org.tribuo.classification.sgd.crf

Class ChainHelper

java.lang.Object

org.tribuo.classification.sgd.crf.ChainHelper

public final class ChainHelper
extends Object

A collection of helper methods for performing training and inference in a CRF.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	ChainHelper.ChainBPResults Belief Propagation results.
static class	ChainHelper.ChainCliqueValues Clique scores within a chain.
static class	ChainHelper.ChainViterbiResults Viterbi output from a linear chain.

Method Summary

Modifier and Type	Method and Description
static ChainHelper.ChainBPResults	beliefPropagation(ChainHelper.ChainCliqueValues scores) Runs belief propagation on a linear chain CRF.
static double	constrainedBeliefPropagation(ChainHelper.ChainCliqueValues scores, int[] constraints) Runs constrained belief propagation on a linear chain CRF.
static double	sumLogProbs(DenseVector input) Sums the log probabilities.
static double	sumLogProbs(double[] input) Sums the log probabilities.
static ChainHelper.ChainViterbiResults	viterbi(ChainHelper.ChainCliqueValues scores) Runs Viterbi on a linear chain CRF.

Methods inherited from class java.lang.Object

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

Method Detail

beliefPropagation

public static ChainHelper.ChainBPResults beliefPropagation(ChainHelper.ChainCliqueValues scores)

Runs belief propagation on a linear chain CRF. Uses the linear predictions for each token and the label transition probabilities.

Parameters:

scores - Tuple containing the label-label transition matrix, and the per token label scores.

Returns:

Tuple containing the normalising constant, the forward values and the backward values.

constrainedBeliefPropagation

public static double constrainedBeliefPropagation(ChainHelper.ChainCliqueValues scores,
                                                  int[] constraints)

Runs constrained belief propagation on a linear chain CRF. Uses the linear predictions for each token and the label transition probabilities.

See:

 "Confidence Estimation for Information Extraction",
 A. Culotta and A. McCallum
 Proceedings of HLT-NAACL 2004: Short Papers, 2004.
 

Parameters:

scores - Tuple containing the label-label transition matrix, and the per token label scores.

constraints - An array of integers, representing the label constraints. -1 signifies no constraint, otherwise it's the label id.

Returns:

The normalization constant for this constrained run.

viterbi

public static ChainHelper.ChainViterbiResults viterbi(ChainHelper.ChainCliqueValues scores)

Runs Viterbi on a linear chain CRF. Uses the linear predictions for each token and the label transition probabilities.

Parameters:

scores - Tuple containing the label-label transition matrix, and the per token label scores.

Returns:

Tuple containing the score of the maximum path and the maximum predicted label per token.

sumLogProbs

public static double sumLogProbs(DenseVector input)

Sums the log probabilities. Must be updated in concert with sumLogProbs(double[]).

Parameters:

input - A DenseVector of log probabilities.

Returns:

log sum exp input[i].

sumLogProbs

public static double sumLogProbs(double[] input)

Sums the log probabilities. Must be updated in concert with sumLogProbs(DenseVector).

Parameters:

input - A double array of log probabilities.

Returns:

log sum exp input[i].