cc.mallet.fst
Class CRF

java.lang.Object
  cc.mallet.fst.Transducer
      cc.mallet.fst.CRF

All Implemented Interfaces:

java.io.Serializable

Direct Known Subclasses:

MEMM

public class CRF
extends Transducer
implements java.io.Serializable

Represents a CRF model.

See Also:

Serialized Form

Nested Class Summary

static class	CRF.Factors A simple, transparent container to hold the parameters or sufficient statistics for the CRF.
static class	CRF.State
protected static class	CRF.TransitionIterator

Nested classes/interfaces inherited from class cc.mallet.fst.Transducer

Transducer.Incrementor

Field Summary

protected int	cachedNumParametersStamp
protected java.util.ArrayList<FeatureInducer>	featureInducers
protected FeatureSelection[]	featureSelections
protected FeatureSelection	globalFeatureSelection
protected java.util.ArrayList<CRF.State>	initialStates
protected Alphabet	inputAlphabet
protected java.util.HashMap<java.lang.String,CRF.State>	name2state
protected int	numParameters
protected Alphabet	outputAlphabet
protected CRF.Factors	parameters
protected java.util.ArrayList<CRF.State>	states
protected int	weightsStructureChangeStamp
protected int	weightsValueChangeStamp

Fields inherited from class cc.mallet.fst.Transducer

CERTAIN_WEIGHT, IMPOSSIBLE_WEIGHT, inputPipe, outputPipe

Constructor Summary

CRF(Alphabet inputAlphabet, Alphabet outputAlphabet)

CRF(CRF other)
Create a CRF whose states and weights are a copy of those from another CRF.

CRF(Pipe inputPipe, Pipe outputPipe)

Method Summary

void	addFullyConnectedStates(java.lang.String[] stateNames) Add a group of states that are fully connected with each other, with parameters equal zero, and labels on their out-going arcs the same name as their destination state names.
void	addFullyConnectedStatesForBiLabels()
void	addFullyConnectedStatesForLabels()
void	addFullyConnectedStatesForThreeQuarterLabels(InstanceList trainingSet)
void	addFullyConnectedStatesForTriLabels()
java.lang.String	addOrderNStates(InstanceList trainingSet, int[] orders, boolean[] defaults, java.lang.String start, java.util.regex.Pattern forbidden, java.util.regex.Pattern allowed, boolean fullyConnected) Assumes that the CRF's output alphabet contains Strings.
void	addSelfTransitioningStateForAllLabels(java.lang.String name)
void	addStartState()
void	addStartState(java.lang.String name)
void	addState(java.lang.String name, double initialWeight, double finalWeight, java.lang.String[] destinationNames, java.lang.String[] labelNames) Default gives separate parameters to each transition.
void	addState(java.lang.String name, double initialWeight, double finalWeight, java.lang.String[] destinationNames, java.lang.String[] labelNames, java.lang.String[] weightNames)
void	addState(java.lang.String name, double initialWeight, double finalWeight, java.lang.String[] destinationNames, java.lang.String[] labelNames, java.lang.String[][] weightNames)
void	addState(java.lang.String name, java.lang.String[] destinationNames) Add a state with parameters equal zero, and labels on out-going arcs the same name as their destination state names.
void	addStatesForBiLabelsConnectedAsIn(InstanceList trainingSet) Add states to create a second-order Markov model on labels, adding only those transitions the occur in the given trainingSet.
void	addStatesForHalfLabelsConnectedAsIn(InstanceList trainingSet) Add as many states as there are labels, but don't create separate weights for each source-destination pair of states.
void	addStatesForLabelsConnectedAsIn(InstanceList trainingSet) Add states to create a first-order Markov model on labels, adding only those transitions the occur in the given trainingSet.
void	addStatesForThreeQuarterLabelsConnectedAsIn(InstanceList trainingSet) Add as many states as there are labels, but don't create separate observational-test-weights for each source-destination pair of states---instead have all the incoming transitions to a state share the same observational-feature-test weights.
void	evaluate(TransducerEvaluator eval, InstanceList testing) Deprecated.
void	freezeWeights(int weightsIndex) Freezes a set of weights to their current values.
void	freezeWeights(java.lang.String weightsName) Freezes a set of weights to their current values.
double[]	getDefaultWeights()
Alphabet	getInputAlphabet()
int	getNumParameters()
Alphabet	getOutputAlphabet()
double	getParameter(int sourceStateIndex, int destStateIndex, int featureIndex) Only gets the parameter from the first group of parameters.
double	getParameter(int sourceStateIndex, int destStateIndex, int featureIndex, int weightIndex)
CRF.Factors	getParameters()
double	getParametersAbsNorm()
Transducer.State	getState(int index)
CRF.State	getState(java.lang.String name)
SparseVector[]	getWeights()
SparseVector	getWeights(int weightIndex)
SparseVector	getWeights(java.lang.String weightName)
int	getWeightsIndex(java.lang.String weightName)
java.lang.String	getWeightsName(int weightIndex)
int	getWeightsStructureChangeStamp()
int	getWeightsValueChangeStamp()
void	induceFeaturesFor(InstanceList instances) When the CRF has done feature induction, these new feature conjunctions must be created in the test or validation data in order for them to take effect.
java.util.Iterator	initialStateIterator()
boolean	isTrainable()
boolean	isWeightsFrozen(int weightsIndex)
protected CRF.State	newState(java.lang.String name, int index, double initialWeight, double finalWeight, java.lang.String[] destinationNames, java.lang.String[] labelNames, java.lang.String[][] weightNames, CRF crf)
int	numStates()
Sequence[]	predict(InstanceList testing) Deprecated.
void	print()
void	print(java.io.PrintWriter out)
void	setAsStartState(CRF.State state)
void	setDefaultWeight(int widx, double val)
void	setDefaultWeights(double[] w)
void	setFeatureSelection(int weightIdx, FeatureSelection fs)
void	setParameter(int sourceStateIndex, int destStateIndex, int featureIndex, double value) Only sets the parameter from the first group of parameters.
void	setParameter(int sourceStateIndex, int destStateIndex, int featureIndex, int weightIndex, double value)
void	setWeights(int weightsIndex, SparseVector transitionWeights)
void	setWeights(SparseVector[] m)
void	setWeights(java.lang.String weightName, SparseVector transitionWeights)
void	setWeightsDimensionAsIn(InstanceList trainingData)
void	setWeightsDimensionAsIn(InstanceList trainingData, boolean useSomeUnsupportedTrick)
void	setWeightsDimensionDensely()
void	unfreezeWeights(java.lang.String weightsName) Unfreezes a set of weights.
void	weightsStructureChanged() This method should be called whenever the CRFs weights (parameters) have their structure/arity/number changed.
void	weightsValueChanged() This method should be called whenever the CRFs weights (parameters) are changed.
void	write(java.io.File f)

