You can provide feedback on the answer card. I've had a few answer cards fixed after providing feedback. Notably "https port" used to yield the http port.
Edit: I think you've actually managed to find a really terrific example of a case where modern ML systems are going to have trouble, because of how that Wikipedia page is worded in relation to the query.
* There isn't a single phrase that answers the query directly, so the ML model would need to make very good use of context (attention), both within multiclause sentences and between paragraphs.
* There are many different numbers on the page, so the model has to determine the right one. It can't just get lucky by guessing here.
* A wrong number (1.7MB) has close proximity to literal keywords in the query (floppy, disk). (The right number [880KB] does too.)
* The model has to understand and properly make use of "most" vs "unusual" in its decision.
Edit: I think you've actually managed to find a really terrific example of a case where modern ML systems are going to have trouble, because of how that Wikipedia page is worded in relation to the query.
* There isn't a single phrase that answers the query directly, so the ML model would need to make very good use of context (attention), both within multiclause sentences and between paragraphs.
* There are many different numbers on the page, so the model has to determine the right one. It can't just get lucky by guessing here.
* A wrong number (1.7MB) has close proximity to literal keywords in the query (floppy, disk). (The right number [880KB] does too.)
* The model has to understand and properly make use of "most" vs "unusual" in its decision.