ichiroex / sat

The accompanying code for "Table Fact Verification with Structure-Aware Transformer" , (EMNLP 2020).

sat

The accompanying code for Table Fact Verification with Structure-Aware Transformer (short paper of EMNLP 2020).

1.Enveriment

python 3.7
pip install sklearn   
pip install torch==1.1   
pip install pytorch-pretrained-bert==0.6.2   
pip install transformers==2.6   
bert-base-uncased/  download and unzip the pre-trained bert here. (the multilingual bert is prefered)

2.train from scratch

a.clone FacTable via: git clone https://github.com/wenhuchen/Table-Fact-Checking.git
b.cp sat_main.py Table-Fact-Checking-master
c.mkdir models
d.train the model: 
    python sat_main.py # with default configuration (vertical scan)
    It takes about 50 mins on a single V100 to run an epoch, and takes 15-18 epochs to achieve best validataion results.

e. All the ablation experiment is runnable. 
   e.g: python sat_main.py --use_vm no # ablation on attention masking
   please refer to sat_main.py for more configuration details.