Methods inherited from class cc.mallet.fst.Transducer

averageTokenAccuracy, canIterateAllTransitions, generatePath, getInputPipe, getMaxLatticeFactory, getOutputPipe, getSumLatticeFactory, isGenerative, label, less_efficient_sumLogProb, no_longer_needed_sumNegLogProb, setMaxLatticeFactory, setSumLatticeFactory, stateIndexOfString, sumLogProb, transduce, transduce

Methods inherited from class java.lang.Object

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

Field Detail

inputAlphabet

protected Alphabet inputAlphabet

outputAlphabet

protected Alphabet outputAlphabet

states

protected java.util.ArrayList<CRF.State> states

initialStates

protected java.util.ArrayList<CRF.State> initialStates

name2state

protected java.util.HashMap<java.lang.String,CRF.State> name2state

parameters

protected CRF.Factors parameters

globalFeatureSelection

protected FeatureSelection globalFeatureSelection

featureSelections

protected FeatureSelection[] featureSelections

featureInducers

protected java.util.ArrayList<FeatureInducer> featureInducers

weightsValueChangeStamp

protected int weightsValueChangeStamp

weightsStructureChangeStamp

protected int weightsStructureChangeStamp

cachedNumParametersStamp

protected int cachedNumParametersStamp

numParameters

protected int numParameters

Constructor Detail

CRF

public CRF(Pipe inputPipe,
           Pipe outputPipe)

CRF

public CRF(Alphabet inputAlphabet,
           Alphabet outputAlphabet)

CRF

public CRF(CRF other)

Create a CRF whose states and weights are a copy of those from another CRF.

