New!! The dataset is now available at Hugging Face 🤗

Finetuning_BART_for_FACET_Summarization

Paper: ACL 2021, Bringing Structure into Summaries: a Faceted Summarization Dataset for Long Scientific Documents

Introduction

There are some things that you need to familiarise yourself with / consider:

• BART model

• Fairseq

• Fairseq tutorial on fine-tuning BART on Seq2Seq task

• Emerald dataset used for faceted summarization

Fine-tuning

Prerequisites:

• PyTorch
• Fairseq
• Download the pretrained BART-large model
• Get the Emerald dataset

Preprocessing data

$> python preprocess_data.py
$> bash bpe.sh
$> bash binarize.sh

For parameters, check the finetune.sh script.

Although we did not find major differences with updating the max_tokens parameter during BART finetuning, in case you want to try it, the code allows to change the parameter (in the scripts/train.py file).

Find more information at fairseq bart repo!