Let's not go the way of Meta's Galactica

[dkoslicki]

If you've seen the new, this language model was shut down rather quickly due to its ridiculous answers. Trying out one similar test query (with vaccines and autism), on our system, I got: https://arax.ncats.io/?r=99378 .

This came from RTX-KG2 (don't know if you want to cross-post this issue @saramsey @amykglen @acevedol).

The underlying SemMedDB snippets look reasonable, but the predicates sure are unfortunate.

[saramsey]

@dkoslicki thank you for bringing this up. Yeah, that's awkward:

But perhaps a useful use-case to figure out how to improve ARAX?

[dkoslicki]

Precisely: an example to sanity check our ranking and filtering

[edeutsch]

I wonder if this is a case where we need some special logic in the KG2 build process to correct some known problems. And this is also a case where we want to use the "NOT" modifier to statements. So that that statement is "vaccines NOT biolink:causes autism spectrum disorder".

[dkoslicki]

Just checked and while this particular query now returns zero results, a modification (different CURIEs) does: https://arax.test.transltr.io/?r=247697. Technical debt, or....?

[dkoslicki]

closing in favor of #2303