Structure-Augmented Keyphrase Generation

This PyTorch code was used in the experiments of the research paper.

• Structure-Augmented Keyphrase Generation. Jihyuk Kim, Myeongho Jeong, Seungtaek Choi, and Seung-won Hwang (EMNLP'21).

Targeting keyphrase generation task, given a document as input, we first extend the given document with related but absent keyphrases from existing keyphrases, to augment missing contexts (generating structure), and then, build a graph of keyphrases and the given document, to obtain structure-aware representation of the augmented text (encoding structure).

If there are any questions, please send Jihyuk Kim an email: [email protected].

Open set KG

For open set KG, we experimented using KP20k, scientific publication dataset. Most of the codes are adapted from keyphrase-generation-rl (Neural Keyphrase Generation via Reinforcement Learning with Adaptive Rewards. ACL 2019. Chan et al). We used preprocessed dataset from KG-KE-KR-M (An Integrated Approach for Keyphrase Generation via Exploring the Power of Retrieval and Extraction. NAACL 2019. Chen et al).

For experiments, you can follow the scripts below.

1. Download & Preprocess data

wget https://www.dropbox.com/s/lgeza7owhn9dwtu/Processed_data_for_onmt.zip?dl=1
unzip Processed_data_for_onmt.zip\?dl\=1 
rm Processed_data_for_onmt.zip\?dl\=1 

mkdir -p data/kp20k_filtered
# train data
cp data/Processed_data_for_onmt/Training/word_kp20k_training_context_filtered.txt data/kp20k_filtered/train_src.txt
cp data/Processed_data_for_onmt/Training/word_kp20k_training_context_nstpws_sims_retrieved_keyphrases_filtered.txt data/kp20k_filtered/train_ret.txt
cp data/Processed_data_for_onmt/Training/word_kp20k_training_keyword_filtered.txt data/kp20k_filtered/train_trg.txt
# validation data
cp data/Processed_data_for_onmt/Validation/word_kp20k_validation_context_filtered.txt data/kp20k_filtered/valid_src.txt
cp data/Processed_data_for_onmt/Validation/word_kp20k_validation_context_nstpws_sims_retrieved_keyphrases_filtered.txt data/kp20k_filtered/valid_ret.txt
cp data/Processed_data_for_onmt/Validation/word_kp20k_validation_keyword_filtered.txt data/kp20k_filtered/valid_trg.txt
# test data
cp data/Processed_data_for_onmt/Testing/word_kp20k_testing_context.txt data/kp20k_filtered/test_src.txt
cp data/Processed_data_for_onmt/Testing/word_kp20k_testing_context_nstpws_sims_retrieved_keyphrases_filtered.txt data/kp20k_filtered/test_ret.txt
cp data/Processed_data_for_onmt/Testing/word_kp20k_testing_keyword.txt data/kp20k_filtered/test_trg.txt

rm -r data/Processed_data_for_onmt*

# Preprocessing
# w/o title
cd open_set
python preprocess.py -data_dir ../data/kp20k_filtered -vocab_size 50000
# w/ title
mkdir -p ../data/kp20k_filtered_title
cp ../data/kp20k_filtered/*.txt ../data/kp20k_filtered_title/
python preprocess.py -data_dir ../data/kp20k_filtered_title -vocab_size 50000 -use_title
cd ../

2. Training

cd open_set
source train.sh [device] [enc_layers] [title]

• [device] denotes GPU index, used as CUDA_VISIBLE_DEVICES=$device.
• [enc_layers] denotes the number of GCN layers. We used 3 in our experiments.
• [title] denotes whether to use title information. For example, source train.sh 0 3 title uses title information, and source train.sh 0 3 do not use title.

3. Prediction and Evaluation

cd open_set
source predict.sh [device] [enc_layers] [title]

Cite the Paper!

To cite the paper splits, please use this BibTex:

@inproceedings{kim2021structure,
  title={Structure-Augmented Keyphrase Generation},
  author={Kim, Jihyuk and Jeong, Myeongho and Choi, Seungtaek and Hwang, Seung-won},
  booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
  pages={2657--2667},
  year={2021}
}