Home

Awesome

Extreme Summarization

This repository contains data and code for our EMNLP 2018 paper "Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization". Please contact me at shashi.narayan@gmail.com for any question.

Please cite this paper if you use our code or data.

@InProceedings{xsum-emnlp,
  author =      "Shashi Narayan and Shay B. Cohen and Mirella Lapata",
  title =       "Don't Give Me the Details, Just the Summary! {T}opic-Aware Convolutional Neural Networks for Extreme Summarization",
  booktitle =   "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing ",
  year =        "2018",
  address =     "Brussels, Belgium",
}

Extreme Summarization (XSum) dataset

You can always build the dataset using the instructions below. The original dataset is also available upon request.

Instructions to download and preprocess the extreme summarization dataset are here.

Looking for a Running Demo of Our System?

A running demo of our abstractive system can be found here.

Pretrained models and Test Predictions (Narayan et al., EMNLP 2018)

Topic-Aware Convolutional Model for Extreme Summarization

This repository contains PyTorch code for our Topic-ConvS2S model. Our code builds on an earlier copy of Facebook AI Research Sequence-to-Sequence Toolkit.

We also release the code for the ConvS2S model. It uses optimized hyperparameters for extreme summarization. Our release facilitates the replication of our experiments, such as training from scratch or predicting with released pretrained models, as reported in the paper.

Installation

Our code requires PyTorch version 0.4.0 or 0.4.1. Please follow the instructions here: https://github.com/pytorch/pytorch#installation.

After PyTorch is installed, you can install ConvS2S and Topic-ConvS2S:

# Install ConvS2S
cd ./XSum-ConvS2S
pip install -r requirements.txt
python setup.py build
python setup.py develop

# Install Topic-ConvS2S
cd ../XSum-Topic-ConvS2S
pip install -r requirements.txt
python setup.py build
python setup.py develop

Training a New Model

Data Preprocessing

We partition the extracted datset into training, development and test sets. The input document is truncated to 400 tokens and the length of the summary is limited to 90 tokens. Both document and summary files are lowercased.

ConvS2S
python scripts/xsum-preprocessing-convs2s.py

It generates the following files in the "data-convs2s" directory:

train.document and train.summary
validation.document and validation.summary
test.document and test.summary

Lines in document and summary files are paired as (input document, corresponding output summary).

TEXT=./data-convs2s
python XSum-ConvS2S/preprocess.py --source-lang document --target-lang summary --trainpref $TEXT/train --validpref $TEXT/validation --testpref $TEXT/test --destdir ./data-convs2s-bin --joined-dictionary --nwordstgt 50000 --nwordssrc 50000

This will create binarized data that will be used for model training. It also generates source and target dictionary files. In this case, both files are identical (due to "--joined-dictionary") and have 50000 tokens.

Topic-ConvS2S
python scripts/xsum-preprocessing-topic-convs2s.py

It generates the following files in the "data-topic-convs2s" directory:

train.document, train.summary, train.document-lemma and train.doc-topics
validation.document, validation.summary, validation.document-lemma and validation.doc-topics
test.document, test.summary, test.document-lemma and test.doc-topics

Lines in document, summary, document-lemma and doc-topics files are paired as (input document, output summary, input lemmatized document, document topic vector).

TEXT=./data-topic-convs2s
python XSum-Topic-ConvS2S/preprocess.py --source-lang document --target-lang summary --trainpref $TEXT/train --validpref $TEXT/validation --testpref $TEXT/test --destdir ./data-topic-convs2s --joined-dictionary --nwordstgt 50000 --nwordssrc 50000 --output-format raw

This will generate source and target dictionary files. In this case, both files are identical (due to "--joined-dictionary") and have 50000 tokens. It operates on the raw format data.

Model Training

By default, the code will use all available GPUs on your machine. We have used CUDA_VISIBLE_DEVICES environment variable to select specific GPU(s).

ConvS2S
CUDA_VISIBLE_DEVICES=1 python XSum-ConvS2S/train.py ./data-convs2s-bin --source-lang document --target-lang summary --max-sentences 32 --arch fconv --criterion label_smoothed_cross_entropy --max-epoch 200 --clip-norm 0.1 --lr 0.10 --dropout 0.2 --save-dir ./checkpoints-convs2s --no-progress-bar --log-interval 10
Topic-ConvS2S
CUDA_VISIBLE_DEVICES=1 python XSum-Topic-ConvS2S/train.py ./data-topic-convs2s --source-lang document --target-lang summary --doctopics doc-topics --max-sentences 32 --arch fconv --criterion label_smoothed_cross_entropy --max-epoch 200 --clip-norm 0.1 --lr 0.10 --dropout 0.2 --save-dir ./checkpoints-topic-convs2s --no-progress-bar --log-interval 10

Generation with Pre-trained Models

ConvS2S

CUDA_VISIBLE_DEVICES=1 python XSum-ConvS2S/generate.py ./data-convs2s --path ./checkpoints-convs2s/checkpoint-best.pt --batch-size 1 --beam 10 --replace-unk --source-lang document --target-lang summary > test-output-convs2s-checkpoint-best.pt

Make sure that ./data-convs2s also has the source and target dictionary files.

Topic-ConvS2S

CUDA_VISIBLE_DEVICES=1 python XSum-Topic-ConvS2S/generate.py ./data-topic-convs2s --path ./checkpoints-topic-convs2s/checkpoint_best.pt --batch-size 1 --beam 10 --replace-unk --source-lang document --target-lang summary --doctopics doc-topics --encoder-embed-dim 512 > test-output-topic-convs2s-checkpoint-best.pt

Make sure that ./data-topic-convs2s has the test files to decode, the source and target dictionary files.

Extract final hypothesis

python scripts/extract-hypothesis-fairseq.py -o test-output-convs2s-checkpoint-best.pt -f final-test-output-convs2s-checkpoint-best.pt
python scripts/extract-hypothesis-fairseq.py -o test-output-topic-convs2s-checkpoint-best.pt -f final-test-output-topic-convs2s-checkpoint-best.pt