Unless they’re highly motivated to do so, most people won’t bother even with hash tagging, in my experience.

Yes, there’s a huge rabbit hole of such optimizations that can be done. Your stemmer step might take care of making those things identical, or you could use some sort of string distance comparison like Levenshtein distance instead of strict equality when deciding whether the term matches some other term.

They certainly are best placed to do so. The problem is rather that most users don’t care.

They view the machine as a black box into which they dump random information and magically get useful results out. Every extra step you introduce into that process cuts compliance by half, unless they’re highly motivated (e.g. their boss orders them to do it.)

If you can’t force them to do the tagging out-of-band like that, it’s probably a non-starter unless the process you are building delivers 100x value over the next-best solution / whatever they’re doing now. 2x or 10x likely isn’t enough.

Many shops fake it with humans on staff paid to do the tagging, so customers don’t have to.