In standard methodology for natural language processing, entities in text are
typically embedded in dense vector spaces with pre-trained models. The
embeddings produced this way are effective when fed into downstream models, but
they require end-task fine-tuning and are fundamentally d