Compositional Generalization for Data-to-Text Generation

This repository releases code and data for Improving Compositional Generalization with Self-Training for Data-to-Text Generation our paper accepted at ACL 2022.

Data

We released our fewshot weather data, please find more detailed descriptions under data/.

Code

To prepare finetuning data for BLEURT (change input file to yours), run:

sh run.sh

To finetune BLEURT, follow these instructions.

Cite

@inproceedings{Mehta2022compgen,
  title = {Improving Compositional Generalization with Self-Training for Data-to-Text Generation},
  author = {Sanket Vaibhav Mehta, Jinfeng Rao, Yi Tay, Mihir Kale, Ankur Parikh, Hongtao Zhong, Emma Strubell},
  year = {2022},
  booktitle = {Proceedings of ACL}
}