Adversarial Augmentation Policy Search for Domain and Cross-Lingual Generalization in Reading Comprehension
PyTorch code for the Findings of EMNLP 2020 paper "Adversarial Augmentation Policy Search for Domain and Cross-Lingual Generalization in Reading Comprehension".
In order to run BayesAugment for SQuaD v2.0 using Roberta-Base, please run the following:
python search_augmentation_policy_bayesopt.py
This code was run on a system with 4 2080 Ti GPUs.
The BayesAugment codebase is heavily derived from the original Adversarial RC codebase.
Additionally, to prepare the adversarial dataset, the pipeline uses/requires the following repositories:
nectar (used in original code)
pattern (used in original code)
Syntactic Paraphrasing Networks (To generate syntactic paraphrases)
Semantic Paraphrases (To generate semantic paraphrases)