Skip to content

Latest commit

 

History

History
134 lines (92 loc) · 3.65 KB

README.md

File metadata and controls

134 lines (92 loc) · 3.65 KB
Raw

Controllable Open-ended Question Generation with A New Question Type Ontology

Code for ACL 2021 paper "Controllable Open-ended Question Generation with A New Question Type Ontology".

Raw data

Question type annotation

Reddit

Our Yahoo dataset is based on the Yahoo Answer L6 dataset. After obtaining the license for the L6 dataset, please email Shuyang (caoshuy@umich.edu) with the proof of license attached to obtain the Yahoo dataset.

Environment

Our experiments are based on PyTorch 1.7.0 and Fairseq at commit 0db28cd with a simple edit. Newer versions of Fairseq might also work. For graph neural networks, we use PyTorch-Geometric 1.7.2.

# virtual environment
conda create -n open_ended_qg python=3.7
conda activate open_ended_qg

# install pytorch
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=10.2 -c pytorch

# install fairseq, note that you need to follow the instructions in fairseq/README.md 
# to install other dependencies (e.g., apex for training)
cd lib/fairseq
pip install -e .
# fix hydra error
pip install hydra-core==1.0.7

# install torch-geometric
pip install torch-scatter==2.0.7 -f https://data.pyg.org/whl/torch-1.7.0+cu102.html
pip install torch-sparse==0.6.9 -f https://data.pyg.org/whl/torch-1.7.0+cu102.html
pip install torch-geometric==1.7.2

Note: we use AllenNLP during data processing, which requires a different PyTorch version. Please use a different virtual environment for AllenNLP.

Data Preprocess

Preprocessed binarized Reddit data can be downloaded from here.

For data preprocessing, please refer to the README in data_preprocess.

Run our models

Please download the generation models from here and put them under $MODEL/generation_models. The binarized dataset should be under $DATA/binarized_data.

To convert the fairseq generation output to text, use convert_output.py:

python convert_output.py --generate-dir <result_dir>
JointGen
cd gen_scripts
./jointgen.sh $DATA/output/jointgen
ExplGen
cd gen_scripts
./explgen.sh $DATA/output/explgen
TplGen
cd gen_scripts
./tplgen_question_generation.sh $DATA/output/tplgen_question
ExplGen: conditioned on top 9 types
cd gen_scripts
./explgen_9types.sh $DATA/output/explgen_9types
TplGen: conditioned on top 9 types
cd gen_scripts
./tplgen_question_generation_9types.sh $DATA/output/tplgen_question_9types

Train our models

Please set BART_PATH as the path to the bart.large model, which can be downloaded here.

export BART_PATH=<path_to_bart_large_dir>/model.pt
JointGen
cd train_scripts
CUDA_VISIBLE_DEVICES=0,1 ./jointgen.sh $BART_PATH $MODEL/jointgen
ExplGen
cd train_scripts
CUDA_VISIBLE_DEVICES=0,1 ./explgen.sh $BART_PATH $MODEL/explgen
TplGen: template generation
cd train_scripts
CUDA_VISIBLE_DEVICES=0,1 ./tplgen_template_generation.sh $BART_PATH $MODEL/tplgen_template_generation
TplGen: question generation
cd train_scripts
CUDA_VISIBLE_DEVICES=0,1 ./tplgen_question_generation.sh $BART_PATH $MODEL/tplgen_question_generation