Well, this seems dangerously inept.
When I spot checked a few of these, I found instances that were not merely marginal, but overtly and obviously false connections. "EST" in the word "DESTROY" in its context is NOT a reference to white supremecism, is it? Maybe the subreddit "BlackPeopleTwitter" pushed that classification along, but the next couple, where "single" is linked with homophobia (in the same subreddit) is simply absurd - its just used in completely different contexts from anything to do with that subject at all! This could underpin automated tyranny without a lot of care; or does the context and purpose of this dataset obviate that? Be aware, that's all, and be careful!
Hi @JWBickel , this is the "potential" dog whistle dataset - which means it was consolidated through simple keyword search against the Allen AI Dog Whistle Glossary - as is explained in the dataset card:
"This dataset contains over 6 million excepts from Reddit comments that contain a dogwhistle term, created via keyword search on terms in the Allen AI Dogwhistle Glossary. Therefore, many of the excerpts are likely using these terms in a non-coded sense. This dataset was used as input to our approach for word-sense disambiguation of dog whistles from standard speech using Large Language Models (LLMs). Please see the paper linked below for more details."
This was the input data to our word-sense disambiguation pipeline, which then filtered out the coded from non-coded cases. So, yes, many of these example are not coded dog whistles. We release this data because we hope other teams with less resource constraints could expand the disambiguated dataset SALT-NLP/silent_signals, or could use the input to test out their own solutions. Hope this helps!
I'm glad I misunderstood the card. Let's stay vigilant about this kind of thing. NOTE: The word "vigilant" in the preceding text is not an indication you should have me arrested. :^D Sigh. It's a sad, weird world, isn't it?