Closed koliaok closed 4 years ago
Hi @koliaok, sorry for the confusion. In the paper, we try to describe our models in a concise way as there is the page limit. The representation encoded_context
is expanded by simply stacking by |Zdsxfert| times as we want the resulting stack to be compatible with the attention matrix pointer_attn
. The stacked vector is used to obtain the weighted sum of dialogue context embedding.
https://github.com/henryhungle/NADST/blob/afdc1d1f7ecb855b03933e441c0b2fcefbc28feb/model/generators.py#L57
The weighted sum is then concatenated to the other 2 representations to obtain a 3xd dimensional representation.
https://github.com/henryhungle/NADST/blob/afdc1d1f7ecb855b03933e441c0b2fcefbc28feb/model/generators.py#L58
This code is not a major part of our idea anyway. It is mainly to address the OOV problem in DST generation using a traditional pointer network.
Ok Thank you for response
Have nicely Henry
On May 18, 2020, at 18:39, (Henry) Hung Le notifications@github.com wrote:
Hi @koliaok https://github.com/koliaok, sorry for the confusion. In the paper, we try to describe our models in a concise way as there is the page limit. The representation encoded_context is expanded by simply stacking by |Zdsxfert| times as we want the resulting stack to be compatible with the attention matrix pointer_attn. The stacked vector is used to obtain the weighted sum of dialogue context embedding. https://github.com/henryhungle/NADST/blob/afdc1d1f7ecb855b03933e441c0b2fcefbc28feb/model/generators.py#L57 https://github.com/henryhungle/NADST/blob/afdc1d1f7ecb855b03933e441c0b2fcefbc28feb/model/generators.py#L57 The weighted sum is then concatenated to the other 2 representations to obtain a 3xd dimensional representation. https://github.com/henryhungle/NADST/blob/afdc1d1f7ecb855b03933e441c0b2fcefbc28feb/model/generators.py#L58 https://github.com/henryhungle/NADST/blob/afdc1d1f7ecb855b03933e441c0b2fcefbc28feb/model/generators.py#L58 This code is not a major part of our idea anyway. It is mainly to address the OOV problem in DST generation using a traditional pointer network.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/henryhungle/NADST/issues/4#issuecomment-630067762, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADG6G3AW6G3VKFAG3PSOGZLRSD655ANCNFSM4NAW7OTQ.
I have some feedback after read your NADST code review
first, your NADST paper page 6, and 3.3 State Decoder
you explained like this "where Wgen ∈ R3d×1 and Zexp is the expanded vector of Z to match dimensions of Zds×fert."
I think more explained Z_exp, don't enough explain it in paper
your real code :
context_vec = (encoded_context.unsqueeze(1).expand_as(expanded_pointer_attn) * expanded_pointer_attn).sum(2)
I think, need to more explain above code from "expanded_pointer_attn variable" for multiply expanded encoding context vector
How about that?