SeedBERT: Recovering Annotator Rating Distributions from an Aggregated Label