Live site (may take a few moments to load): https://literary-gender-bender.herokuapp.com/
This is a website that flips text so that all male characters become female and all female characters become male. At a high level, it does so by switching out all of the male pronouns and names for female pronouns and names and vice versa.
This is the second iteration of a project concieved in a hybrid art and computer science class. The original prompt was to create a book consisting of generative text. My book contained genderbent versions of classic novels. The original project can be seen at: http://cmuems.com/2018/60212f/yalbert/11/16/yalbert-book/
The the original project occcured over a week long assignment. This second iteration involved improving the genderbending algorithm, reorganizing the code, and formatting it for web. This was done over roughly half a week. (Optimizations for the algorithm were done here:https://github.com/maayanalbert/gender-bender-local)
Genderbending algorithm written in Python -> Django backend -> Boostrap frontend -> hosted on Heroku
Many of the files in this repository are part of the basic Django framework. The interesting ones are:
- gender_bender_web/genderbender/templates/genderbender/index.html (frontend)
- gender_bender_web/genderbender/views.py (backend)
- gender_bender_web/genderbender/genderBender.py (genderbending algorithm)
- gender_bender_web/genderbender/nameDictMaker.py (genderbending algorithm)
- gender_bender_web/genderbender/pronounDictMaker.py (genderbending algorithm)
I used a find and replace system that basically identified all of the gendered pronouns and names and replaced them with those of the opposite gender. Although this method fails on some edge cases due to irregularity in the English language (ML would be the best way to yeild completely accurate results) it works for most texts.
Major considerations I had when building this were efficiency (because I was potentially processing long strings) and how to acheive beleivable name replacements.
There are four primary stages to this genderbending algorithm, 1) parsing the text 2) compiling a dictionary of pronouns mapped to their opposite gendered equivalents 3) compiling a dictionary of names mapped to their opposite gendered equivalents and 4) replacing original names and pronouns based on the aforementioned dictionaries.
If we were to just use the raw text in its original format (a string), replacing the words in the text would be extremely costly because one would have to rewrite the entire string with each word replacement. Therefore, I created a custom string splitting function that clumps all words together and isolates all non-letters (eg: ["Harry", ":", " ", "the", " ", "wizard"]). This makes replacements much cheaper and made it easier to search for names in order to compile the name dictionary.
Relevant files:
- gender_bender_web/genderbender/genderBender.py
This step was pretty straightforward. I simply read in a txt file containing matching pronouns and compiled a dictionary based on its contents. A consideration when adding pronouns to the dictionary was making sure I had all iterations of the word (uppercase, lowercase, plural, etc).
Relevant files:
- gender_bender_web/genderbender/pronounDictMaker.py;
- gender_bender_web/genderbender/pronoun_corpus/pronouns.txt
This was the most challenging part of the project. I struggled to find a decent corpus of names until I came accross records from the US Social Security Administristration that contained the 10000 most popular baby names arranged in order of populary since 1880 (https://www.ssa.gov/oact/babynames/limits.html).
To find name replacements, I first iterated through the word array described above and picked out any word that appeared in the name file for the year the book was written. For each name found, I went through all of the opposite gendered names and used Levenshtein's distance algorithm to create a shortlist of those that were similar to the original. I then picked the most popular name from the shortlist and mapped it to the original name.
To speed up this process, I partitioned the names from the corpus into subdictionaries based first letter (eg. A={Amy, Amanda, Ada}, B={Bernadine, Becky}, etc). Therefore, in order to find the opposite gendered equivalent, I only had to look through the names with the same letter as the original.
Relevant files:
- gender_bender_web/genderbender/nameDictMaker.py;
- gender_bender_web/genderbender/name_corpus/*
This part was also quite straightforward. I simply iterated through the word array and if a word showed up as a key in either dictionary, I mapped it to its corresponding value.
Relevant files:
- gender_bender_web/genderbender/genderBender.py
From a technical standpoint, I liked how there wasn't allways a clear answer for how best to solve the problems. Most of the project was self driven, so I had to figure out where I was going to get my data and how I wanted to structure it on my own. I found the process of solving these problems and seeing the results of my work very rewarding.
From a social standpoint, I like how this project raises awareness about gender expectations. Reading through the genderbent novels, I found certain characters strange before realizing that it was because their genders had been switched. Through reading genderbent versions of familiar novels or articles, this site can help people see and come to terms with their own gender biases.
If I were to work on this project more, I'd like to continue testing and improving the genderbending algorithm. In addition, I'd like to explore creating a browser plugin that genderbends all of the text on a given website.