The pytorch implementation of Question Generation over Knowledge Base via Modeling Diverse Subgraphs with Meta-learner (EMNLP 2022).

1. Environments

• Create a virtual environment first via:

$ conda activate -n your_env_name python 3.8.5 pip

• Install all the required tools using the following command:

$ conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

$ pip install -r requirements.txt

2. Dataset

• WQ : dataset/ contains the files for WQ dataset.
• PQ : dataset/ contains the files for PQ dataset.

1. Prepare dataset and positive samples for training retriever.

• Prepare dataset for retriever: $ python retriever/preprocess_retriever.py
• Generate positive samples for retriever: $ python retriever/relation_path.py

2. Train the GCL-based retriever.

 $ python retriever/main_gcl.py

Specifically, other retrievers run as follows:

• Run DGI-based retriever: $ python retriever/main_dgi.py
• Run RGCN-based retriever: $ python retriever/main_rgcn.py
• Run GED retriever: $ python retriever/main_ged.py
• Run All-RP retriever: $ python retriever/relation_path.py

3. Prepare dataset for DSM and process dataset for creating learning tasks .

• Prepare dataset for DSM:

  $ python preprocess.py -input_dir dataset/WQ --output_dir './output_WQ' --model_name_or_path 'facebook/bart-base'
  

• Process dataset to create learning tasks:

  dataset.py is used to process dataset to create learning tasks.

4. To run the DSM, execute (Note: we take the GCL-based retriever as an example, and other retrievers are similar.):

$ python bart_train.py --epoch 30 --input_dir dataset/WQ --output_dir './output_WQ' --update_lr 5e-5 --meta_lr 3e-5 --model_name_or_path 'facebook/bart-base'

We evaluate two classical KBQA models named GRAFT-Net and NSM on WebQSP. To evaluate the quality of the generated questions by DSM, we replace part of the (question, answer) pairs in WebQSP with the generated questions. The code for GRAFT-Net and NSM can be downloaded from https://pan.baidu.com/s/1RGOZW23FbfcpkShvJkGRUw?pwd=ddsm.