Skip to content

ymcui/NLP-Review-Scorer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

NLP Review Scorer

Disclaimer: This is only a toy. You should seriously treat your rebuttal despite the what scores are given below. Wish you good luck with your paper submission!

Also, as the notebook will run under YOUR CONTROL, please rest assured that your review won't be recorded in any form and I have no access to it.

I know some of you are thinking about how to convert paper review to a numerical score. Yes, the time has come.

In this notebook, you will be able to convert your paper review to overall score (hopefully in range 1~5) as well as reviewer confidence.

In my own experience, the prediction on reviewer confidence is not that accurate.

News

July 12, 2019 New model trained on 5.7K reviews is available. Seems to be more accurate.

July 11, 2019 Initial version released, trained on 3K reviews.

Quick Introduction

The model is trained on real reviews from PeerRead dataset as well as in-house collected reviews for training. Note that, we only include the reviews with open access, and the private reviews without author permissions are not included. The implementation was based on run_classifier.py in BERT repository with slight modifications.

As the review data is rather private, I won't be able to release them.

Prerequisites

How-To

  1. Copy (do not need to download) the one of the following model to your Google Drive.
Model Training Data MAE @ Dev Link
v2 (latest) 5.7k 0.35 Google Drive
v1 3k 0.5 Google Drive
  1. Then, go to Google Colab for further instructions

Sample Output (v2 version)

Note that, in real situations, your input review will be much longer than these examples!

***********REVIEW**************
This is a very good paper, outstanding paper, brilliant paper.
I have never seen such a good paper before.
It was well-written and the models are novel.
The evaluations are sound and the results achieve state-of-the-art performance.
It should be definitely accepted or I will be angry.
***********SCORE***************
Paper	Recommendation	Confidence
EMNLP	4.5141506	3.8331783
********************************

***********REVIEW**************
The paper was rather bad that I don't want to see it again.
The idea was trivial and the evaluations are not convincing to me at all.
We should reject this paper or I won't review for this venue in the future.
***********SCORE***************
Paper	Recommendation	Confidence
EMNLP	1.3770846	4.0270653
********************************

Disclaimer

This is not a product by Joint Laboratory of HIT and iFLYTEK Research (HFL).

Acknowledgement

I personally thank Google Colab for providing free computing resources for researchers.

Issue

If there is any problem, please submit a GitHub Issue.