Introduction

We propose a deep learning based model and well-organized dataset for context aware paper citation recommendation. Our model comprises a document encoder and a context encoder, which uses Graph Convolutional Networks (GCN) layer and Bidirectional Encoder Representations from Transformers (BERT), which is a pretrained model of textual data. By modifying the related PeerRead,AAN dataset, we propose a new dataset called FullTextPeerRead, FullTextANN containing context sentences to cited references and paper metadata.

The code is based on that(BERT, GCN).

Paper

A Context-Aware Citation Recommendation Model with BERT and Graph Convolutional Networks

• Main Result

Data

• Full Context PeerRead : Created by processing allenai-PeerRead
• Full Context ANN : Created by processing arXiv Vanity (Not disclosed due to copyright.)

There are two types of data, AAN and PeerRead. Both columns are identical.

Header	Description
target_id	citing paper id
source_id	cited paper id
left_citated_text	text to the left of the citation tag when citing
right_citated_text	text to the right of the citation tag when citing
target_year	release target paper year
source_year	release source paper year

run_classifier.py

The main script to train bert, bert-gcn model

python run_classifier.py [options]

• General Parameters:

  • --model (Required): The mode to run the run_classifier.py script in. Possible values: bert or bert_gcn
  • --dataset (Required): The dataset to run the run_classifier.py script in. Possible values: AAN or PeerRead
  • --frequency (Required): Parse datasets more frequently
  • --max_seq_length : Length of cited text to use
  • --gpu : The gpu to run code

• BERT Parameters: You can refer to it here.

  • --do_train, --do_predict, --data_dir, --vocab_file, --bert_config_file, --init_checkpoint, ...

gcn_pretrain.py

If you want to use bert-gcn you have to run it.

python gcn_pretrain.py [options]

 

• GCN Parameters: You can refer to it here.
  • --gcn_model, --gcn_lr, --gcn_epochs, --gcn_hidden1, --gcn_hidden2, ...