Program-Enhanced-Table-Fact-Checking

Code for the EMNLP 2020 paper "Program Enhanced Fact Verification with Verbalization and Graph Attention Network"

Environment

  python  3.7.3
  pytorch 1.3.1

Data Preparation

Please refer to the data preprocessing in TABFACT baselines to get the following data.

0. The preprocessed data for LPA

$program_dir$

1. The preprocessed data for Table-BERT

Note that we take the horizontal version.

$table_dir$

A. Program Selection

In './program_selection', please run 'run_BERT_margin_loss.py' for program selection.

You can always train your own program selection model, and get the selected programs for the next verbalization step.

To simplify this procedure, we provide a checkpoint of program selection model so that you can just load it and get the selected programs for verbalization:

python run_BERT_margin_loss.py --do_gen --load_dir $the downloaded ckpt$ --data_dir $program_dir$

B. Verbalization with Program Execution

In './verbalization', please run 'run_verbalize_attn.py' for verbalization.

python run_verbalize_attn.py --selected_prog_dir $the output of prev program selection$ --table_bert_dir $table_dir$ --save_root_dir $determine your verb output path$

C. Graph-based Verification Network

To train the grah-based verification model (here we can load a table-bert baseline ckpt to accelerate the training process):

python run_gvn.py --do_train --do_eval --tune_table_after10k --load_dir $table-bert baseline ckpt$ --data_dir $table_dir$ --data_dir_verb $verb output path$

To load the checkpoint and only run eval:

python run_gvn.py --do_eval --do_test --do_simple_test --do_complex_test --do_small_test --load_dir_whole $the downloaded ckpt$ --data_dir $table_dir$ --data_dir_verb $verb output path$

Contact

For any questions, please send email to authors.