(Obviously) I don't fully understand your architechure, but isn't there a (relatively) simple check that it shouldn't test for a word that it has never taught?
(And on the bright side, I'm sure I'll remember unserem now!)
Well, yes, but the system filters definitions based on part of speech. For example, you don't want to show "to guide or conduct" (lead, verb) for a sentence talking about the metal lead (lead, noun).
In this case, unserem was assigned the part of speech determiner by the NLP pipeline (shoutout to https://stanfordnlp.github.io/stanza/, it's brilliant), but the dictionary only has an entry for unserem as a pronoun.
Perhaps a better design is to always show all entries, and just put the entries whose POS match the word in the sentence (if any) above the others.
In this case, unserem was assigned the part of speech determiner by the NLP pipeline (shoutout to https://stanfordnlp.github.io/stanza/, it's brilliant), but the dictionary only has an entry for unserem as a pronoun.
Perhaps a better design is to always show all entries, and just put the entries whose POS match the word in the sentence (if any) above the others.