cc.mallet.fst.semi_supervised
Class EntropyLattice

java.lang.Object
  cc.mallet.fst.semi_supervised.EntropyLattice

public class EntropyLattice
extends java.lang.Object

Runs subsequence constrained forward-backward to compute the entropy of label sequences.

Reference: Gideon Mann, Andrew McCallum "Efficient Computation of Entropy Gradient for Semi-Supervised Conditional Random Fields" HLT/NAACL 2007

Author:

Gideon Mann, Gaurav Chandalia, Gregory Druck

Nested Class Summary

class

EntropyLattice.LatticeNode Contains alpha, beta values at a particular input position and state pair.

Field Summary

protected double	entropy
protected int	inputLength
protected int	latticeLength
protected EntropyLattice.LatticeNode[][]	nodes
protected int	numStates
protected Transducer	transducer

Constructor Summary

EntropyLattice(FeatureVectorSequence fvs, double[][] gammas, double[][][] xis, Transducer transducer, Transducer.Incrementor incrementor, double scalingFactor)
Runs constrained forward-backward.

Method Summary

double	backwardLattice(double[][] gammas, double[][][] xis) Computes the backward entropies (H^beta).
double	forwardLattice(double[][] gammas, double[][][] xis) Computes the forward entropies (H^alpha).
double	getEntropy()
EntropyLattice.LatticeNode	getLatticeNode(int ip, int si)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

latticeLength

protected int latticeLength

inputLength

protected int inputLength

transducer

protected Transducer transducer

numStates

protected int numStates

nodes

protected EntropyLattice.LatticeNode[][] nodes

entropy

protected double entropy

Constructor Detail

EntropyLattice

public EntropyLattice(FeatureVectorSequence fvs,
                      double[][] gammas,
                      double[][][] xis,
                      Transducer transducer,
                      Transducer.Incrementor incrementor,
                      double scalingFactor)

Runs constrained forward-backward.

If incrementor is null then do not update expectations due to these computations.

The contribution of entropy to the expectations is multiplies by the scaling factor.

Method Detail

getEntropy

public double getEntropy()

forwardLattice

public double forwardLattice(double[][] gammas,
                             double[][][] xis)

Computes the forward entropies (H^alpha).

backwardLattice

public double backwardLattice(double[][] gammas,
                              double[][][] xis)

Computes the backward entropies (H^beta).

getLatticeNode

public EntropyLattice.LatticeNode getLatticeNode(int ip,
                                                 int si)