0

I am trying to perform a classification procedure where my training data looks like this:

(state, (feature_1, feature_2, feature_3, ..., feature_n))

Thus, given a set of features, I need to predict what state/label/class those features most likely correspond to.

I have the nice CRFSuite model set up for making CRFs very fast, but is a CRF really ideal for this kind of learning? I used CRF in the past for sequences of states, that is the label of the $nth$ state may also depend on the label / features of the previous $n-1$ states. For example, here is a training sequence I used for trying to predict a child's phonetic output given the adult IPA transcription:

e   Adult=e __BOS__
i   Adult=-
d   Adult=d
r   Adult=-
i   Adult=i
ə   Adult=-
n   Adult=- __EOS__

A CRF makes sense for this data because phonology/phonetics is very regular--what sound is chosen highly affects future sound choices, e.g. a vowel will probably be followed by a consonant and not another vowel.

I (believe) understand that a CRF is actually just a sequential form of a Maxent model. So if all my training sequences are always length $1$, will I basically just have a Maxent model called a CRF?

This question CRF for named entity recognition addressed using a CRF for named entity recognition, but I am guessing it uses sequences of states?

Community
  • 1
  • 1
user3898238
  • 957
  • 3
  • 11
  • 25

1 Answers1

0

Well for this purpose , i think the most simple CRF that is Logistic regression a discriminative classifier can do the job. Your requirement is unlike named entity because there is not a chain of states or observation. It is simple one state and limited features. I think logistic regression and after that bias/variance trade off is the way to go.

Dhruv Premi
  • 38
  • 6