Question about encoder the sentences or entities when using Genre.

[Yuanhy1997]

To build my own trie, I first encode all the entities in my KB. When encoding, I used the following code like the code in the rebuilding trie issue:

Trie([2]+model.encode(entity.lower())[1:])

In my opinion, 2 stands for the 'eos' token, while in another issue, you said 2 stands for 'bos' token which is confusing. And another problem is, I do not understand why there is [1:], because in my application, the output of model.encode() is not prepended with any 'bos' or 'eos' token, its beginning is the bpe token of my input entity and the final token is '2'. Here is an example of the model.encode() output:

[ 4469, 29048, 19961, 29382, 13736, 36, 7779, 10337, 43, 2]

Can you please explain that?

[nicola-decao]

Can you please show me an example where this fails?

[Yuanhy1997]

This does not fail, your code works fine, but I am confusing that how exactly your code works. In my opinion, when training the decoder inputs and labels are:

inputs: eos, bos, y1, y2,... labels: bos, y1, y2,..., eos

while in GENRE, I think it becomes

inputs: eos, y1, y2,... labels: y1, y2,..., eos

I dont know if I'm right. If so, can you please explain why this difference happens.

[nicola-decao]

I do not understand. If you are not sure if you are right what do you want me to answer? We used a pre-trained BART model, the encode and decode are the same as the original BART model.