HabNet

The code and data for our paper "Hierarchical Bi-Directional Self-Attention Networks for Paper Review Rating Recommendation" which is accepted by COLING 2020.

Code overview

There are two folders of code: HabNet is the code for the main task of predicting final acceptance decisions for papers; HabNet_MC is the code for the sub_task of predicting ratings for reviews. The steps to run them are the same as shown in following section "Usage".

Dataset

The OpenReview dataset (processed) collected by us are in the folder "OpenReview_Data" which includes two files "ICLR_Review_all_with_decision_processed.csv" and "ICLR_Review_all_processed.csv". There is also a copy of these two files in the "data" folder of HabNet and HabNet_MC respectively. "ICLR_Review_all_with_decision_processed.csv" is used for the main task of predicting acceptance decsions for papers, its copy is located in "data" folder of HabNet; "ICLR_Review_all_processed.csv" is used for the sub-task of predicting ratings for reviews, its copy is located in "data" folder of HabNet_MC.

Code

Requirements

• Python 3.6.8
• Tensorflow = 1.13.1
• Pandas
• Nltk
• Tqdm
• Glove pre-trained word embeddings

Usage

First step: run script to prepare the data:

python data_prepare.py

Second step: train and evaluate the model:
(make sure Glove embeddings are ready before training, put glove.6B.50d.txt in HabNet, and put glove.6B.100d.txt in HabNet_MC)

wget http://nlp.stanford.edu/data/glove.6B.zip
unzip glove.6B.zip

python train.py

Citation

If you use our code or dataset, please cite our COLING 2020 paper, our paper is available at: https://www.aclweb.org/anthology/2020.coling-main.555/

Acknowledgements

Our code is based on DiSAN, we thank the authors of DiSAN for their open-source code.