This repo contains the resources for implementing the paper "Zero-Shot Readability Assessment of Korean ESG Reports using BERT."(pending for results).
The paper constructs a small binary dataset for benchmarking purposes.
The dataset is consisted of text, text_length and label(ESG, NEWS). ESG label stands for Korean ESG reports collected from KSA, and NEWS label stands for Korean Finance News inherited from the KLUE Dataset. Below are the details of the dataset. (Codes used to preprocess the text can be found at functions/tp_func)
| #Label | #Num of Data | #Average Length |
|---|---|---|
| ESG | 1,387 | 16,879 |
| NEWS | 1,500 | 917 |
The functions include implementations of readability assessment scores (FOG, RSRS). It also includes novel readability assessment scores introduced from our paper: sentimentAssessment and biRSRS. Both scores leverage pretrained language models, and is designed to function in a zero-shot manner, not requiring additional labeld data or training of any sort.
Guijin Son, Naeun Yoon, Kaeun Lee
If you have any questions please feel free to reach out at spthsrbwls123@yonsei.ac.kr.