This is the replication package for the paper On ML-Based Program Translation: Perils and Promises, accepted to ICSE'23 NIER. It includes all of the files we tested, the fixes we made, and the results following each.
The paper was authored by the following:
- Aniketh Malyala: aniketh.malyala@gmail.com (Silver Creek High School)
- Katelyn Zhou: zhoukatelyn@gmail.com (Silver Creek High School)
- Baishakhi Ray: rayb@cs.columbia.edu (Columbia University in the City of New York)
- Saikat Chakraborty: saikatc@microsoft.edu (Microsoft Research)
In the file HOWTOREPLICATE.md, we detail how and what to replicate in terms of our results. We used the Facebook Research TransCoder model in order to obtain our results.
The folder java2python includes every single file we ran through the model to translate from Java to Python and the manual fixes we made along with the translation results for everything.
The folder python2java includes the same as above just from opposing languages.
The data we used comes directly from the data that the TransCoder model was trained upon. The files were originally from Geeks4Geeks. We took a random sample of 50 files from both the Java training data and Python training data for a total of 100 files altogether.
Here is the link to the Java Training Data: https://github.com/facebookresearch/TransCoder/tree/main/data/evaluation/geeks_for_geeks_successful_test_scripts/java
Here is the link to the Python Training Data: https://github.com/facebookresearch/TransCoder/tree/main/data/evaluation/geeks_for_geeks_successful_test_scripts/python