Exploring Continual Learning of Compositional Generalization in NLI

Code and Data for the TACL paper (to appear) Exploring Continual Learning of Compositional Generalization in NLI

We introduce the Continual Compositional Generalization in Inference (C2Gen NLI) challenge, where a model continuously acquires knowledge of constituting primitive inference tasks as a basis for compositional inferences. We explore how continual learning affects compositional generalization in NLI, by designing a continual learning setup for compositional NLI inference tasks.

Data

Data Format

{
"verdical_label": "negative",        
"sick_label": "neutral",    
"sent1": "A man fails to make a snowball", 
"sent2": "A man plays with a ball", 
"mid_sent": "A man makes a snowball",
"label": "neutral"   
}

Data Download (OneDrive)

link: https://mailnankaieducn-my.sharepoint.com/:u:/g/personal/fuxiyan_mail_nankai_edu_cn/EeNfHHyVvkBAjbfoXO_OnKYBn0gjhRf3EeuK_0Te5_LwGw?e=q8g41M

Code

preliminary: split data as you require from the provided dataset

environment: python3.7, pytorch1.7.1

run: the script is provided in the run.sh

Acknowledgement: The code of continual learning strategies come from VisCOLL

Citations

Please cite our paper if you are using this dataset.

@article{fu2024exploring,
  title={Exploring Continual Learning of Compositional Generalization in NLI},
  author={Fu, Xiyan and Frank, Anette},
  journal={arXiv preprint arXiv:2403.04400},
  year={2024}
}