So at work recently they unveiled an ml app that takes long winded descriptions of machine parts ( usually due to the 1000 different ways someone can describe a specific widget) and distilled it down to its core description. All good.
My question is, is it possible to have some ml code that can look at whats been used so far for training, and based on some human guidance, generate plausible missing permutations of words to fill the gap to generate a 90-95% or higher accuracy by using the new synthetic training data? And then you could stick it in a self tuning loop to keep things tight.
Its just a thought at the moment, sometimes i dont know all the details, but can see the problem and just need the right stuff to fix the problem.
Thoughts welcome.