Our icon was made by an AI with craiyon.com
Downloading evaluation data from reddit and evaluating it
in main.py results in the following graphs which show that the
judgement mechanism in use - detoxify - rates the texts
as much less toxic and obscene after safeworder has been applied
| Less obscene subreddit | Obscene subreddit |
|---|---|
 |
 |
Use this tool if you want to replace nsfw, profane, toxic, offensive words or expressions with a
more appropriate content. You can use it out of the box or customize it to your needs.
In some cases you may want to not just replace the words with characters like * or -. This
package allows you to replace the words with any other expression you want.
You will also get the indices of where replacement happens which allows you to
highlight, replace, remove or alter the words in the original text.
pip install safeworder
from safeworder import NSFWReplacer
rp = NSFWReplacer()
safe_text, replacements = rp.replace("You are an idiot")
print(safe_text, replacements)
>>> "You are a simpleton", {(11, 16): "simpleton"}
As a return value you get a tuple of the safer text and a dictionary containing the ranges in the original string where a replacement happened and with which safe expression.
Either create an excel (.xlsx) file that and in the table Sheet1 specify the safe expressions in the first row that expressions in the same column should be replaced with, or use a json file that maps each safe expression to a list of its unsafe counterparts. Have a look at safeworder/mappings/ or tests/obscenity.json for reference and use the following code to replace the toxicity and obscenity dictionary respectively.
from safeworder import NSFWReplacer
orep = "obsc.json"
trep = "toxic.xlsx"
rp = NSFWReplacer(obscenity_replacement=orep, toxicity_replacement=trep)
If detoxify does not flag the text, nothing will be replaced. Some words have a negative connotation in general from how detoxify was trained, but in everyday use the word does not have to be profane. Here is how to offset this higher baseline suspicion of the model.
print(r.replace("You suck!"))
>>> ('You inhale!', {(4, 8): 'inhale'})
print(r.replace("he was sucking lemonade through the straw"))
>>> ('he was inhaling lemonade through the straw', {(7, 14): 'inhaling'})
As one can see in the second case suck should not be replaced. Here is the fix:
tolerances_suck = {"suck": 0.98, "sucks": 0.98, "sucked": 0.98, "sucking": 0.98}
r = NSFWReplacer(tolerances=tolerances_suck)
print(r.replace("You suck!"))
>>> ('You inhale!', {(4, 8): 'inhale'})
print(r.replace("he was sucking lemonade through the straw"))
>>> ('he was sucking lemonade through the straw', {})