clj-crfsuite

Clojure interface to CRFSuite - a fast linear-chain CRF library.

This is a thin wrapper around jcrfsuite so we can use clojure data structures.

Usage

Lein artifact

A linear chain CRF produces one symbol per output feature. For a simple task like POS-tagging:

NP --- conj --- NP --- VP --- dobj
|       |       |      |       |
|       |       |      |       |
John   and     mary   like    apples

The CRF is supplied english words (or some features corresponding to english words). The output is the POS tags.

Training:

Args:

xs : A list of sequences. A sequence is a clojure sequence of feature-vectors or feature-maps (a map from feature -> value).
ys : Tags corresponding to sequences in xs.
model-file : Location to dump model to.

user=> (train [
               [{:feat1 2.0, :feat2 3.0}
                {:feat1 3.0, :feat2 4.0}] ; the first sequence of features
			   [{:feat1 2.0, :feat2 3.0}
			    {:feat1 4.0, :feat2 5.0}] ; the second sequence of features
			   ]
		      [
			   ["y1", "y2"], ; tag-sequence
			   ["y1", "y3"]  ; tag-sequence
		       ]
		      "trainmodel.crfsuite") ; destination
.
.
.
L-BFGS resulted in convergence
Total seconds required for training: 0.000

Storing the model
user=>

You can also supply a vector of features (or an ndarray like core.matrix matrices):

      (train [
              [
               [1.0, 2.0],
               [3.0, 4.0]
               ]
              [
               [1.0, 2.0],
               [3.0, 4.0]
               ]
              ]
             
             [["y3", "y4"]
              ["y3", "y4"]]

             "trainmodel.crfsuite")

Tagging:

Pass a sequence of features to be tagged with a model stored in traimodel.crfsuite. Returns a sequence of tag, probability pairs.

user=> (tag [{:feat1 2.0, :feat2 3.0} {:feat1 3.0, :feat2 4.0}] "trainmodel.crfsuite")
[
  [{:tag "y1"} {:probability 0.43526113123400284}] ;; instances of type clj-crfsuite.core.Tag
  [{:tag "y3"} {:probability 0.36107961034023944}]
])

License