README.md

SciBERT for Uncertainty Prediction

The repository contains the codes to fine-tune SciBERT for uncertainty prediction.

How to Use

Our trained model can predict the number of uncertainty expressions given a sentence and some meta-information about the economic journal, including gender of the authors.

Step 1) Download the trained model

Our trained model can be downloaded from this google drive link.

To download it to your server using command line, you can use the following command:

file_id="1xJ_5myOvjGKGSH3hGRLd9hu5s50ydhTO"
wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=${file_id}' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=${file_id}" -O trained_scibert_uncertainty.pt && rm -rf /tmp/cookies.txt

mv trained_scibert_uncertainty.pt models/

Step 2) Run the inference mode of the trained model

We added 10 example data points in data/example_data.csv.

You can run our trained model on this example data using the following command:

python eval.py --splits_filename data/example_data.csv

Training Procedure

Model Architecture

Our model follows the following pipeline: [Use the image from overleaf]

Training

To train a model for classification/regression tasks, use train.py:

python train.py --dataset_df_dir data/ --splits_filename train.csv val.csv test.csv \
    --text_col input --y_col label --class_weight automatic \
    --model_save_dir models/ \
    --log_dir log/ --iter_time_span 1000 \
    --pretrained_model roberta-large --lr 1e-5 --max_length 512 --csv_output_path output/roberta_large_output.csv \
    --n_epochs 5

Note: To obtain the entire training data, please contact the correspondence author of "Editing a Woman’s Voice" (2021).

Evaluation

To evaluate the trained model on new data, use eval.py:

python eval.py --splits_filename test.csv --text_col input --y_col label --num_numeric_features 0 --numeric_features_col \
        --model_load_path models/roberta_large.pt \
        --log_dir log/ \
        --csv_output_path output/roberta_large_output.csv \
        --output_type binary --max_length 512 --pretrained_model roberta-large \
        --batch_size 8 --dropout_rate 0.1