Featurizers

[Omid]

Hi again, sorry if I ask these many questions, I want to use my features. I am following the example of POSTagger, this is the default wordFeaturizer in CRF Model:

def myFeaturizer(counts: Counter2[String, String, Double]) = {
    val dsl = new WordFeaturizer.DSL[String](counts)
    import dsl._
    (
      unigrams(word + clss, 1)
         +bigrams(clss, 2)
          +bigrams(tagDict, 2)
        + suffixes()
       + prefixes()
       +  props
    )
  }

If I want to add my features, which is a list of String features for a word (e.g. dependency path or any string). what exactly should I add to the function? Here is my Map of String and features:

val myFeatures: Map[String,List[String]] //for a word,list of string feature

[David Hall]

Questions are good. They help me figure out where documentation is lacking, if nothing else.

You need to make a WordFeaturizer; look at IdentityWordFeaturizer (which isn't quite identity because of some Unking--and it's kind of ugly so I should rewrite it) which is a reasonable example. Probably an even better example is the BrownClusterFeaturizer.

val myFeaturizer = ???

You can add featurizers together with + and (if the dsl is imported like the above) you can get cartesian products/cross products with *

-- David

[Omid]

I looked at the codes, as I can see a Feature get an string value like:

case class BrownClusterFeature(f: String) extends Feature

And a Featurizer gets sometimes: "lengths: Array[Int]" or "wordCounts: Counter[W, Double]" or Counter2 ... what exactly is the input of a featurizer ? and what does it do? would you explain it more and give and example like a Featurizer that get the word pair with POS and make faeture of it.

Thanks

[David Hall]

A WordFeaturizer is just this interface:

trait WordFeaturizer[W] {

  def anchor(words: IndexedSeq[W]):WordFeatureAnchoring[W]

}

and WordFeatureAnchoring is just this interface:

trait WordFeatureAnchoring[W] {

  def words: IndexedSeq[W]

  def featuresForWord(pos: Int):Array[Feature]

}

So you want something like:

case class MyFeature(f: String) extends Feature

class MyFeaturizer(features: Map[String, Seq[String]]) extends WordFeaturizer[String] {

   def anchor(w: IndexedSeq[W]) = new WordFeatureAnchoring[String] {

      def words = w

      

      def featuresForWord(pos: Int) = if(pos < 0 || pos >= length) Array.empty else features(words(pos)).map(MyFeature).toArray

   }

  

}

[Omid]

Thanks,

length in pos >= length is the length of features?

[David Hall]

length of the sentence

[Omid]

Then, when we pass a function with these:

  unigrams(word + clss, 1)
         +bigrams(clss, 2)
          +bigrams(tagDict, 2)

what are "word" and "clss" and "tagDict" ?

[David Hall]

word is word identity. clss is WordClassFeaturizer. tagDict is the most common tag for each word.  Look at WordFeaturizer.DSL

[Omid]

Thanks a lot :)

[David Hall]

you're most welcome