This is the implementation of the paper Efficiently Summarizing Text and Graph Encodings of Multi-Document Clusters.

1. This code is tested using Python 3.6, Pytorch 1.4, and CUDA 10.1

2. Install Apex:

   • cd ~; git clone https://github.com/NVIDIA/apex
   • cd apex
   • pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--deprecated_fused_adam" ./

3. Install nccl:

   • cd ~; git clone https://github.com/NVIDIA/nccl.git
   • cd nccl
   • make -j10 src.build CUDA_HOME=/usr/local/cuda-10.1
   • sudo apt install build-essential devscripts debhelper fakeroot
   • make -j10 pkg.debian.build CUDA_HOME=/usr/local/cuda-10.1
   • sudo apt install ./build/pkg/deb/*.deb

4. Setup fairSeq and import some files:

   • cd ~
   • pip uninstall -y enum34 # Prevent AttributeError: module 'enum' has no attribute 'IntFlag'
   • git clone --branch v0.9.0 https://github.com/pytorch/fairseq
   • mkdir ~/fairseq/models
   • cd ~/fairseq/models
   • wget 'https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz'
   • tar -xzf bart.large.tar.gz
   • cd ~/fairseq
   • wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json'
   • wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe'
   • wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt'
   • Download encoder-updated.json file from https://github.com/amazon-research/BartGraphSumm/blob/main/data/encoder-updated.json and put it under ~/fairseq

5. Install NLTK and Spacy:

   • pip install nltk spacy more_itertools
   • python -m spacy download en_core_web_sm
   • python -m nltk.downloader stopwords
   • python -m nltk.downloader punkt

6. For ROUGE:

   • sudo apt-get install -y cpanminus
   • cpanm —force XML::Parser
   • cd ~
   • pip install -U git+https://github.com/pltrdy/pyrouge
   • git clone https://github.com/pltrdy/files2rouge.git
   • cd files2rouge
   • python setup_rouge.py
   • python setup.py install
   • pyrouge_set_rouge_path ~/.files2rouge

7. Get PreSumm (for fast parallel ROUGE impl — note that this does not split summaries into sentences and therefore gives worse ROUGE-L scores than files2rouge)

   • cd ~
   • git clone https://github.com/nlpyang/PreSumm.git

8. Other:

   • mkdir -p ~/results
   • pip install rouge

1. cd ~; mkdir data; cd data

2. multi-news-500 (Preprocessed and truncated data):

   • Get it from original source and rename the folder as multi-news-500; Also, rename xxx.txt.src.yyy as xxx.source and xxx.txt.tgt.yyy as xxx.target

3. multi-news-full-clean (Preprocessed but not truncated):

   • Get it from original source and rename the folder as multi-news-full-clean; Also, rename the files inside this folder as follows: xxx.txt.src as xxx.source and xxx.txt.tgt as xxx.target

4. multi-news-full-raw (not processed and not truncated) -- This is only needed for graph construction in BART-Long-Graph models.

   • Get it from original source and rename the folder as multi-news-full-raw; Also, rename the files inside this folder as follows: xxx.src as xxx.source and xxx.tgt as xxx.target

• cd ~; git clone [email protected]:amazon-research/BartGraphSumm.git
• cd ~/BartGraphSumm/src/fairseq
• pip install --editable .
• cd ../

Try the following command to train and evaluate the BART baseline model on Multi-News-500 dataset

cd ~/BartGraphSumm/src
make -f bart-large.mk TASK=~/data/multi-news-500 OUTPUT_DIR=~/results/bart-large-multinews-model1 rouge

The ROUGE F1 scores (R1/R2/RL) can be found at

cat ~/results/bart-large-multinews-model1/test.rouge-stdout | grep ">"
>> ROUGE-F(1/2/3/l): 49.22/18.88/23.88

The scores correspond to the numbers from BART (input length=500) in Table 4 in the paper.

