Awesome

MFMA

This repository provides a factual consistency checking model from our NAACL-Findings 2022 paper Masked Summarization to Generate Factually Inconsistent Summaries for Improved Factual Consistency Checking.

1. Usage of pre-trained factual consistency checking model

MFMA is a pre-trained factual consistency checking model(trained as a binary classifier) for abstractive summaries trained with the augmented negative samples using Mask-and-Fill with Article(MFMA).

You only need huggingface transformers library to load the pre-trained model.

from transformers import AutoModelforSequenceClassification
model = AutoModelforSequenceClassification("henry931007/mfma")

2. Training MFMA Instructions

Create a python 3.8 environment and then install the requirements.

Install packages using "requirements.txt"

conda create -name msm python=3.8
pip install -r requirements.txt

python train_fb.py --mask_ratio1 $MASK_ARTICLE \
                   --mask_ratio2 $MASK_ARTICLE \

python infer_fb.py --mask_ratio1 $MASK_ARTICLE \
                   --mask_ratio2 $MASK_ARTICLE \

python train_metric.py --datadir $DATA_PATH \

References

@inproceedings{lee2022mfma,
      title={Masked Summarization to Generate Factually Inconsistent Summaries for Improved Factual Consistency Checking}, 
      author={Hwanhee Lee and Kang Min Yoo and Joonsuk Park and Hwaran Lee and Kyomin Jung},
      year={2022},
      month={july},
      booktitle={Findings of the Association for Computational Linguistics: NAACL 2022},
}