Skip to content

Code for paper "Extract, Denoise and Enforce: Evaluating and Improving Concept Preservation for Text-to-Text Generation" EMNLP 2021 and "Constrained Abstractive Summarization: Preserving Factual Consistency with Constrained Generation" arXiv 2020

Notifications You must be signed in to change notification settings

morningmoni/EDE

Repository files navigation

The repo provides the code for paper "Extract, Denoise and Enforce: Evaluating and Improving Concept Preservation for Text-to-Text Generation" EMNLP 2021 and "Constrained Abstractive Summarization: Preserving Factual Consistency with Constrained Generation" arXiv 2020

Code

[update] Code for Constrained Abstractive Summarization is in folder CAS/

The code is based on the seq2seq examples of huggingface transformers (3.0 <= version < 4.0)

The most important files are as follows:

DDBA.py: core functions for constrained generation, including a PyTorch implementation of DBA [1], adapted from the official MXNet implementation

finetune.py: model training

run_eval.py: model inference with or without constraints

transformers_local/generation_utils.py: modified the functions related to model decoding for enforcing constraints

transformers_local/modeling_bart.py: implemented BART+copy mechanism and other side functions

[1] "Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation", NAACL 2018

About

Code for paper "Extract, Denoise and Enforce: Evaluating and Improving Concept Preservation for Text-to-Text Generation" EMNLP 2021 and "Constrained Abstractive Summarization: Preserving Factual Consistency with Constrained Generation" arXiv 2020

Topics

Resources

Stars

Watchers

Forks

Languages