Awesome

GNN-TabFact

This is the state-of-the-art model on TabFact Dataset dataset, we leverage the idea proposed in NumGNN into the encoding of tabular data.

Requirements:

• HuggingFace Transformers 2.6.0
• tensorboardX
• pandas 0.25.1
• AllenNLP 0.9.0

Architecture

• Cross-Attention Between Table and Statement to obtain the representation
• Construct the greater/less mask for the table numeric columns
• Use the the dense greater/less connection to propagate the information in each cell
• Obtain the graph repsentation obtained by the NumGNN
• Finally use it to do the two-way classification.

Performance

We demonstrate our results as follows:

Training and Evaluating

Creating a folder for saving the model

mkdir models

Downloading the pre-trained model from Amazon S3, also link the folder of all_csv from TabFact dataset.

ln -s TABFACT/data/all_csv .
cd models
wget https://gnntabfact.s3-us-west-2.amazonaws.com/gnn_fp16_numeric.zip
unzip gnn_fp16_numeric.zip

Loading the trained GNN Model and reproduce our results:

CUDA_VISIBLE_DEVICES=0 python gnn.py --model bert-base-multilingual-uncased --do_test --encoding gnn --load_from models/gnn_fp16_numeric/model_ep4.pt

Retrain your own GNN Model on TabFact:

CUDA_VISIBLE_DEVICES=0 python gnn.py --model bert-base-multilingual-uncased --do_train --encoding gnn --output_dir models/gnn_fp16_numeric_test --attention cross --lr_default 5e-6