PaperRobot: Incremental Draft Generation of Scientific Ideas

PaperRobot: Incremental Draft Generation of Scientific Ideas [Sample Output]

Accepted by 57th Annual Meeting of the Association for Computational Linguistics (ACL 2019)

Table of Contents

• Overview
• Requirements
• Quickstart
• Citation

Overview

Requirements

Environment:

• Python 3.6 CAUTION!! The model might not be saved and loaded properly under Python 3.5
• Ubuntu 16.04/18.04 CAUTION!! The model might not run properly on windows because windows uses backslashes on the path while Linux/OS X uses forward slashes

Pacakges

You can click the following links for detailed installation instructions.

• Pytorch 1.1
• NumPy 1.16.3
• SciPy 1.2.1
• NetworkX 2.3

Data:

• PubMed Paper Reading Dataset This dataset gathers 14,857 entities, 133 relations, and entities corresponding tokenized text from PubMed. It contains 875,698 training pairs, 109,462 development pairs, and 109,462 test pairs.

• PubMed Term, Abstract, Conclusion, Title Dataset This dataset gathers three types of pairs: Title-to-Abstract (Training: 22,811/Development: 2095/Test: 2095), Abstract-to-Conclusion and Future work (Training: 22,811/Development: 2095/Test: 2095), Conclusion and Future work-to-Title (Training: 15,902/Development: 2095/Test: 2095) from PubMed. Each pair contains a pair of input and output as well as the corresponding terms(from original KB and link prediction results).

Quickstart

Existing paper reading

CAUTION!! Because the dataset is quite large, the training and evaluation of link prediction model will be pretty slow.

Preprocessing:

Download and unzip the paper_reading.zip from PubMed Paper Reading Dataset . Put paper_reading folder under the Existing paper reading folder.

Training

Hyperparameter can be adjusted as follows: For example, if you want to change the number of hidden unit to 6, you can append --hidden 6 after train.py

python train.py

To resume training, you can apply the following command and put the previous model path after the --model

python train.py --cont --model models/GATA/best_dev_model.pth.tar

Test

Put the finished model path after the --model The test.py will provide the ranking score for the test set.

python test.py --model models/GATA/best_dev_model.pth.tar

New paper writing

Preprocessing:

Download and unzip the data_pubmed_writing.zip from PubMed Term, Abstract, Conclusion, Title Dataset . Put data folder under the New paper writing folder.

Training

Put the type of data after the --data_path. For example, if you want to train an abstract model, put data/pubmed_abstract after --data_path. Put the model directory after the --model_dp

python train.py --data_path data/pubmed_abstract --model_dp abstract_model/

To resume training, you can apply the following command and put the previous model path after the --model

python train.py --data_path data/pubmed_abstract --cont --model abstract_model/memory/best_dev_model.pth.tar

For more other options, please check the code.

Test

Put the finished model path after the --model The test.py will provide the score for the test set.

python test.py --data_path data/pubmed_abstract --model abstract_model/memory/best_dev_model.pth.tar

Predict an instance

Put the finished model path after the --model The input.py will provide the prediction for customized input.

python input.py --data_path data/pubmed_abstract --model abstract_model/memory/best_dev_model.pth.tar

Citation

@inproceedings{wang-etal-2019-paperrobot,
    title = "{P}aper{R}obot: Incremental Draft Generation of Scientific Ideas",
    author = "Wang, Qingyun  and
      Huang, Lifu  and
      Jiang, Zhiying  and
      Knight, Kevin  and
      Ji, Heng  and
      Bansal, Mohit  and
      Luan, Yi",
    booktitle = "Proceedings of the 57th Conference of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1191",
    pages = "1980--1991"
}