Method Detail

getInputAlphabet

public Alphabet getInputAlphabet()

getOutputAlphabet

public Alphabet getOutputAlphabet()

weightsStructureChanged

public void weightsStructureChanged()

This method should be called whenever the CRFs weights (parameters) have their structure/arity/number changed.

weightsValueChanged

public void weightsValueChanged()

This method should be called whenever the CRFs weights (parameters) are changed.

newState

protected CRF.State newState(java.lang.String name,
                             int index,
                             double initialWeight,
                             double finalWeight,
                             java.lang.String[] destinationNames,
                             java.lang.String[] labelNames,
                             java.lang.String[][] weightNames,
                             CRF crf)

addState

public void addState(java.lang.String name,
                     double initialWeight,
                     double finalWeight,
                     java.lang.String[] destinationNames,
                     java.lang.String[] labelNames,
                     java.lang.String[][] weightNames)

addState

public void addState(java.lang.String name,
                     double initialWeight,
                     double finalWeight,
                     java.lang.String[] destinationNames,
                     java.lang.String[] labelNames,
                     java.lang.String[] weightNames)

addState

public void addState(java.lang.String name,
                     double initialWeight,
                     double finalWeight,
                     java.lang.String[] destinationNames,
                     java.lang.String[] labelNames)

Default gives separate parameters to each transition.

addState

public void addState(java.lang.String name,
                     java.lang.String[] destinationNames)

Add a state with parameters equal zero, and labels on out-going arcs the same name as their destination state names.

addFullyConnectedStates

public void addFullyConnectedStates(java.lang.String[] stateNames)

Add a group of states that are fully connected with each other, with parameters equal zero, and labels on their out-going arcs the same name as their destination state names.

addFullyConnectedStatesForLabels

public void addFullyConnectedStatesForLabels()

addStartState

public void addStartState()

addStartState

public void addStartState(java.lang.String name)

setAsStartState

public void setAsStartState(CRF.State state)

addStatesForLabelsConnectedAsIn

public void addStatesForLabelsConnectedAsIn(InstanceList trainingSet)

Add states to create a first-order Markov model on labels, adding only those transitions the occur in the given trainingSet.

addStatesForHalfLabelsConnectedAsIn

public void addStatesForHalfLabelsConnectedAsIn(InstanceList trainingSet)

Add as many states as there are labels, but don't create separate weights for each source-destination pair of states. Instead have all the incoming transitions to a state share the same weights.

addStatesForThreeQuarterLabelsConnectedAsIn

public void addStatesForThreeQuarterLabelsConnectedAsIn(InstanceList trainingSet)

Add as many states as there are labels, but don't create separate observational-test-weights for each source-destination pair of states---instead have all the incoming transitions to a state share the same observational-feature-test weights. However, do create separate default feature for each transition, (which acts as an HMM-style transition probability).

addFullyConnectedStatesForThreeQuarterLabels

public void addFullyConnectedStatesForThreeQuarterLabels(InstanceList trainingSet)

addFullyConnectedStatesForBiLabels

public void addFullyConnectedStatesForBiLabels()

addStatesForBiLabelsConnectedAsIn

public void addStatesForBiLabelsConnectedAsIn(InstanceList trainingSet)

Add states to create a second-order Markov model on labels, adding only those transitions the occur in the given trainingSet.

addFullyConnectedStatesForTriLabels

public void addFullyConnectedStatesForTriLabels()

addSelfTransitioningStateForAllLabels

public void addSelfTransitioningStateForAllLabels(java.lang.String name)

addOrderNStates

public java.lang.String addOrderNStates(InstanceList trainingSet,
                                        int[] orders,
                                        boolean[] defaults,
                                        java.lang.String start,
                                        java.util.regex.Pattern forbidden,
                                        java.util.regex.Pattern allowed,
                                        boolean fullyConnected)

Assumes that the CRF's output alphabet contains Strings. Creates an order-n CRF with input predicates and output labels given by trainingSet and order, connectivity, and weights given by the remaining arguments.

Parameters:

trainingSet - the training instances

orders - an array of increasing non-negative numbers giving the orders of the features for this CRF. The largest number n is the Markov order of the CRF. States are n-tuples of output labels. Each of the other numbers k in orders represents a weight set shared by all destination states whose last (most recent) k labels agree. If orders is null, an order-0 CRF is built.

defaults - If non-null, it must be the same length as orders, with true positions indicating that the weight set for the corresponding order contains only the weight for a default feature; otherwise, the weight set has weights for all features built from input predicates.

start - The label that represents the context of the start of a sequence. It may be also used for sequence labels. If no label of this name exists, one will be added. Connection wills be added between the start label and all other labels, even if fullyConnected is false. This argument may be null, in which case no special start state is added.

