This repository contains the code for reproducing the results reported in the following paper:
Orhan AE (2021) Compositional generalization in semantic parsing with pretrained transformers. arXiv:2109.15101.
Please see the slurm scripts inside the scripts
directory for replicating each experiment reported in the paper (11 models x 2 benchmarks = 22 experiments in total). The file names should be self-explanatory. All experiments use the same code base contained in run_translation.py
, which is modified from the translation example in the Huggingface examples repository. The output files for each experiment can be found in the results
directory.