ACL_2025_PETScan

Instructions for running PETScan.

1. Prepare an experiment folder consisting of files called "train_X.csv", "val_X.csv", and "test_X_LANG.csv". If testing on multiple languages, ensure that each train/val file has their own range of X values, but the testing files all have the same numbers. i.e., if testing on chinese and english, your files will look like:

   • train_0.csv, train_1.csv... train_4.csv... val_0.csv, val_1.csv...val_4.csv, test_0_chinese.csv, test_1_chinese.csv, ....
   • train_20.csv, train_21.csv... train_24.csv... val_20.csv, val_21.csv...val_24.csv, test_0_english.csv, test_1_english.csv, ...

   For our data, the ranges used for file names is:

   • 0-4 - chinese
   • 20-24 - english
   • 40-44 - spanish
   • 60-64 - yoruba
   • 80-84 - turkish

2. Make sure local/run.sh is correctly configured, and if this is the first time using the experiment runner, ensure that the 'requirements' line is uncommented.

3. Ensure src/launch.py values are correctly set - including the exp_dir path. IMPORTANT - the lines labelled 'CHANGE' are to instantiate what language pair you wish to experiment with.

   For 'DIFF' environment variable:

   • if L1 is chinese
     • +20 english
     • +40 spanish
     • +60 yoruba
     • +80 turkish
   • if L1 is english
     • -20 chinese
     • +20 spanish
     • +40 yoruba
     • +60 turkish
   • if L1 is spanish
     • -40 chinese
     • -20 english
     • +20 yoruba
     • +40 turkish
   • if L1 is yoruba
     • -60 chinese
     • -40 english
     • -20 spanish
     • +20 turkish
   • if L1 is turkish
     • -80 chinese
     • -60 english
     • -40 spanish
     • -20 yoruba

4. From a terminal, 'cd' into the PETScan folder.

5. Enable permissions to execute the script ('chmod +x local/run.sh')

6. launch training with sh local/run.sh