Confusion about the output (role type)

[thunlp]

Confusion about the output (role type)

[wzq016]

entity information is included in dynamic maxpooling stage and the output is for candidate entity.

In test stage, we calculate every roles and select role with highest score, so we didn't use label in test stage.

[ZacBi]

  Thanks for your response, but it doesn't clear my confusion. For your paper you said:

  In Logic Union Module you calculate a weighted average sum over all hidden states in the sentence to get a role-oriented embedding(context vector for each role).
  So, I understand that your method/framework can work if and only if an input sentence contained one candidate entity, otherwise it doesn't make sense. Perhaps, use dynamic max-pooling for each candidate entity to construct its own instance embedding x.

[wzq016]

Hidden embeddings h is related with selected trigger and entity, since we feed position embedding to encoder together with word embedding. So role-oriented embedding is related with candidate entity.

[ZacBi]

Fine... it looks like we don't reach the same point. I have to view your implementation before more discussions. Thank you again.