edu.stanford.nlp.naturalli

Class ClauseSplitterSearchProblem

java.lang.Object

edu.stanford.nlp.naturalli.ClauseSplitterSearchProblem

public class ClauseSplitterSearchProblem
extends java.lang.Object

A search problem for finding clauses in a sentence.

For usage at test time, load a model from ClauseSplitter.load(String), and then take the top clauses of a given tree with topClauses(double, int), yielding a list of SentenceFragments.

   
     ClauseSearcher searcher = ClauseSearcher.factory("/model/path/");
     List<SentenceFragment> sentences = searcher.topClauses(threshold);
   
 

For training, see ClauseSplitter.train(Stream, File, File).

Author:

Gabor Angeli

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static interface	ClauseSplitterSearchProblem.Action An action being taken; that is, the type of clause splitting going on.
static interface	ClauseSplitterSearchProblem.Featurizer Mostly just an alias, but make sure our featurizer is serializable!
class	ClauseSplitterSearchProblem.State A search state.
static class	ClauseSplitterSearchProblem.TrainingOptions The options used for training the clause searcher.

Field Summary

Fields

Modifier and Type	Field and Description
boolean	assumedTruth The assumed truth of the original clause.
static ClauseSplitterSearchProblem.Featurizer	DEFAULT_FEATURIZER The default featurizer to use during training.
protected static java.util.Map<java.lang.String,java.util.List<java.lang.String>>	HARD_SPLITS A specification for clause splits we _always_ want to do.
protected static java.util.Set<java.lang.String>	INDIRECT_SPEECH_LEMMAS A set of words which indicate that the complement clause is not factual, or at least not necessarily factual.
int	sentenceLength The length of the sentence, as determined from the tree.
SemanticGraph	tree The tree to search over.

Constructor Summary

Constructors

Modifier	Constructor and Description
	ClauseSplitterSearchProblem(SemanticGraph tree, boolean assumedTruth) Create a clause searcher which searches naively through every possible subtree as a clause.
protected	ClauseSplitterSearchProblem(SemanticGraph tree, boolean assumedTruth, java.util.Optional<Classifier<ClauseSplitter.ClauseClassifierLabel,java.lang.String>> isClauseClassifier, java.util.Optional<java.util.function.Function<Triple<ClauseSplitterSearchProblem.State,ClauseSplitterSearchProblem.Action,ClauseSplitterSearchProblem.State>,Counter<java.lang.String>>> featurizer) Create a searcher manually, suppling a dependency tree, an optional classifier for when to split clauses, and a featurizer for that classifier.

Method Summary

Modifier and Type	Method and Description
protected void	search(IndexedWord root, java.util.function.Predicate<Triple<java.lang.Double,java.util.List<Counter<java.lang.String>>,java.util.function.Supplier<SentenceFragment>>> candidateFragments, Classifier<ClauseSplitter.ClauseClassifierLabel,java.lang.String> classifier, java.util.Map<java.lang.String,? extends java.util.List<java.lang.String>> hardCodedSplits, java.util.function.Function<Triple<ClauseSplitterSearchProblem.State,ClauseSplitterSearchProblem.Action,ClauseSplitterSearchProblem.State>,Counter<java.lang.String>> featurizer, java.util.Collection<ClauseSplitterSearchProblem.Action> actionSpace, int maxTicks) The core implementation of the search.
void	search(java.util.function.Predicate<Triple<java.lang.Double,java.util.List<Counter<java.lang.String>>,java.util.function.Supplier<SentenceFragment>>> candidateFragments) Search, using the default weights / featurizer.
void	search(java.util.function.Predicate<Triple<java.lang.Double,java.util.List<Counter<java.lang.String>>,java.util.function.Supplier<SentenceFragment>>> candidateFragments, Classifier<ClauseSplitter.ClauseClassifierLabel,java.lang.String> classifier, java.util.Map<java.lang.String,java.util.List<java.lang.String>> hardCodedSplits, java.util.function.Function<Triple<ClauseSplitterSearchProblem.State,ClauseSplitterSearchProblem.Action,ClauseSplitterSearchProblem.State>,Counter<java.lang.String>> featurizer, int maxTicks) Search from the root of the tree.
java.util.List<SentenceFragment>	topClauses(double thresholdProbability, int maxClauses) Get the top few clauses from this searcher, cutting off at the given minimum probability.

Methods inherited from class java.lang.Object

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

Field Detail

HARD_SPLITS

protected static final java.util.Map<java.lang.String,java.util.List<java.lang.String>> HARD_SPLITS

A specification for clause splits we _always_ want to do. The format is a map from the edge label we are splitting, to the preference for the type of split we should do. The most preferred is at the front of the list, and then it backs off to the less and less preferred split types.

INDIRECT_SPEECH_LEMMAS

protected static final java.util.Set<java.lang.String> INDIRECT_SPEECH_LEMMAS

A set of words which indicate that the complement clause is not factual, or at least not necessarily factual.

tree

public final SemanticGraph tree

The tree to search over.

assumedTruth

public final boolean assumedTruth

The assumed truth of the original clause.

sentenceLength

public final int sentenceLength

The length of the sentence, as determined from the tree.

