I'd assume you could train a cross encoder network like

https://sbert.net/docs/cross_encoder/usage/usage.html

to "predict" if two entities are the same?