ravoxsg / summafusion

Source code for SummaFusion (EMNLP 2022).

SummaFusion

Source code for the paper Towards Summary Candidates Fusion.

Mathieu Ravaut, Shafiq Joty, Nancy F. Chen.

Accepted for publication at EMNLP 2022.

Setup

1 - Download the code

git clone https://github.com/Ravoxsg/SummaFusion.git
cd SummaFusion

2 - Install the dependencies

conda create --name summa_fusion python=3.8.8
conda activate summa_fusion
pip install -r requirements.txt

Dataset

We use HuggingFace datasets library to access and save each dataset. We save it as .txt file for the sources, and another one for the summaries, with 1 data point per line.

For instance to download and save SAMSum (default code):

cd src/candidate_generation/
bash dataset.sh

Note that for Reddit TIFU, we make a custom 80/10/10 train/val/test split.
To match our results on Reddit TIFU, first double check that you have the following:
For training set, size is 33,704 and the first data point summary is:
got a toy train from first grade. used an old hot wheels ramp to fling it into the air and smash my ceiling fan globe.
For the validation set, size is 4,213 and the first data point summary is:
married a redditor. created a reddit account. lost many hours to reddit.
For the test set, size is 4,222 and the first data point summary is:
laughed at baby boner...it turned into a super soaker.

If you want to work in few-shot, you need to prepare the (train, val) few-shot pairs. For each dataset and each few-shot size (among {10,100,1000}), we sample 3 pairs, corresponding to seeds {42,43,44}.

For instance on SAMSum 100-shot (default code):

bash few_shot.sh

DEMO

If you just want a demo (in a single file) of SummaFusion on a single data point (default: XSum), run:

cd src/summafusion/
CUDA_VISIBLE_DEVICES=0 python demo.py

EVALUATION pipeline

1 - Generate summary candidates

SummaFusion takes as input a set of summary candidates from a given sequence-to-sequence model PEGASUS with diverse beam search.

You need such a fine-tuned checkpoint before generating the candidates.

For instance on SAMSum 100-shot validation set (default code):

CUDA_VISIBLE_DEVICES=0 bash candidate_generation.sh

Generating summary candidates should take a few minutes in few-shot, and up to a few hours on the full validation or test sets of XSum, Reddit or SAMSum.

2 - Score the candidates

As part of SummaFusion, we train a classifier on the summary candidates and thus need candidate-level information.

For instance to score candidates on SAMSum 100-shot validation set with ROUGE-1/2/L (default code):

bash scores.sh

Scoring all candidates should take a few seconds in few-shot, and up to a few minutes on the validation or test sets of XSum, Reddit or SAMSum.

3 - Download the SummaFusion model checkpoint

XSum full-shot checkpoint: here
XSum 100-shot checkpoint (seed: 42): here
Reddit full-shot checkpoint: here
Reddit 100-shot checkpoint (seed: 42): here
SAMSum full-shot checkpoint: here
SAMSum 100-shot checkpoint (seed: 42): here

If you are using a full-shot checkpoint, place it into:

src/summafusion/saved_models/{dataset}/

And if it is a few-shot checkpoint, place it into:

src/summafusion/saved_models/{dataset}_few_shot/

where {dataset} is in {xsum,reddit,samsum} corresponds to the dataset name.

4 - Run SummaFusion

For instance, to run SummaFusion on SAMSum 100-shot validation set (default code):

cd ../summafusion/
CUDA_VISIBLE_DEVICES=0 bash evaluate.sh

Make sure that the argument --load_model_path points to the name of the checkpoint you want to use.

TRAINING pipeline

1 - Fine-tune base models

We follow a cross-validation approach similar to SummaReranker.

First we split each training set into two halves.

For instance on SAMSum 100-shot:

cd ../base_model_finetuning/
bash build_train_splits.sh

Then we train a model on each half, and a third model on the entire training set.

CUDA_VISIBLE_DEVICES=0 bash train_base_models.sh

2 - Generate summary candidates

Then, we need to get summary candidates on the training, validation and test sets.

For instance on SAMSum 100-shot:

cd ../candidate_generation/
CUDA_VISIBLE_DEVICES=0 bash candidate_generation_train.sh

Generating summary candidates should take a few minutes in few-shot, and up to a few days for XSum full-shot.

3 - Score the candidates

Next, we need to score the summary candidates on the training, validation and test sets for each of the metrics. This is needed for the candidate-level classification part of SummaFusion.

For instance on SAMSum 100-shot with ROUGE-1/2/L:

CUDA_VISIBLE_DEVICES=0 bash scores_train.sh

Scoring all candidates should take a few seconds in few-shot, a few minutes in full-shot.

4 - Train SummaFusion

For instance, to train Summafusion on SAMSum 100-shot:

cd ../summafusion/
CUDA_VISIBLE_DEVICES=0 bash train.sh

Citation

If you find our paper or this project helps your research, please kindly consider citing our paper in your publication.

@article{ravaut2022towards,
  title={Towards Summary Candidates Fusion},
  author={Ravaut, Mathieu and Joty, Shafiq and Chen, Nancy F},
  journal={arXiv preprint arXiv:2210.08779},
  year={2022}
}

summafusion's Issues

About the scored_summaries in AbstractiveFusionDataset

I found that in your data set training for SummaFusion i see that the AbstractiveFusionDataset has score attributes but in the paper you don't mention about score, i thought that only Source and Summary Candidates is input for Second-stage why it's has score too and what is that score, could you figure me out?