Comment by evrydayhustling

We used to use BERT-based embeddings + semantic distance for classification / decision problems in new domains. There was a lot of interest at the time in these kinds of pre-generative but portable models -- Meta's Prophet was another example that came up a lot.