This repository contains the official code for the paper:

Data Factors for Better Compositional Generalization

Xiang Zhou, Yichen Jiang and Mohit Bansal

EMNLP 2023

The code is tested on Python 3.7.9 and PyTorch 1.8.1.

Other dependencies are listed in requirements.txt and can be installed by running pip install -r requirements.txt

This repository uses wandb for logging experiments. So before running your experiments, you need to log in you wandb account.

The preprocessed datasets and resources used in our experiments can be downloaded at here.

To replicate the experiments below, please download the datasets and extract it under the root folder.

We provide scripts to run experiments in our paper. All the scripts are under the scripts folder.

For the experiments on increasing the lexical or length complexity of the dataset in Sec. 3, the corresponding scripts are

• Increasing primitive complexity in SCAN SCAN_transformer_huge.sh. The corresponding baseline for SCAN can be replicated from the instructions at owenzx/met-primaug
• Increasing primitive complexity in GeoQuery GEO_transformer_prim.sh The corresponding baseline is at GEO_transformer_baseline.sh
• Increasing length complexity in SCAN SCAN_length.sh

For the experiments changing the curriculum in Table 4 in Sec. 4, the corresponding scripts are

• On the original dataset: GEO_curriculum_transformer.sh and GEO_curriculum_transformer_prim.sh
• On the 20x augmentation dataset: GEO_large_curriculum_transformer.sh and GEO_large_curriculum_transformer_prim.sh

For the AugZero experiments in Sec. 4, the corresponding script is

• On SCAN: SCAN_transformer_zero.sh
• On Geoquery: GEO_transformer_zero.sh

For the difficulty-related experiments in Sec. 5, the corresponding scripts are

• Prototype-based difficulty experiments (Fig. 4): ATIS_transformer_difficulty.sh, SMCAL_transformer_all_difficulty.sh and SMCAL_cs_transformer_all_difficulty.sh
• Increasing example difficulty in SCAN (Fig. 5): SCAN_transformer_jump_complete_len.sh and SCAN_tranformer_jump_complete_pe.sh

The code in this repository is based on https://github.com/ekinakyurek/lexical

@inproceedings{zhou2023data,
  title={Data Factors for Better Compositional Generalization},
  author={Xiang Zhou and Yichen Jiang and Mohit Bansal},
  booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2023}
}