de.up.ling.irtg.maxent

Class MaximumEntropyIrtg

java.lang.Object

de.up.ling.irtg.InterpretedTreeAutomaton

de.up.ling.irtg.maxent.MaximumEntropyIrtg

All Implemented Interfaces:

Serializable

public class MaximumEntropyIrtg
extends InterpretedTreeAutomaton

An IRTG with a log-linear probability model. In addition to the ordinary IRTG rules (cf. InterpretedTreeAutomaton), a maxent IRTG allows you to specify a set of feature functions, cf. FeatureFunction. In the grammar file, these are declared using the feature keyword, and come after the interpretation declarations and before the rules of the grammar.

You can use any subclass of FeatureFunction in your grammar, as long as it is on the classpath. You can then add concrete instances of your feature function classes to the grammar using one of the following two forms:

• feature f1: de.up.ling.irtg.maxent.ChildOfFeature('VP','PP'): Constructs a new object of class ChildOfFeature, passing the strings "VP" and "PP" as the first and second argument to the constructor of the class.
• feature f2: YourClass::staticMethod("a", "b"): Calls the static method YourClass#staticMethod with the given string arguments. The static method is supposed to return an object of a subclass of FeatureFunction, which is then used as the feature function instance with name f2.

Note that different feature instances must have different names (in the example: f1 and f2) so the system can keep their weights apart.

Supervised learning of the weights of a maxent IRTG is performed using trainMaxent(de.up.ling.irtg.corpus.Corpus). Computation of a weighted chart (in which rule weights corresponds to log-linear scores of trees) is done with parseInputObjects(java.util.Map). From this, you can compute the best derivation tree using TreeAutomaton.viterbi() and similar methods.

Author:

koller, Danilo Baumgarten

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
MaximumEntropyIrtg(TreeAutomaton<String> automaton, Map<String,FeatureFunction> featureMap) Constructor

Method Summary

Modifier and Type	Method and Description
FeatureFunction	getFeatureFunction(int index) Returns the feature function referenced by index
FeatureFunction	getFeatureFunction(String name) Returns the feature function referenced by name
List<String>	getFeatureNames() Returns the list of the feature function names
FeatureFunction[]	getFeatures()
double	getFeatureWeight(int i) Returns the weight of a specific feature function referenced by i
double[]	getFeatureWeights() Returns the array of the feature function weights
int	getNumFeatures() Returns the number of features
TreeAutomaton	parseInputObjects(Map<String,Object> inputs) Parses an input of representations and their name and computes a chart for this input The member variable useIrtgParser indicates which parser to use True: the parser of InterpretedTreeAutomaton will be used False: an implementation of a CKY-parser will be used
void	readWeights(Reader reader) Reads the feature function weights from a reader, e.g., string or file The data must be formatted as Java properties.
void	setFeatures(Map<String,FeatureFunction> featureMap) Sets the feature functions
void	setFeatureWeight(int index, double weight) Sets the weight of a specific feature function
void	setFeatureWeights(double[] weights) Sets the array of the feature function weights
static void	setLoggingLevel(Level level)
String	toString() Returns a string representing the object and its elements
boolean	trainMaxent(Corpus corpus)
boolean	trainMaxent(Corpus corpus, ProgressListener listener) Trains the weights for the rules according to the training data.
void	writeWeights(Writer writer) Writes the feature function weights to a writer, e.g., string or file The data will be formatted as Java properties

Methods inherited from class de.up.ling.irtg.InterpretedTreeAutomaton

addAllInterpretations, addInterpretation, bulkParse, bulkParse, decode, decodeToAutomaton, equals, filterBinarizedForAppearingConstants, filterForAppearingConstants, forAlgebras, fromPath, fromString, getAutomaton, getInterpretation, getInterpretations, interpret, interpret, normalizeRuleWeights, parse, parseCondensedWithPruning, parseSimple, parseString, parseWithSiblingFinder, read, readCorpus, setDebug, trainEM, trainEM, trainEM, trainML, trainVB, trainVB, trainVB

