FactEdit - Post-Editing Factual Errors in Generated Summaries

This repository contains code and data for EMNLP 2022 paper Correcting Diverse Factual Errors in Abstractive Summarization via Post-Editing and Language Model Infilling work with Hannaneh Hajishirzi, William Cohen and Yulia Tsvetkov.

Data and Models

All training data and pretrained models can be found here: https://drive.google.com/drive/folders/1VeALcCBLIx0H3VQF2_pEJJ5ieQWtEpo9?usp=sharing

Check out scripts/ for various training, inference and evaluations scripts.

Training Infilling Model

Edit data and output paths in scripts/cnndm_run_infill.sh

bash scripts/cnndm_run_infill.sh

Generate Training Data for correction Model

Edit data and output paths in scripts/cnndm_predict_infill.sh

bash scripts/cnndm_predict_infill.sh

Training and Evaluating Fact Correction Model

Edit data and output paths in scripts/cnndm_run_corr.sh, scripts/cnndm_predict_corr.sh

bash scripts/cnndm_run_corr.sh
bash scripts/cnndm_predict_corr.sh
bash eval.sh