Exclude "bad words" from Corpora? #256
Comments
|
Wordfilter seems to be picking up on word components that are slurs (i.e.,
the scunthorpe problem). You could probably lower the false positive rate
by splitting by non-alphabetic characters & anchoring your check to the
start and end of each string.
(It seems like only the word clues contain actual slurs. Some of them are
probably being used in the more common non-slur context: specifically
'spade', 'spook', and 'gash'. I'm not sure if we care about being
context-aware here, since it's pretty easy to take corpus elements out of
context.)
…On Fri, Mar 17, 2017 at 2:56 AM Hugo ***@***.***> wrote:
I ran wordfilter <https://github.com/dariusk/wordfilter> on Corpora:
./data/foods/herbs_n_spices.json may contain bad words: [u'spices']
./data/religion/fictional_religions.json may contain bad words: [u'Esoteric Order of Dagon']
./data/religion/religions.json may contain bad words: [u'Ancient Egyptian religion', u'Syrian-Egyptic Gnosticism']
./data/societies_and_groups/animal_welfare.json may contain bad words: [u'Japan', u'Pakistan', u'Egypt']
./data/words/word_clues/clues_five.json may contain bad words: [u'spade', u'tardy', u'blame', u'spice', u'flame', u'spook']
./data/words/word_clues/clues_four.json may contain bad words: [u'gash', u'lame', u'gimp']
./data/words/word_clues/clues_six.json may contain bad words: [u'spooks', u'spooky', u'blames', u'flames', u'script', u'spices', u'retard']
(This excludes keys matching "Description", "description", "descriptions",
"scripts", "wine_descriptions".)
1. Should any of those found words be removed from Corpora?
2. Would it be useful to pop this into the CI?
a. Just for information, or
b. To fail a build if it finds something?
If yes to 2a, should some of those be added to a whitelist?
If yes to 2b, there needs to be a whitelist for any which aren't removed.
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#256>, or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAd6GVeeOK0dqL9OyyY4zac_gvuQpbXgks5rmi6YgaJpZM4MgPaA>
.
|
|
This may also be relevant to the topic: |
|
Yes, wordfilter falls for Scunthorpe by design:
|
|
Which makes sense for most applications of wordfilter -- with twitter bots,
we have both automation and lack of oversight, so the threat model is that
offensive content will be produced either by accident or through the
manipulation of some griefer, and the loss from rejecting even a false
positive is no big deal. For corpora (where everything must be approved by
a human anyhow, and where every entry is the result of manual effort), the
threat model is that someone has accidentally failed to scrub a word from a
list & the person approving the request also failed to notice.
Option 2b is likely to create a lot of work (and possibly a lot of strange
gaps, if whitelists don't get updated often enough) while not actually
providing more utility than option 2a. Changing the filter rules to
eliminate some false positives will make 2b more reasonable.
Someone submitting a pull request to this repo is very unlikely to be
carefully constructing a list of hybrid slurs that wordfilter will miss and
injecting them into otherwise innocent-looking json files. Someone on
twitter absolutely would do that kind of thing.
…On Fri, Mar 17, 2017 at 12:02 PM Hugo ***@***.***> wrote:
Yes, wordfilter falls for Scunthorpe by design:
Also note that due to the complexities of the English language, I am
considering anything containing the substring of a bad word to be
blacklisted. For example, even though "homogenous" is not a bad word, it
contains the substring "homo" and it gets filtered. The reason for this is
that new slang pops up all the time using compound words and I can't
possibly keep up with it. I'm willing to lose a few words like "homogenous"
and "Pakistan" in order to avoid false negatives.
https://github.com/dariusk/wordfilter#documentation
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#256 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAd6GcJ5inZA-LQYbqZdwo9a6RSugxhYks5rmq5JgaJpZM4MgPaA>
.
|
|
Yup -- I considered automatically testing for |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I ran wordfilter on Corpora:
(This excludes keys matching "Description", "description", "descriptions", "scripts", "wine_descriptions".)
a. Just for information, or
b. To fail a build if it finds something?
If yes to 2a, should some of those be added to a whitelist?
If yes to 2b, there needs to be a whitelist for any which aren't removed.
The text was updated successfully, but these errors were encountered: