Measuring Bias

Current implementation has two measures:

1. Demographic statistics: Computes the number of times a given demographic group appears in a given text (eg: {'female': 7, 'male': 6}).
2. Co-occurance matrix: Computes the number of times a given demographic group term (eg: "she") co-occurs with a target term (eg: "caring") in a given input text (eg: {'('female', 'caring')': 2, ('male', 'caring'): 0, ..}).

To run the program:

python3 bias_measure.py \
--text example_text.txt \
--target_group_name professions

• Values for --target_group_name can be currently chosen from professions or adjectives.
• --text should be set to path to the .txt file containing the text to be evaluated. (eg: example_text.txt)

To save the results:

python3 bias_measure.py \
--text example_text.txt \
--target_group_name professions \
--folder_to_save_results "."

• folder_to_save_results must contain the path to the folder to save the results to as a .csv file. This is an optional argument. (eg: ./src)

To update the Target Terms:

1. Update bias_terms.py with:
   1. List containing target terms (eg: adjectives = ["reactive", "sweet", .. ])
   2. Update target_dictionary with the new List containing target terms (eg: target_dictionary = {"adjectives": adjectives, .. })

To update the Demographic Terms:

1. Update bias_terms.py with:
   1. List containing demographic terms (eg: female_terms = ["her", "she", "wife", .. ])
   2. Update/Create gender_dictionary with the new List containing demographic terms (eg: gender_dictionary = {"female": female_terms, "male": male_terms})

Requirements:

itertools pandas collections nltk