Methods inherited from class java.lang.Object

getClass, hashCode, notify, notifyAll, wait, wait, wait

Constructor Detail

MaximumEntropyIrtg

public MaximumEntropyIrtg(TreeAutomaton<String> automaton,
                          Map<String,FeatureFunction> featureMap)

Constructor

Parameters:

automaton - the TreeAutomaton build by grammar rules

featureMap - the map contains feature functions accessed by their names. These functions are used to calculate probabilities for the RTG

Method Detail

setFeatures

public final void setFeatures(Map<String,FeatureFunction> featureMap)

Sets the feature functions

Parameters:

featureMap - the mapping of names to feature functions

setFeatureWeights

public void setFeatureWeights(double[] weights)

Sets the array of the feature function weights

Parameters:

weights - the array of feature weights

setFeatureWeight

public void setFeatureWeight(int index,
                             double weight)

Sets the weight of a specific feature function

Parameters:

index - the position of the weight in the array

weight - the new weight

Throws:

NoFeaturesException - if no features are present

getFeatureWeight

public double getFeatureWeight(int i)

Returns the weight of a specific feature function referenced by i

Parameters:

i - the reference of a feature function

Returns:

double the weight of the feature function

Throws:

NoFeaturesException - if no features are present

getFeatureWeights

public double[] getFeatureWeights()

Returns the array of the feature function weights

Returns:

double[] containing the weights of all feature functions

Throws:

NoFeaturesException - if no features are present

getFeatureNames

public List<String> getFeatureNames()

Returns the list of the feature function names

Returns:

List() containing the names of all feature functions

getFeatureFunction

public FeatureFunction getFeatureFunction(String name)

Returns the feature function referenced by name

Parameters:

name - the name of the feature function

Returns:

the feature function with the name name if no corresponding function is found

Throws:

NoFeaturesException - if no features are present

getFeatureFunction

public FeatureFunction getFeatureFunction(int index)

Returns the feature function referenced by index

Parameters:

index - the index of the feature function

Returns:

the feature function with the index

Throws:

NoFeaturesException - if no features are present

getNumFeatures

public int getNumFeatures()

Returns the number of features

Returns:

number of features

parseInputObjects

public TreeAutomaton parseInputObjects(Map<String,Object> inputs)

Parses an input of representations and their name and computes a chart for this input The member variable useIrtgParser indicates which parser to use True: the parser of InterpretedTreeAutomaton will be used False: an implementation of a CKY-parser will be used

Overrides:

parseInputObjects in class InterpretedTreeAutomaton

Parameters:

inputs - mapping of representations and their names

Returns:

TreeAutomaton the computed chart

trainMaxent

public boolean trainMaxent(Corpus corpus)

trainMaxent

public boolean trainMaxent(Corpus corpus,
                           ProgressListener listener)

Trains the weights for the rules according to the training data.

Parameters:

corpus - the training data containing sentences and their parse tree

Returns:

true iff L-BFGS optimization was successful

readWeights

public void readWeights(Reader reader)
                 throws IOException

Reads the feature function weights from a reader, e.g., string or file The data must be formatted as Java properties. It's assumed that the feature functions are already set. Setting the feature functions will create a new array of weights.

Parameters:

reader - the reader to read the data from

Throws:

IOException - if the reader cannot read the data properly

writeWeights

public void writeWeights(Writer writer)
                  throws IOException

Writes the feature function weights to a writer, e.g., string or file The data will be formatted as Java properties

Parameters:

writer - the writer to store the data into

Throws:

IOException - if the writer cannot store the data properly

toString

public String toString()

Returns a string representing the object and its elements

Overrides:

toString in class InterpretedTreeAutomaton

Returns:

String the string representing the object

getFeatures

public FeatureFunction[] getFeatures()

Returns:

the features

setLoggingLevel

public static void setLoggingLevel(Level level)