features.add_feature_edge( lambda y_, y: 0 )

[goog]

Hey, shuyo! the 351 line in crf.py :features.add_feature_edge( lambda y_, y: 0 ) , is there a problem? THX!!!

[shuyo]

I cannot remember why I wrote the line ... I guess it is because I did for a sample skeleton, maybe ... :D

[goog]

BTW, which algorithm did you apply to calculate the gradient_likelihood function?

[shuyo]

No. I reckon it needs no specific algorithm to calculate the gradient of likelihood on CRF because it has a closed form.

[goog]

Hey! could you say it more detailed ,and I'm not similar with it. Thanks!

[goog]

I fixed it like this:

    pre_y =''
    for j,label in enumerate(features.labels):  #label for a line
        for token in tokens:
            features.add_feature( lambda x, y, l=label, t=token: 1 if y==l and x[0]==t else 0 )
            # def self.features.append(f) .
        for info in infos:
            features.add_feature( lambda x, y, l=label, i=info: 1 if y==l and x[1]==i else 0 )
        if j:
            features.add_feature_edge( lambda y_, y,p=pre_y,l=label: 1 if y_==p and y==l else 0 )
        pre_y=label 

[goog]

Hey, pal! there is a little problem in function _calc_fmlist which didnt take positions into consideration

[shuyo]

Sutton's CRF tutorial introduced the gradient of the likelihood.
http://people.cs.umass.edu/~mccallum/papers/crf-tutorial.pdf

[goog]

Thanks! BTW, which difference(違い) is used by The gradient is computation?

[shuyo]

Like what problem?

[goog]

is the gradient computed using differences ?

[shuyo]

What are "differences"?
If you means Numerical Differentiation, then no.

[goog]

Hi,shuyo! I have some questions
1,In the partial derivatives of (1.21) in Sutton's CRF tutorial, why there is a probability? It's difficult for me to to understand.
2, how to store models or models mergence for several different train datasets
3, when 'fmlist doesn't depend on x_i' , is it like HMM and which scenario can this case be applied in.
4, the subscripts of logbetas in gradient_likelihood seems inconsistent with my book
I wanna fully understand your code and if possible add more detailed comments on the code for people who like me (a newbie for CRF) please! Thanks in advance!!!