We introduce a new probabilistic method and task for quantifying the privacy risk of a document with personal attributes using large language models.

We integrate discrete diffusion models with neurosymbolic predictors for scalable and calibrated learning and reasoning