FLEXTAF: Enhancing Table Reasoning with Flexible Tabular Formats

This repository contains code for the paper "FLEXTAF: Enhancing Table Reasoning with Flexible Tabular Formats".

If you use FLEXTAF in your work, please cite it as follows:

@article{zhang2024flextaf,
  title={FLEXTAF: Enhancing Table Reasoning with Flexible Tabular Formats},
  author={Zhang, Xuanliang and Wang, Dingzirui and Dou, Longxu and Wang, Baoxin and Wu, Dayong and Zhu, Qingfu and Che, Wanxiang},
  journal={arXiv preprint arXiv:2408.08841},
  year={2024}
}

Build Environment

conda create -n flex python=3.10
conda activate flex
pip install -r requirements.txt

Pre-Process Data

Download and put each dataset in ./dataset, and run dataset/slurm/preprocess.slurm.

Download Model

Download the models and put them in ./model.

FLEXTAF-Vote

Reasoning

Run the table reasoning with reason/slurm/inference.slurm, in which can select the tabular format.

Vote

Ensemble the results of multiple formats with reason/slurm/vote.slurm

FLEXTAF-Single

Classification

This step is to train the classifier to predict the most suitable tabular format.

Obtain training data

Firstly run the table reasoning on the training set to get the training data with reason/slurm/inference.slurm and reason/slurm/vote.slurm.

Train

Train the classifier with classify/slurm/multi_label_finetune.slurm.

Reasoning

Predict the suitable tabular format with classify/slurm/multi_label_classify.slurm. If with results of all candidate tabular formats, the performance of Flex-Formats-Single can be also obtained at this step.