/language-model-bias

Analyzing and reducing bias in word-level language models

Primary LanguageJupyter Notebook

The model is run using the AWD LSTM language model

To run the model with main.py with bias regularization -

  • Set the value of the the booleans --bias_reg_encoder and --bias_reg_decoder as True
  • Set the value of --bias_reg_en_factor and --bias_reg_de_factor
  • Input the --gender_pair_file from the repo if you are using PTB, wikitext-2 or CNN/Dailymail. If you are using a different corpus, you will need to create your own based on the corpus.

To run the bias scores on a text corpus -

  • `python fixed_context_bias_score.py filepath dm -w 100 -n 1
  • `python infinite_context_bias_score.py filepath dm -w 100 -n 1