/* Copyright (C) 2010 Univ. of Massachusetts Amherst, Computer Science Dept.
This file is part of "MALLET" (MAchine Learning for LanguagE Toolkit).
http://www.cs.umass.edu/~mccallum/mallet
This software is provided under the terms of the Common Public License,
version 1.0, as published by http://www.opensource.org. For further
information, see the file `LICENSE' included with this distribution. */
package cc.mallet.fst.semi_supervised.constraints;
import java.util.ArrayList;
import java.util.BitSet;
import cc.mallet.fst.SumLattice;
import cc.mallet.fst.semi_supervised.StateLabelMap;
import cc.mallet.types.FeatureVector;
import cc.mallet.types.FeatureVectorSequence;
import cc.mallet.types.Instance;
import cc.mallet.types.InstanceList;
/**
* GE Constraint on the probability of self-transitions in the FST.
*
* @author Gregory Druck
*/
public class SelfTransitionGEConstraint implements GEConstraint {
private double selfTransProb;
private double numTokens;
private double expectation;
private double weight;
/**
* @param selfTransProb Probability of self-transition
* @param weight Weight of this constraint in the objective function
*/
public SelfTransitionGEConstraint(double selfTransProb, double weight) {
this.selfTransProb = selfTransProb;
this.weight = weight;
this.numTokens = 0;
this.expectation = 0;
}
private SelfTransitionGEConstraint(double selfTransProb, double weight, double numTokens, double expectation) {
this.selfTransProb = selfTransProb;
this.weight = weight;
this.numTokens = numTokens;
this.expectation = expectation;
}
public GEConstraint copy() {
return new SelfTransitionGEConstraint(selfTransProb, weight, numTokens, expectation);
}
public boolean isOneStateConstraint() {
return false;
}
public void setStateLabelMap(StateLabelMap map) {}
// no pre-processing possible here
public void preProcess(FeatureVector fv) {}
public BitSet preProcess(InstanceList data) {
// count number of tokens
BitSet bitSet = new BitSet(data.size());
bitSet.set(0, data.size(), true);
for (Instance instance : data) {
FeatureVectorSequence fvs = (FeatureVectorSequence)instance.getData();
this.numTokens += fvs.size();
}
return bitSet;
}
public double getCompositeConstraintFeatureValue(FeatureVector fv, int ip, int si1, int si2) {
if (si1 == si2) {
return this.weight * (selfTransProb / expectation);
}
else {
return this.weight * ((1-selfTransProb) / (numTokens - expectation));
}
}
public double getValue() {
double selfTransEx = this.expectation / this.numTokens;
if (selfTransProb == 1) {
return weight * Math.log(selfTransEx);
}
else if (selfTransProb == 0) {
return weight * Math.log(1-selfTransEx);
}
return weight * (selfTransProb * (Math.log(selfTransEx) - Math.log(selfTransProb))
+ ((1-selfTransProb) * (Math.log(1-selfTransEx)-Math.log(1-selfTransProb))));
}
public void zeroExpectations() {
this.expectation = 0;
}
public void computeExpectations(ArrayList<SumLattice> lattices) {
double[][][] xis;
for (int i = 0; i < lattices.size(); i++) {
SumLattice lattice = lattices.get(i);
xis = lattice.getXis();
int numStates = xis[0].length;
FeatureVectorSequence fvs = (FeatureVectorSequence)lattice.getInput();
for (int ip = 0; ip < fvs.size(); ++ip) {
for (int si = 0; si < numStates; si++) {
this.expectation += Math.exp(xis[ip][si][si]);
}
}
}
System.err.println("Self transition expectation: " + (this.expectation/this.numTokens));
}
}