All candidate pair scores?

[sunilitggu]

Hi I was trying to understand the code. I found that you are feeding an abstract separately for each candidate pair of the abstract in the model. However, in the paper, it is written once for all candidate pairs.

Am I missing something?

Thanks

[patverga]

Currently the code is doing some redundant computation, re-encoding the abstract for each entity pair. However, it is computing the full pairwise score tensor without any entity pair specific features which you can access here https://github.com/patverga/bran/blob/master/src/models/transformer.py#L468 . You can compute that tensor once and then aggregate scores for each of the entity pairs. This can be done efficiently using a gather/scatter but is not currently implemented.

[leebird]

Hi,

In the above link, a few lines above, it seems e1_mask and e2_mask are not used in any way. How does the model know the locations of the tokens of the current entity pairs?

Thanks!

[leebird]

Ah, never mind, just saw the ep_dist list.

[nefujiangping]

Ah, never mind, just saw the ep_dist list.

Did you know what exactly ep_dist_batch represents? I saw that ep_dist_batch is added to pairwise_scores in code (here) directly.

Thanks!