This is the Pytorch implementation for the paper Generating Diverse and Consistent QA pairs from Contexts with Information-Maximizing Hierarchical Conditional VAEs (ACL 2020, long paper) : [Paper] [Slide] [Video].

Contribution of this work

• We propose a novel hierarchical variational framework for generating diverse QA pairs from a single context, which is, to our knowledge, the first probabilistic generative model for questionanswer pair generation (QAG).
• We propose an InfoMax regularizer which effectively enforces the consistency between the generated QA pairs, by maximizing their mutual information. This is a novel approach in resolving consistency between QA pairs for QAG.
• We evaluate our framework on several benchmark datasets by either training a new model entirely using generated QA pairs (QA-based evaluation), or use both ground-truth and generated QA pairs (semi-supervised QA). Our model achieves impressive performances on both tasks, largely outperforming existing QAG baselines.

This code is written in Python. Dependencies include

• python >= 3.6
• pytorch >= 1.4
• json-lines
• tqdm
• pytorch_scatter
• transfomers

Download data from here. It contains SQuAD training file(data/squad/train-v1.1.json) and our own dev/test split(data/squad/my_dev.json, data/squad/my_test.json). We preprocess it and convert to examples.pkl and features.pkl. Those pickle files are in data/pickle-file folder. If you want to download the original data, run the following commands

mkdir squad
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json -O ./squad/train-v1.1.json
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json -O ./squad/dev-v1.1.json

Train Info-HCVAE with the following command. The checkpoint will be save at ./save/vae-checkpoint.

cd vae
python main.py

Generate QA pairs from unlabeled paragraphs. If you generate QA pairs from SQuAD, use option --squad.

cd vae
python translate.py --data_file "DATA DIRECTORY for paragraph" --checkpoint "directory for Info-HCVAE model" --output_file "output file directory" --k "the number of QA pairs to sample for each paragraph" --ratio "the percentage of context to use"

It requires 3 1080ti GPUS (11GB memory) to reproduce the results. You should download data from here and place it under the root directory. Uncompress it and the "data" folder contains all the files required for QAE and Semi-supervised Learning.

cd qa-eval
python main.py --devices 0_1_2 --pretrain_file $PATH_TO_qaeval --unlabel_ratio 0.1 --lazy_loader --batch_size 24

It requires 4 1080ti GPUS (11GB memory) As QAE, you should download the data from here and place it under the root directory.

cd qa-eval
python main.py --devices 0_1_2_3 --pretrain_file $PATH_TO_semieval --unlabel_ratio 1.0 --lazy_loader --batch_size 32

Download data from here and uncompress it under the root directory. The folder data/harv_synthetic_data_qae contains generated QA pairs from Harvesting QA dataset without any filtering. Another folder data/harv_synthetic_data_semi contains the same generated QA pairs but with postprocessing. We replace the generated answer with pretrained BERT QA model if its F1 is lower than the threshold.

To cite the code/data/paper, please use this BibTex

@inproceedings{lee2020generating,
  title={Generating Diverse and Consistent QA pairs from Contexts with Information-Maximizing Hierarchical Conditional VAEs},
  author={Lee, Dong Bok and Lee, Seanie and Jeong, Woo Tae and Kim, Donghwan and Hwang, Sung Ju},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year={2020}
}