Create an "inverse mask"

Is there any way I could create essentially an inverse mask, where I give the model words and it spits out the probability that those words would fit the sentence.

Person1: I hate you.
Person2: I understand that, how can we fix it.
Provide the probability that the following adjectives would apply to the response of Person2:
Angry: mask
Cooperative: mask
Hateful: mask
Sarcastic: mask

It would then hopefully output the probabilities.

Edit: I currently have something doing this using an llama based model, but the output so far has been unpredictable and unreliable. It essentially feeds the above as a json form and some instructions and asks it to complete the JSON.