DEFAULT_FEATURIZER

public static final ClauseSplitterSearchProblem.Featurizer DEFAULT_FEATURIZER

The default featurizer to use during training.

Constructor Detail

ClauseSplitterSearchProblem

protected ClauseSplitterSearchProblem(SemanticGraph tree,
                                      boolean assumedTruth,
                                      java.util.Optional<Classifier<ClauseSplitter.ClauseClassifierLabel,java.lang.String>> isClauseClassifier,
                                      java.util.Optional<java.util.function.Function<Triple<ClauseSplitterSearchProblem.State,ClauseSplitterSearchProblem.Action,ClauseSplitterSearchProblem.State>,Counter<java.lang.String>>> featurizer)

Create a searcher manually, suppling a dependency tree, an optional classifier for when to split clauses, and a featurizer for that classifier. You almost certainly want to use ClauseSplitter.load(String) instead of this constructor.

Parameters:

tree - The dependency tree to search over.

assumedTruth - The assumed truth of the tree (relevant for natural logic inference). If in doubt, pass in true.

isClauseClassifier - The classifier for whether a given dependency arc should be a new clause. If this is not given, all arcs are treated as clause separators.

featurizer - The featurizer for the classifier. If no featurizer is given, one should be given in search(java.util.function.Predicate, Classifier, Map, java.util.function.Function, int), or else the classifier will be useless.

See Also:

ClauseSplitter.load(String)

ClauseSplitterSearchProblem

public ClauseSplitterSearchProblem(SemanticGraph tree,
                                   boolean assumedTruth)

Create a clause searcher which searches naively through every possible subtree as a clause. For an end-user, this is almost certainly not what you want. However, it is very useful for training time.

Parameters:

tree - The dependency tree to search over.

assumedTruth - The truth of the premise. Almost always True.

Method Detail

topClauses

public java.util.List<SentenceFragment> topClauses(double thresholdProbability,
                                                   int maxClauses)

Get the top few clauses from this searcher, cutting off at the given minimum probability.

Parameters:

thresholdProbability - The threshold under which to stop returning clauses. This should be between 0 and 1.

maxClauses - A hard limit on the number of clauses to return.

Returns:

The resulting SentenceFragment objects, representing the top clauses of the sentence.

search

public void search(java.util.function.Predicate<Triple<java.lang.Double,java.util.List<Counter<java.lang.String>>,java.util.function.Supplier<SentenceFragment>>> candidateFragments)

Search, using the default weights / featurizer. This is the most common entry method for the raw search, though topClauses(double, int) may be a more convenient method for an end user.

Parameters:

candidateFragments - The callback function for results. The return value defines whether to continue searching.

search

public void search(java.util.function.Predicate<Triple<java.lang.Double,java.util.List<Counter<java.lang.String>>,java.util.function.Supplier<SentenceFragment>>> candidateFragments,
                   Classifier<ClauseSplitter.ClauseClassifierLabel,java.lang.String> classifier,
                   java.util.Map<java.lang.String,java.util.List<java.lang.String>> hardCodedSplits,
                   java.util.function.Function<Triple<ClauseSplitterSearchProblem.State,ClauseSplitterSearchProblem.Action,ClauseSplitterSearchProblem.State>,Counter<java.lang.String>> featurizer,
                   int maxTicks)

Search from the root of the tree. This function also defines the default action space to use during search. This is NOT recommended to be used at test time.

Parameters:

candidateFragments - The callback function.

classifier - The classifier for whether an arc should be on the path to a clause split, a clause split itself, or neither.

featurizer - The featurizer to use during search, to be dot producted with the weights.

See Also:

search(Predicate)

search

protected void search(IndexedWord root,
                      java.util.function.Predicate<Triple<java.lang.Double,java.util.List<Counter<java.lang.String>>,java.util.function.Supplier<SentenceFragment>>> candidateFragments,
                      Classifier<ClauseSplitter.ClauseClassifierLabel,java.lang.String> classifier,
                      java.util.Map<java.lang.String,? extends java.util.List<java.lang.String>> hardCodedSplits,
                      java.util.function.Function<Triple<ClauseSplitterSearchProblem.State,ClauseSplitterSearchProblem.Action,ClauseSplitterSearchProblem.State>,Counter<java.lang.String>> featurizer,
                      java.util.Collection<ClauseSplitterSearchProblem.Action> actionSpace,
                      int maxTicks)

The core implementation of the search.

Parameters:

root - The root word to search from. Traditionally, this is the root of the sentence.

candidateFragments - The callback for the resulting sentence fragments. This is a predicate of a triple of values. The return value of the predicate determines whether we should continue searching. The triple is a triple of

1. The log probability of the sentence fragment, according to the featurizer and the weights
2. The features along the path to this fragment. The last element of this is the features from the most recent step.
3. The sentence fragment. Because it is relatively expensive to compute the resulting tree, this is returned as a lazy Supplier.

classifier - The classifier for whether an arc should be on the path to a clause split, a clause split itself, or neither.

featurizer - The featurizer to use. Make sure this matches the weights!

actionSpace - The action space we are allowed to take. Each action defines a means of splitting a clause on a dependency boundary.