forbidden - If non-null, specifies what pairs of successive labels are not allowed, both for constructing norder states or for transitions. A label pair (u,v) is not allowed if u + "," + v matches forbidden.

allowed - If non-null, specifies what pairs of successive labels are allowed, both for constructing norder states or for transitions. A label pair (u,v) is allowed only if u + "," + v matches allowed.

fullyConnected - Whether to include all allowed transitions, even those not occurring in trainingSet,

Returns:

The name of the start state.

getState

public CRF.State getState(java.lang.String name)

setWeights

public void setWeights(int weightsIndex,
                       SparseVector transitionWeights)

setWeights

public void setWeights(java.lang.String weightName,
                       SparseVector transitionWeights)

getWeightsName

public java.lang.String getWeightsName(int weightIndex)

getWeights

public SparseVector getWeights(java.lang.String weightName)

getWeights

public SparseVector getWeights(int weightIndex)

getDefaultWeights

public double[] getDefaultWeights()

getWeights

public SparseVector[] getWeights()

setWeights

public void setWeights(SparseVector[] m)

setDefaultWeights

public void setDefaultWeights(double[] w)

setDefaultWeight

public void setDefaultWeight(int widx,
                             double val)

isWeightsFrozen

public boolean isWeightsFrozen(int weightsIndex)

freezeWeights

public void freezeWeights(int weightsIndex)

Freezes a set of weights to their current values. Frozen weights are used for labeling sequences (as in transduce), but are not be modified by the train methods.

Parameters:

weightsIndex - Index of weight set to freeze.

freezeWeights

public void freezeWeights(java.lang.String weightsName)

Freezes a set of weights to their current values. Frozen weights are used for labeling sequences (as in transduce), but are not be modified by the train methods.

Parameters:

weightsName - Name of weight set to freeze.

unfreezeWeights

public void unfreezeWeights(java.lang.String weightsName)

Unfreezes a set of weights. Frozen weights are used for labeling sequences (as in transduce), but are not be modified by the train methods.

Parameters:

weightsName - Name of weight set to unfreeze.

setFeatureSelection

public void setFeatureSelection(int weightIdx,
                                FeatureSelection fs)

setWeightsDimensionAsIn

public void setWeightsDimensionAsIn(InstanceList trainingData)

setWeightsDimensionAsIn

public void setWeightsDimensionAsIn(InstanceList trainingData,
                                    boolean useSomeUnsupportedTrick)

setWeightsDimensionDensely

public void setWeightsDimensionDensely()

getWeightsIndex

public int getWeightsIndex(java.lang.String weightName)

numStates

public int numStates()

Specified by:

numStates in class Transducer

getState

public Transducer.State getState(int index)

Specified by:

getState in class Transducer

initialStateIterator

public java.util.Iterator initialStateIterator()

Specified by:

initialStateIterator in class Transducer

isTrainable

public boolean isTrainable()

getWeightsValueChangeStamp

public int getWeightsValueChangeStamp()

getWeightsStructureChangeStamp

public int getWeightsStructureChangeStamp()

getParameters

public CRF.Factors getParameters()

getParametersAbsNorm

public double getParametersAbsNorm()

setParameter

public void setParameter(int sourceStateIndex,
                         int destStateIndex,
                         int featureIndex,
                         double value)

Only sets the parameter from the first group of parameters.

setParameter

public void setParameter(int sourceStateIndex,
                         int destStateIndex,
                         int featureIndex,
                         int weightIndex,
                         double value)

getParameter

public double getParameter(int sourceStateIndex,
                           int destStateIndex,
                           int featureIndex)

Only gets the parameter from the first group of parameters.

getParameter

public double getParameter(int sourceStateIndex,
                           int destStateIndex,
                           int featureIndex,
                           int weightIndex)

getNumParameters

public int getNumParameters()

predict

@Deprecated
public Sequence[] predict(InstanceList testing)

Deprecated.

This method is deprecated.

evaluate

@Deprecated
public void evaluate(TransducerEvaluator eval,
                                InstanceList testing)

Deprecated.

This method is deprecated.

induceFeaturesFor

public void induceFeaturesFor(InstanceList instances)

When the CRF has done feature induction, these new feature conjunctions must be created in the test or validation data in order for them to take effect.

print

public void print()

Overrides:

print in class Transducer

print

public void print(java.io.PrintWriter out)

write

public void write(java.io.File f)