The official code of paper Awakening Latent Grounding from Pretrained Language Models for Semantic Parsing.
Please first install PyTorch, and then install all the dependencies by running:
pip install -r requirements.txt
Please remember to unzip the json.zip
in the data/wtq_grounding
folder. And the file structure should be like:
data/wtq_grounding
├── json
│ ├── 202.json
│ ├── 203.json
│ ├── ...
├── dev.json
├── test.json
└── ...
Please run the script train_spider_ground.sh
to train the grounding model on Spider dataset.
Please run the script train_wtq_ground.sh
to train the grounding model on WTQ dataset.
Please run the script eval_spider_ground.sh
to evaluate the grounding model on Spider dataset. Note that you should replace the model checkpoint checkpoints/spider_grounding_model/model.pt
with yours.
You should get the following results after following the training script:
avg loss = 0.2189
table accuracy = 0.8453
column accuracy = 0.7602
value accuracy = 0.9449
overall accuracy = 0.7050
table P = 0.847, R = 0.857, F1 = 0.852
column P = 0.842, R = 0.838, F1 = 0.840
value P = 0.948, R = 0.932, F1 = 0.940
average F1 = 0.8773
Please run the script eval_wtq_ground.sh
to evaluate the grounding model on WTQ dataset. Note that you should replace the model checkpoint checkpoints/wtq_grounding_model/model.pt
with yours.