Task-aware representation of sentences (TARS), is a simple and effective method for few-shot and even zero-shot learning for text classification. However, it was extended to perform Zero-Shot NERC.
Basically, TARS tries to convert the problem to a binary classification problem, predicting if a given text belongs to a specific class.
TARS doesn't need the descriptions of the entities, so if you can't provide the descriptions of the entities maybe this is the approach you're looking for.
The TARS linker will use the entities specified in the
TARS end2end Linker
Default entities to use in case no custom ones are set One of: - 'conll-short' - 'ontonotes-long' - 'ontonotes-short' - 'wnut_17-long' - 'wnut_17-short'
As TARS use only the labels, take just the name of the entities and not the description
Load TARS model if its not initialized