My opinion: don't
Don't attempt to adapt/translate/localize them, at least at the beginning. There isn't even universal agreement for the meaning of emoji within a single community of practice.
I chat with people on different mobile platforms and am constantly excusing myself just because their older Android draws the emoji differently than mine, so the subtle meaning conveyed in the way the iris points in an eyeball means the difference between complete understanding me unintentional upsetting.
Toy example: the "imp" emoji displayed by Apple (left) and the custom implementation by some Samsung Android devices (right):
One is "clearly angry" and one is "clearly playful", to me, because of my culture and personal experience.
Furthermore, while the meaning of emoji is influenced by culture, but is also highly dependent on the relationship between the two people chatting.
So it will be incredibly difficult to figure out which emojis means what to whom, and is probably not worth the risk. Just be aware of what meanings some emoji can carry in different cultures and be sensitive to that.
Perhaps worth mentioning
Perhaps a useful idea to mitigate against the problems above could be the GitHub Reactions approach: limit emoji in system-generated responses to a subset of relatively unambiguous emoji (however you define that for your user base).
You're making a chatbot, which implies that you're doing machine learning. Use that as a data source and improve your use of emoji over time, and find trends on how different classes of user react to / appreciate / use emoji in context.
That is also something you can A/B test yourself.