Q. About leading_symbolic

[isbada]

Hi, @XuezheMax
I have noticed that in the loss function of NER.py , there is a parameter named 'leading_symbolic'(value=1),
In the source code of loss funtion , you use this parameter as following:

 _, preds = torch.max(output[:, :, leading_symbolic:], dim=2)
preds += leading_symbolic

Could you explain Why you change the shape of output of the network in this way?
looking forward to your reply~

[XuezheMax]

Hi, The argument 'leading_symbolic' is only used for decoding. For CRF, we have some symbolic labels, such as "START" of sequence, "END" of sequence and "PADDING" label. In decoding, we would not like our model to output these symbolic ones. The way we avoid to output the symbolic labels is to put them in the first positions in the index and filter them out during decoding. Is it clear?

…

On Sat, Dec 30, 2017 at 5:34 AM, isbada ***@***.***> wrote: Hi, @XuezheMax <https://github.com/xuezhemax> I have noticed that in the loss function of NER.py , there is a parameter named 'leading_symbolic'(value=1), In the source code of loss funtion , you use this parameter as following: _, preds = torch.max(output[:, :, leading_symbolic:], dim=2) preds += leading_symbolic Could you explain Why you change the shape of output of the network in this way? looking forward to your reply~ — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#7>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADUtliAAWmMleSDTpX8C7bWjbLHZgNiWks5tFhGhgaJpZM4RPkVR> .

-- ------------------ Best regards, Ma，Xuezhe Language Technologies Institute, School of Computer Science, Carnegie Mellon University Tel: +1 206-512-5977

[isbada]

@XuezheMax Thanks for your reply.
I have understood it now