format of entries for entity linking training about genre HOT 2 CLOSED

The training of the EL task is done via a sequence-to-sequence objective. The input text is plain text and the output is a markup language like this example:

Input: In 1921, Einstein received a Nobel Prize.
Output: In 1921, { Einstein } [ Albert Einstein ] received a { Nobel Prize } [ Nobel Prize in Physics ] .

from genre.

Hello

Thanks for your response but my question was more about the construction of a dataset to train GENRE.

In your ED model, there is only 1 mention by entry, which is unusual for EL task (where the entry is 1 plain document) and ED task (because the context between the differents mentions of a document is lost).
I don't know if this choice is a constraint for fairseq or an optimisation of the model's objective.

I would like to know if for the EL task, I have to process 1 mention by entry or if I have to process 1 document by entry.

from genre.