tomhosking / exconsumm

Extractive and Compressive Neural Summarization Based on Summary State Representations (NAACL 2019)

Under Construction - please wait

ExCoNSumm :Jointly Extracting and Compressing Documentswith Summary State Representations

This repository releases our code for ExConSum. It uses dynet and the code is in C++.

Please contact [email protected] or [email protected] for any question.

Data

CNN and Dailymail data

The datasets and the word emebeddings are the same as for REFRESH except for the compressed version. You can get them from https://github.com/EdinburghNLP/Refresh.

The Compressive and extactive Oracles dataset used in the paper are available at:

wget ftp://"ftp.priberam.pt|anonymous"@ftp.priberam.pt/SUMMAPublic/Corpora/Exconsumm/ExCoNSum-CNN-DailyMail-Data.tgz

Training

./exconsumm -cconfig/extractive_cnn.cfg -t --dynet-autobatch 1 --dynet-mem 7500

Evaluation

./exconsumm -coutput/.cfg --dynet-autobatch 1 --dynet-mem 7500

Build the system

Install dynet prerequisites eigen and MKL check https://dynet.readthedocs.io/en/latest/install.html.

The first time you clone the repository, you need to sync the dynet/ submodule.

Select the backend eiher "cuda" for GPU or "eigen" for CPU, on CPU you should use MKL as described in the dynet instalation page above.

git submodule init
git submodule update

mkdir build
cd build
cmake .. -DEIGEN3_INCLUDE_DIR=/path/to/eigen -DBACKEND=cuda|eigen
make -j2

System described in the paper:

Jointly Extracting and Compressing Documents with Summary State Representations, Afonso Mendes, Shashi Narayan, Sebastião Miranda, Zita Marinho, André F. T. Martins and Shay B. Cohen (NAACL 2019, accepted, to appear). Also at arXiV: https://arxiv.org/abs/1904.02020