---
license: apache-2.0
---

This model is a binary classifier developed to analyze comment authorship patterns on Korean news articles.
For further details, refer to our paper on Journalism: [News comment sections and online echo chambers: The ideological alignment between partisan news stories and their user comments](https://journals.sagepub.com/doi/full/10.1177/14648849211069241)

* This model is a BERT classification model to classify Korean user generated comments into binary labels of liberal or conservative.
* This model was trained on approximately 37,000 user generated comments collected from NAVER\'s news portal. The dataset was collected in 2019; as such, note that comments related to recent political topics might not be classified correctly.
* This model is a finetuned model based on ETRI\'s KorBERT. 


### How to use
* The model requires an edited version of the transformers class `BertTokenizer`, which can be found in the file `KorBertTokenizer.py`.
* Usage example:

~~~python
from KorBertTokenizer import KorBertTokenizer
from transformers import BertForSequenceClassification
import torch

tokenizer = KorBertTokenizer.from_pretrained('conviette/korPolBERT')
model = BertForSequenceClassification.from_pretrained('conviette/korPolBERT')

def classify(text):
    inputs = tokenizer(text, padding='max_length', max_length=70, return_tensors='pt')

    with torch.no_grad():
        logits=model(**inputs).logits
        predicted_class_id = logits.argmax().item()
        return model.config.id2label[predicted_class_id]


input_strings = ['좌파가 나라 경제 안보 말아먹는다',
                 '수꼴들은 나라 일본한테 팔아먹었냐']

for input_string in input_strings:
    print('===\n입력 텍스트: {}\n분류 결과: {}\n==='.format(input_string, classify(input_string)))
~~~


### Model performance

* Accuracy: 0.8322
* F1-Score: 0.8322
* For further technical details on the model, refer to our paper for the W-NUT workshop (EMNLP 2019), [The Fallacy of Echo Chambers: Analyzing the Political Slants of User-Generated News Comments in Korean Media](https://aclanthology.org/D19-5548/).