Yes, this really needed its own thread. In researching what vocab we have, and what roles do and no not have vocab, I realized that the vocab system is completely broken and odd.
Notice how some of those cops are vocab and some are not? It turns out that the game does not just mark the vocab that we say. It is more situation based than that. Someone better at explaining the specifics will explain below.
But, long story short, we should probably use a different system.
ok it's not letting me edit this. I did a writeup and it's gone now so im gonna keep this shorter im tired
EM uses NLP to detect nouns in a sentence to check against predetermined role names and aliases. If it matches, it highlights the word.
The problem lies with NLP having a lot of false negatives with anything other than perfect grammar. If it recognizes the word as a noun properly, all is well. If it does not recognize a word that should be highlighted as a keyword as a noun, it will not even be checked against the dictionary of role names, aliases, and descriptions.
In short, writing with improper grammar makes it more likely that there will be a false negative and that a keyword that by all means should be highlighted is not, because the keyword does not pass the NLP check.
Another problem is that NLP recognizes some adjective-noun phrases like "confused cop" as one noun, so the phrase "confused cop" is checked against role names and aliases, and nothing is found as only "confused" exists as a role name, not confused cop.
I imagine the decision to use NLP was made to reduce false positives, or maybe just to try something new, but I think simple keyword detection (like what exists for terminology keywords like 'wifom') would work better. I believe that when it comes to giving explanatory information to the players, false positives are better than false negatives.
