication package of the work "Towards Automating Code Review Activities".

The purpose of this repository is to provide everything necessary to replicate our results.

Requirements

• Python >= 3.5 and <= 3.8
• OpenNMT-tf

Use the following commands to install the dependencies.

pip install --upgrade pip
pip install OpenNMT-tf

Content

Datasets

The datasets folder contains two datasets used for experimenting with two Transformer models. Both datasets are split in three parts training (80%), validation (10%), and test (10%) sets; Each Transformer model (I.e., 1-encoder and 2-encoders folders) has three sub-folders train, eval, and test with the actual data.

• The dataset for the 1-encoder model is composed of Reviewed Code Pairs (RCPs). A RCP is a <m_s, m_r> pair composed by the abstracted code of the method extracted from the Java file submitted by a contributor for review (m_s) and by the abstracted code of its revised version (m_r). We provide data in the form of textual file organized per row, therefore, a pair <m_s, m_r> refers to the same line number of the following two files:

  • src file contains the m_s instances;
  • tgt file contains the m_r instances.

• The dataset for the 2-encoders model is, instead, composed of Reviewed Commented Code Triplets (RCCTs), each triple have the form <m_s, m_r, r_nl>, where m_r is the abstracted code of the method implementing the natural language recommendation (r_nl) provided by a reviewer for the method in the code submitted (m_s). Therefore, to build a triple <m_s, m_r, r_nl> three files are provided:

  • src1 file contains the r_nl instances;
  • src2 file contains the m_s instances;
  • tgt file contains the m_r instances.

NOTE. The natural language recommendations have been cleaned and abstracted as described in our manuscript.

Code

The code folder contains all scripts used to train and test the models.

For both models, the best configuration we found through the tuning of the hyperparameters is provided in the respective folders. To start the training of a model it is sufficient to run the trainin.py file. It first will create the vocabularies needed and then it will start the training. The trained model will be saved in the run folder.

Once the model is trained, it is possible to test it on the test set by running the infer.py file. This script creates the predictions.txt file containing all the model's predictions.

It is, also, possible to change the beam search size modifying the beam_width and num_hypotheses parameters in the training/data.yml file.

Other files

This repository also contains:

• idioms.csv : the list of idioms we used during abstraction phase;
• is_relevant.ipynb : a jupyter file showing the logic used to remove the not relevant comments;
• 1_encoder_perfect_predictions.xlsx and 2_encoders_perfect_predictions.xlsx : the qualitative analysis of perfect predictions.
• bleu4_boxplot.zip : a compressed file containing the boxplots of the BLUE4 score of the predictions. Read the README in the compressed file for additional information.
• generated-predictions.zip : a compressed file containing the predictions generated by both models. Read the README in the compressed file for additional information.
• FiltersTable.png : a table showing the numeric results of the filters we applied to the data. For each filter we reported the number of triplets removed from it.