Minority voices ‘filtered’ out of Google Natural Language Processing models(unite.ai)
unite.ai
Minority voices ‘filtered’ out of Google Natural Language Processing models
https://www.unite.ai/minority-voices-filtered-out-of-google-natural-language-processing-models/
33 comments
It was intentional at least in so far as that bias is a known problem with Google's machine learning processes, and that no substantial effort was made to avoid the bias.
Running over a pedestrian might be a second order effect of driving a car, except that driving a car includes applying the brakes to avoid it.
Running over a pedestrian might be a second order effect of driving a car, except that driving a car includes applying the brakes to avoid it.
> an unintended second order effect from trying to remove offensive content from the corpus.
How odd that "offensive" and "minority" turned out to overlap so much. How could that have possibly happened?
How odd that "offensive" and "minority" turned out to overlap so much. How could that have possibly happened?
Use of the n-word is an obvious example.
Which Black folk use colliquially, either to imply that the addressee needs to be more humilitous or apply greater self-improvement about a subject at hand, or to re-enforce camraderie.
Also "minority" tends to be associated with a lower socioeconomic status, which may imply cruder language with more swear words in general, etc.
Sounds to me like a "damned if you do, damned if you don't" basically impossible task without making ML aware of cultural intricacies, trigger warnings, age appropriate approaches and parental controls at the same time.
Like, it's not like this filtering was done for the purpose of silencing anyone — Google (among others) really learned the hard way to not feed smut to ML models, as it _will_ get regurgitated, always as a possible PR disaster in the making:
https://www.buzzfeed.com/fionarutherford/heres-why-some-peop...
https://www.huffpost.com/entry/microsoft-tay-racist-tweets_n...
Still interesting research, but playing the devil's advocate, I think I can see why this part of the corpus was more extensively filtered away.
Sad issue, but I don't think we'd have a sane way out without massive, manual whitelisting.
Like, it's not like this filtering was done for the purpose of silencing anyone — Google (among others) really learned the hard way to not feed smut to ML models, as it _will_ get regurgitated, always as a possible PR disaster in the making:
https://www.buzzfeed.com/fionarutherford/heres-why-some-peop...
https://www.huffpost.com/entry/microsoft-tay-racist-tweets_n...
Still interesting research, but playing the devil's advocate, I think I can see why this part of the corpus was more extensively filtered away.
Sad issue, but I don't think we'd have a sane way out without massive, manual whitelisting.
actually, if your goal is "handle 99.9% of circumstances", your job is to find the edge cases and make sure they are as normal as all other cases.
thats not whats happening in ML. whats happening is "find the easiest amount of information to train a model"
thats not whats happening in ML. whats happening is "find the easiest amount of information to train a model"
I believe this is the word list that authors are objecting to the use of:
https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and...
The list does seem a bit... umm... oddly specific in places, probably due to its history as being first compiled for a photo sharing site.
https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and...
The list does seem a bit... umm... oddly specific in places, probably due to its history as being first compiled for a photo sharing site.
This is great! It has been a long time since I learned new dirty worlds. The inclusion of misspelled words and popular culture references is also intriguing.
Until today, the only thing I could imagine under "Alaskan pipeline" was, well, Alaskan pipeline. How naive I was.
Just a note on something I do have personal experience with:
The authors also observed that the text of many patents are initially obtained via imperfect examples of Optical Character Recognition (OCR), with their accompanying errors in English possibly passed through to the C4 data with little or no annotation that would distinguish it from acceptable English.
While patent offices do release PDFs of all their patent docs (and it isn't just patents; it's all the back-and-forth between the examiners & the applicant, too), a huge percentage are images of paper documents. You can always tell just by double-clicking on a word -- if the word doesn't highlight, it's an image.
OCR output from these things is generally terrible. There is no way it should ever be input to any ML model.
The authors also observed that the text of many patents are initially obtained via imperfect examples of Optical Character Recognition (OCR), with their accompanying errors in English possibly passed through to the C4 data with little or no annotation that would distinguish it from acceptable English.
While patent offices do release PDFs of all their patent docs (and it isn't just patents; it's all the back-and-forth between the examiners & the applicant, too), a huge percentage are images of paper documents. You can always tell just by double-clicking on a word -- if the word doesn't highlight, it's an image.
OCR output from these things is generally terrible. There is no way it should ever be input to any ML model.
‘Our examination of the excluded data suggests that documents associated with Black and Hispanic authors and documents mentioning sexual orientations are significantly more likely to be excluded by C4.EN’s blocklist filtering, and that many excluded documents contained non-offensive or non-sexual content (e.g., legislative discussions of same-sex marriage, scientific and medical content).’
Very disheartening, but I'm not surprised.
Very disheartening, but I'm not surprised.
