kedz / summarization-datasets

Pre-processing and in some cases downloading of datasets for the paper "Content Selection in Deep Learning Models of Summarization."

summarization-datasets

Pre-processing and in some cases downloading of datasets for the paper "Content Selection in Deep Learning Models of Summarization."

Requires python 3.6 or greater.

To install run:

$ python setup.py install

If you haven't installed spacy before in your current environment you should also run:

python -m spacy download en

Also it might be good to set your number of OMP threads to a small number, e.g. export OMP_NUM_THREADS=2

CNN/DailyMail Dataset

To run:

python summarization-datasets/preprocess_cnn_dailymail.py \
    --data-dir data/

This will create the CNN/DM data in a directory data/cnn-dailymail. This dataset is quite large and will take a while to preprocess. Grab a coffee!

NYT Dataset

You must obtain the raw documents for this dataset from the LDC: https://catalog.ldc.upenn.edu/LDC2008T19 Assuming you have the original NYT tar file in a directory called raw_data, run the following:

cd raw_data
tar zxvf nyt_corpus_LDC2008T19.tgz
cd ..
python summarization-datasets/preprocess_nyt.py \
    --nyt raw_data/nyt_corpus \
    --data-dir data

This will create preprocessed NYT data in data/nyt/.

DUC Dataset

To obtain this data, first sign the release forms/email NIST (details here: https://duc.nist.gov/data.html).

You should obtain from NIST, two files for the 2001/2002 data and a username and password. Assuming you have the NIST data in the folder called raw_data, you should have following:

raw_data/DUC2001_Summarization_Documents.tgz
raw_data/DUC2002_Summarization_Documents.tgz

You will also need to download additional data from nist which you can do using a script that will be in your bin directory after installation:

$ duc2002-test-data.sh USERNAME PASSWORD raw_data

where USERNAME and PASSWORD should have been given to you by NIST to access their website data. This should create a file raw_data/DUC2002_test_data.tar.gz

Now run the preprocessing scripts:

python summarization-datasets/preprocess_duc_sds.py \
    --duc2001 raw_data/DUC2001_Summarization_Documents.tgz \
    --duc2002-documents raw_data/DUC2002_Summarization_Documents.tgz \
    --duc2002-summaries raw_data/DUC2002_test_data.tar.gz 
    --data-dir data

This will create preprocessed duc data in data/duc-sds/.

Reddit Dataset

To run:

python summarization-datasets/preprocess_reddit.py \
    --data-dir data/

This will create the Reddit data in a directory data/reddit.

AMI Dataset

To run:

python summarization-datasets/preprocess_ami.py \
    --data-dir data/

This will create the AMI data in a directory data/ami.

PubMed Dataset

To run:

python summarization-datasets/preprocess_pubmed.py \
    --data-dir data/

This will create the PubMed data in a directory data/pubmed.