BEpaRTy

This model is a fine-tuned version of bert-base-uncased on 2.5 million tweets from 825 U.S. congressional politicians. It is capable of predicting the political orientation of tweets. The model's testing accuracy is 89.54%, with an F1 score of 0.8939. This model serves as the initial step in the paper titled "A Two-Step Method to Classify Political Partisanship Using a Deep Learning Model."

Model description

Intended uses & limitations

This model is capable of predicting the political orientation of tweets. It classifies tweets into 0 (Democratic) and 1 (Republican).

Training and evaluation data

2.5 million tweets from 825 U.S. congressional politicians.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• optimizer: default
• training_precision: default

Training results

On the training dataset (80% of the dataset):

          precision    recall  f1-score   support

       0     0.9655    0.9782    0.9718   1001625
       1     0.9779    0.9651    0.9714   1002311

accuracy                         0.9716   2003936

macro avg 0.9717 0.9716 0.9716 2003936 weighted avg 0.9717 0.9716 0.9716 2003936

On the testing dataset (20% of the dataset): precision recall f1-score support

       0     0.8852    0.9091    0.8970    250835
       1     0.9063    0.8818    0.8939    250149

accuracy                         0.8954    500984

macro avg 0.8957 0.8954 0.8954 500984 weighted avg 0.8957 0.8954 0.8954 500984

Framework versions

• Transformers 4.26.1
• TensorFlow 2.6.5
• Tokenizers 0.13.2

Citation

Please cite the following work: Hu, L. (2024). A Two-Step Method for Classifying Political Partisanship Using Deep Learning Models. Social Science Computer Review, 42(4), 961-976.