"I don't have a problem with those people, but my grandma/the Chinese market/our stakeholders might, so it is a risk I cannot take."
Same old story, history repeats. We are just cementing the abstract into code.
Same old story, history repeats. We are just cementing the abstract into code.
d1a2n(1)
I'm not sure I buy this interpretation. As the article says, these are "identity mentions", not definitive classifications of the author's identity. Don't these results just indicate that some identity labels are more likely to be used in offensive ways?
‘Some filters are relatively straightforward, such as removing Lorem ipsum placeholder text. However, we find that another filter which removes documents that contain a token from a banned word list, disproportionately removes documents in dialects of English associated with minority identities (e.g., text in African American English, text discussing LGBTQ+ identities).’
A word so lovely that we really really don't want to risk putting it in our report...
A word so lovely that we really really don't want to risk putting it in our report...
Yeah, I have similar concerns about the different dialects thing. It's true that "the meaning of seemingly 'bad' words heavily depends on the social context", but most realistic systems are gonna be deployed in multiple social contexts and ideally shouldn't be offensive in any of them.
It's not that there are different contexts, it's that with particular identity groups, the "rules" for that context vary so much over time (and context) as well.
For most of the 90's saying black was "wrong" in favor of African American. Now we've gone full circle, and the same people who made a sour-face at black 25 years ago are capitalizing it.
That's probably the most benign, easy case. Of course ML can't keep up with loaded-terms and slurs; most people can hardly keep track of it all.
For most of the 90's saying black was "wrong" in favor of African American. Now we've gone full circle, and the same people who made a sour-face at black 25 years ago are capitalizing it.
That's probably the most benign, easy case. Of course ML can't keep up with loaded-terms and slurs; most people can hardly keep track of it all.
If they didn't filter those words out, then presumably the headline would be about 'perpetuating' instead of about 'filtering'.
> A word so lovely that we really really don't want to risk putting it in our report...
I can't tell if that is sarcasm or not. Is the 'lovely word' disclosed somewhere else?
I can't tell if that is sarcasm or not. Is the 'lovely word' disclosed somewhere else?
Or... it could mean that the dominant culture defaults to describing identity labels as offensive. The labels themselves would only be offensive if someone deems them so. Remember when "gay" was a slur, but now it's an acceptable term for homosexual? That is, unless you're an edgy 13yo boy on the internet and think calling something 'gay' is funny.
Again, this data doesn't show that the labels are offensive, just that they're more likely to be used in text which was classified as offensive. If there's lots of 13 year old boys on the Internet using "gay" as an insult, filtering out those insults from the corpus (and it's gottta be correct to filter them) could fully explain the high PMI for the word "gay".
> used in text which was classified as offensive
Yes, and determining if a text is offensive is a very socially and culturally determined. Who determines the offensiveness of the context, and how? To shift gears a bit, how many rap song lyrics were filtered out from the corpus, for example? Does use of the n-word make an entire text de facto offensive? What about references to the common name of the moth Lymantria dispar dispar? What about the football team from Washington, D.C.?
A lot of this just feels like a repeat of the Net Nanny internet filter days, when keywords were idiotically filtered out.
Yes, and determining if a text is offensive is a very socially and culturally determined. Who determines the offensiveness of the context, and how? To shift gears a bit, how many rap song lyrics were filtered out from the corpus, for example? Does use of the n-word make an entire text de facto offensive? What about references to the common name of the moth Lymantria dispar dispar? What about the football team from Washington, D.C.?
A lot of this just feels like a repeat of the Net Nanny internet filter days, when keywords were idiotically filtered out.
These are all important questions, but I don't think the article meaningfully engages with any of them beyond noting their existence. A lot of people are in a position where they have to judge whether a text is offensive; if your system autofills racial slurs, and you try to explain that it's not a problem because in some contexts the slurs have been reclaimed, you're gonna get in serious trouble.
> That is, unless you're an edgy 13yo boy on the internet and think calling something 'gay' is funny.
You've gotta be living in a bubble to think this is something only 13 year-old boys say on the internet in the privacy of their bedroom.
You've gotta be living in a bubble to think this is something only 13 year-old boys say on the internet in the privacy of their bedroom.
Why would someone want to curate their data if they want to build a good model?
Are they worried that a scientist is going to be exposed to a bad word?
Are they imposing some cultural brand of morality upon the AI?
Are they worried that a scientist is going to be exposed to a bad word?
Are they imposing some cultural brand of morality upon the AI?
> has been extensively ‘filtered’ to remove black and Hispanic authors, as well as material related to gay and lesbian identities
But after it is explained that this is an unintended second order effect from trying to remove offensive content from the corpus. I'm not trying to justify the outcome (it's an issue regardless of intent) I just didn't think there was any need to strongly imply it was intended.