Decomposition-Table-Reasoning

Code for EMNLP 2021 paper "Exploring Decomposition for Table-based Fact Verification"

Pseudo-Decomposition Dataset

Available in ./Data, including:

• train_decomp_type_detect_lm.json: training data for decomposition type detection
• train_decomp_generation_lm.json: training data for decomposition generation

Decomposition Type Detection

• Fine-tune BERT model for decomposition type detection (5-way classification).
• Predict decomposition type for all samples in the TabFact dataset.

Decomposition Generation

• Fine-tune GPT-2 model for decomposition generation.
• Generate decomposition for all samples in the TabFact dataset.

Subproblem Solver

• After getting the decomposed sub-problems for samples in the TabFact, we solve them to obtain corresponding answers.

• How to verify the decomposed sub-statements?

  We used the TAPAS model fine-tuned on the TABFACT (Chen et al., 2020) for sub-statement verification, and the model is available here.

• How to answer the decomposed sub-questions?

  We used the TAPAS model fine-tuned on the WikiTableQuestions (Pasupat and Liang, 2015) for sub-question answering, and the model is available here.

Combine Intermediate Evidence for Fact Verification

• Data

  cd ./Data/TabFact_data
  tar -zxvf table.tar.gz
  

• Download model checkpoint here, and put them in the ./ckpt folder.

• Re-train the model

  If you would like to re-train the verification model:

  cd ./Code
  python run.py --do_train --do_eval --tune_tapas_after10k --load_tapas_model ../ckpt/base.pt --data_dir ../Data/TabFact_data
  

• Test the model

  Or you can evaluate the model using the provided checkpoint:

  cd ./Code
  python run.py --do_eval --do_test --do_simple_test --do_complex_test --do_small_test --tune_tapas_after10k --load_model ../ckpt/model.pt --data_dir ../Data/TabFact_data
  

Contact

For any questions, please send email to the author.