1. Create a longformer model
   • cd ~; cp -r ~/fairseq/models/bart.large ~/fairseq/models/bart.large.long
   • cd ~/BartGraphSumm/src; python convert_model_to_long.py --input_model_path ~/fairseq/models/bart.large/model.pt --output_model_path ~/fairseq/models/bart.large.long/model.pt
2. Create data of length 500 tokens:
   • cd ~/BartGraphSumm/src
   • python prepare_data.py --data_path ~/data/multi-news-full-clean --output_path ~/data/multi-news-500-sentmarkers --max_length 500 --sentence_level_markers

make -f bart-large-long.mk TASK=~/data/multi-news-500-sentmarkers MAX_TOKENS=1024 OUTPUT_DIR=~/results/bart-large-multinews-model2 rouge

The ROUGE F1 scores (R1/R2/RL) can be found at

cat ~/results/bart-large-multinews-model2/test.rouge-stdout | grep ">"
>> ROUGE-F(1/2/3/l): 48.54/18.56/23.78

The scores correspond to the numbers from Bart-Long in Table 1 in the paper.

1. Create BART-Long model with additional encoder for encoding the graph information
   • cp -r ~/fairseq/models/bart.large ~/fairseq/models/bart.large.long.graph.linear
   • cd ~/BartGraphSumm/src; python convert_model_to_long.py --input_model_path ~/fairseq/models/bart.large/model.pt --output_model_path ~/fairseq/models/bart.large.long.graph.linear/model.pt --linear_graph
2. Create the graph data and its linearized form
   • Create a new virtual env with pytorch 1.5 --> our graph construction code relies on latest allennlp modules which has pytorch 1.5 requirements. So, please follow the below setups in creating a new virtual environment
   • sudo apt-get install virtualenv
   • cd ~/; virtualenv -p python3.6 graph_env; source graph_env/bin/activate
   • pip install numpy allennlp==1.0.0 allennlp_models==1.0.0 networkx==2.4 matplotlib==3.3.0
   • cd ~/BartGraphSumm/src
   • python graph_construction.py --data_path ~/data/multi-news-full-raw --output_path ~/data/multi-news-full-graph --split train
   • python graph_construction.py --data_path ~/data/multi-news-full-raw --output_path ~/data/multi-news-full-graph --split val
   • python graph_construction.py --data_path ~/data/multi-news-full-raw --output_path ~/data/multi-news-full-graph --split test
   • Load back the Pytorch 1.4 environment
   • Now create the data by concatenating plan text input with graph information for each sample in the multi-news dataset: python prepare_data.py --data_path ~/data/multi-news-full-clean --output_path ~/data/multi-news-500-500 --max_length 500 --mode standard_with_graph_knowledge --graph_data_path ~/data/multi-news-full-graph --sentence_level_markers

make -f bart-large-graph-linear.mk TASK=~/data/multi-news-500-500 MAX_TOKENS=1500 MAX_EPOCH=8 OUTPUT_DIR=~/results/bart-large-multinews-model3 rouge

The ROUGE F1 scores (R1/R2/RL) can be found at

cat ~/results/bart-large-multinews-model3/test.rouge-stdout | grep ">"
>> ROUGE-F(1/2/3/l): 49.03/19.04/24.04

The scores correspond to the numbers from Bart-Long-Graph (500 tokens graph text) in Table 1 in the paper.

To create the data with 1000 tokens for graph information: python prepare_data.py --data_path ~/data/multi-news-full-clean --output_path ~/data/multi-news-500-1000 --max_length 1000 --mode standard_with_graph_knowledge --graph_data_path ~/data/multi-news-full-graph --sentence_level_markers

and train the model:

make -f bart-large-graph-linear.mk TASK=~/data/multi-news-500-1000 MAX_TOKENS=2500 MAX_EPOCH=8 LR=3e-05 OUTPUT_DIR=~/results/bart-large-multinews-model4 rouge

The ROUGE F1 scores (R1/R2/RL) can be found at

cat ~/results/bart-large-multinews-model4/test.rouge-stdout | grep ">"
>> ROUGE-F(1/2/3/l): 49.24/18.99/23.97

The scores correspond to the numbers from Bart-Long-Graph (1000 tokens graph text) in Table 1 in the paper.

See CONTRIBUTING for more information.

This library is licensed under the MIT-0 License. See the LICENSE file.

If you find this code helpful, please consider citing the following paper:

@inproceedings{pasunuru2021efficient,
    title={Efficiently Summarizing Text and Graph Encodings of Multi-Document Clusters},
    author={Pasunuru, Ramakanth and Liu, Mengwen and Bansal, Mohit and Ravi, Sujith and Dreyer, Markus},
    booktitle={NAACL},
    year={2021}
}