How to create features with duplicate keys ?

[binhnq94]

I see in (crfsuite document)[http://www.chokkan.org/software/crfsuite/manual.html] that key of feature can be duplicate:

B-NP    w[1..4]=a:2 w[1..4]=man w[1..4]=eats
B-NP    w[1..4]=a w[1..4]=a w[1..4]=man w[1..4]=eats
B-NP    w[1..4]=a:2.0 w[1..4]=man:1.0 w[1..4]=eats:1.0

How to create features with duplicate keys if i using sklearn-crfsuite ?

[kmike]

There are no duplicate features here: for example, in a first row w[1..4]=a and w[1..4]=man are feature names; w[1..4]=a has a value = 2 (:2), and w[1..4]=man has a default value = 1. = sign is just a convention to build readable feature names.

[binhnq94]

In crfsuite document:
-> This is the BNF notation representing the data format.

<line>           ::= <item> | <eos>
<item>           ::= <label> ('\t' <attribute>)+ <br>
<eos>            ::= <br>
<label>          ::= <string>
<attribute>      ::= <name> | <name> ':' <scaling>
<name>           ::= (<letter> | "\:" | "\\")+
<scaling>        ::= <numeric>
<br>             ::= '\n'

It mean 2 is scaling or weight.

A sample data for CRFsuite:

[kmike]

Yes, 2 is weight, this is what I've written as well.. Could you please provide an example of a duplicate key? It seems I don't understand your question.

[binhnq94]

Sample sequence: a man eats some food
If we config feature of some sequence labeling problems, we have some type of features:

• type 1: w[-3], w[-2], w[-1], w[0], w[1], w[2], w[3] is a window around the current word. Features for man:

{
   w[-1]: a,
   w[0]: man, 
   w[1]: eats, 
   w[2]: food
}

• type 2: w[-3]|w[-2]|w[-1], w[1]|w[2]|w[3] are junctive features:

{
   w[-3]|w[-2]|w[-1]: a,
   w[1]|w[2]|w[3]: eats|some|foods, 
}

• type 3: w[-3...-1], w[1...3] are disjunctive features:

{
   w[-3...-1]: a,
   w[1...3]: eats,
   w[1...3]: some,
   w[1...3]: foods
}

And your package cant support type 3.

[kmike]

See data format examples at http://python-crfsuite.readthedocs.io/en/latest/pycrfsuite.html#pycrfsuite.ItemSequence - the format you're using is just a shortcut:

{“string_key”: “string_value”, ...} dict; that’s the same as {“string_key=string_value”: 1.0, ...}

All features have float values, both in crfsuite and python-crfsuite; {"w[-1]": "a"} is just a shortcut for {"w[-1]=a": 1.0}, and in crfsuite example = is just a convention. 1.0 is an assumed value; in crfsuite you override it by putting after :, in python-crfsuite you can override it by using another feature format ({"string_key": float_value}).

So if you want such disjunctive features, pass {"w[1...3]=eats": 1.0} instead of {"w[1...3]": "eats"}. If you don't need weights, you can also use a list feature format.

[binhnq94]

Ok. I think that if i pass {"w[1...3]=eats": 1.0}, it be like i pass w[1...3]=eats=1.0:1.0.

If you don't need weights, you can also use a list feature format.

-> How i use a list feature format with sklearn-crfsuite ?

[kmike]

{"w[1...3]=eats": 1.0} is exactly the same as {"w[1...3]": "eats"}.

If all of your features have value 1.0, and you build feature names manually, you can pass a list: ["w[1...3]=eats", "w[1...3]=some", ...] instead of a dict {"w[1...3]": "eats", "w[1...3]": "some", ...}, they are exactly the same.

Please check data formats here: http://python-crfsuite.readthedocs.io/en/latest/pycrfsuite.html#pycrfsuite.ItemSequence

[binhnq94]

Understood. Thank you.