Skip to content

Code for ACL2020 paper "Heterogeneous Graph Neural Networks for Extractive Document Summarization"

Notifications You must be signed in to change notification settings

dqwang122/HeterSumGraph

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HeterSumGraph

Code for ACL 2020 paper: Heterogeneous Graph Neural Networks for Extractive Document Summarization

fastNLP version will come soon.

Some code are borrowed from PG and Transformer. Thanks for their work.

Thanks for issue #28 to point out the flaw of the implementation of GAT layers. The previous version ignores the hidden states of destination nodes when the source and destination nodes have different node types. Since this change will affect the released checkpoints, we update the code in dev branch.

Dependency

  • python 3.5+

  • PyTorch 1.0+

  • DGL 0.4

  • rouge 1.0.0

    • A full Python Implementation of the ROUGE Metric which is used in validation phase
  • pyrouge 0.1.3

  • others

    • nltk
    • numpy
    • sklearn

Data

We have preprocessed CNN/DailyMail, NYT50 and Multi-News datasets for TF-IDF features used in the graph creation, which you can find here.

For CNN/DailyMail and Multi-News, we also provide the json-format datasets in this link. However, due to the license, NYT(The New York Times Annotated Corpus) can only be available from LDC. And we follow the preprocessing code of Durrett et al. (2016) to get the NYT50 datasets.

The example looks like this:

{
  "text":["deborah fuller has been banned from keeping animals ... 30mph",...,"a dog breeder and exhibitor... her dogs confiscated"],
  "summary":["warning : ... at a speed of around 30mph",... ,"she was banned from ... and given a curfew "],
  "label":[1,3,6]
}

and each line in the file is an example. For the text key, the value can be list of string (single-document) or list of list of string (multi-document). The example in training set can ignore the summary key since we only use label during the training phase. All strings need be lowercase and tokenized by Stanford Tokenizer, and nltk.sent_tokenize is used to get sentences.

After getting the standard json format, you can prepare the dataset for the graph by PrepareDataset.sh in the project directory. The processed files will be put under the cache directory.

The default file names for training, validation and test are: train.label.jsonl, val.label.jsonl and test.label.jsonl. If you would like to use other names, please change the corresponding names in PrepareDataset.sh, Line 321-322 in train.py and Line 188 in evaluation.py. (Default names is recommended)

Train

For training, you can run commands like this:

python train.py --cuda --gpu 0 --data_dir <data/dir/of/your/json-format/dataset> --cache_dir <cache/directory/of/graph/features> --embedding_path <glove_path> --model [HSG|HDSG] --save_root <model path> --log_root <log path> --lr_descent --grad_clip -m 3

We also provide our checkpoints on CNN/DailyMail, NYT50 and Multi-News in this link. Besides, the outputs can be found here(NYT50 has been removed due to its license).

Test

For evaluation, the command may like this:

python evaluation.py --cuda --gpu 0 --data_dir <data/dir/of/your/json-format/dataset> --cache_dir <cache/directory/of/graph/features> --embedding_path <glove_path>  --model [HSG|HDSG] --save_root <model path> --log_root <log path> -m 3 --test_model multi --use_pyrouge

Some options:

  • use_pyrouge: whether to use pyrouge for evaluation. Default is False (which means rouge).
    • Please change Line17-18 in tools/utils.py to your own ROUGE path and temp file path.
  • limit: whether to limit the output to the length of gold summaries. This option is only set for evaluation on NYT50 (which uses ROUGE-recall instead of ROUGE-f). Default is False.
  • blocking: whether to use Trigram blocking. Default is False.
  • save_label: only save label and do not calculate ROUGE. Default is False.

To load our checkpoint for evaluation, you should put it under the save_root/eval/ and make the name for test_model to start with eval. For example, if your save_root is "checkpoints", then the checkpoint "cnndm.ckpt" should be put under "checkpoints/eval" and the test_model is evalcnndm.ckpt.

ROUGE Installation

In order to get correct ROUGE scores, we recommend using the following commands to install the ROUGE environment:

sudo apt-get install libxml-perl libxml-dom-perl
pip install git+git://github.com/bheinzerling/pyrouge
export PYROUGE_HOME_DIR=the/path/to/RELEASE-1.5.5
pyrouge_set_rouge_path $PYROUGE_HOME_DIR
chmod +x $PYROUGE_HOME_DIR/ROUGE-1.5.5.pl

You can refer to https://github.com/andersjo/pyrouge/tree/master/tools/ROUGE-1.5.5 for RELEASE-1.5.5 and remember to build Wordnet 2.0 instead of 1.6 in RELEASE-1.5.5/data:

cd $PYROUGE_HOME_DIR/data/WordNet-2.0-Exceptions/
./buildExeptionDB.pl . exc WordNet-2.0.exc.db
cd ../
ln -s WordNet-2.0-Exceptions/WordNet-2.0.exc.db WordNet-2.0.exc.db

About

Code for ACL2020 paper "Heterogeneous Graph Neural Networks for Extractive Document Summarization"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published