public SentenceStatistics mean() {
double sumConfidence = 0;
int countWithConfidence = 0;
Counter<String> avePredictions =
new ClassicCounter<>(MapFactory.<String, MutableDouble>linkedHashMapFactory());
// Sum
for (SentenceStatistics stat : this.statisticsForClassifiers) {
for (Double confidence : stat.confidence) {
sumConfidence += confidence;
countWithConfidence += 1;
}
assert Math.abs(stat.relationDistribution.totalCount() - 1.0) < 1e-5;
for (Map.Entry<String, Double> entry : stat.relationDistribution.entrySet()) {
assert entry.getValue() >= 0.0;
assert entry.getValue() == stat.relationDistribution.getCount(entry.getKey());
avePredictions.incrementCount(entry.getKey(), entry.getValue());
assert stat.relationDistribution.getCount(entry.getKey())
== stat.relationDistribution.getCount(entry.getKey());
}
}
// Normalize
double aveConfidence = sumConfidence / ((double) countWithConfidence);
// Return
if (this.statisticsForClassifiers.size() > 1) {
Counters.divideInPlace(avePredictions, (double) this.statisticsForClassifiers.size());
}
if (Math.abs(avePredictions.totalCount() - 1.0) > 1e-5) {
throw new IllegalStateException("Mean relation distribution is not a distribution!");
}
assert this.statisticsForClassifiers.size() > 1
|| this.statisticsForClassifiers.size() == 0
|| Counters.equals(
avePredictions,
statisticsForClassifiers.iterator().next().relationDistribution,
1e-5);
return countWithConfidence > 0
? new SentenceStatistics(avePredictions, aveConfidence)
: new SentenceStatistics(avePredictions);
}
/**
* 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.
*
* @see edu.stanford.nlp.naturalli.ClauseSplitterSearchProblem#search(Predicate)
* @param candidateFragments The callback function.
* @param classifier The classifier for whether an arc should be on the path to a clause split, a
* clause split itself, or neither.
* @param featurizer The featurizer to use during search, to be dot producted with the weights.
*/
public void search(
// The output specs
final Predicate<Triple<Double, List<Counter<String>>, Supplier<SentenceFragment>>>
candidateFragments,
// The learning specs
final Classifier<ClauseSplitter.ClauseClassifierLabel, String> classifier,
final Map<String, List<String>> hardCodedSplits,
final Function<Triple<State, Action, State>, Counter<String>> featurizer,
final int maxTicks) {
Collection<Action> actionSpace = new ArrayList<>();
// SIMPLE SPLIT
actionSpace.add(
new Action() {
@Override
public String signature() {
return "simple";
}
@Override
public boolean prerequisitesMet(SemanticGraph originalTree, SemanticGraphEdge edge) {
char tag = edge.getDependent().tag().charAt(0);
return !(tag != 'V' && tag != 'N' && tag != 'J' && tag != 'P' && tag != 'D');
}
@Override
public Optional<State> applyTo(
SemanticGraph tree,
State source,
SemanticGraphEdge outgoingEdge,
SemanticGraphEdge subjectOrNull,
SemanticGraphEdge objectOrNull) {
return Optional.of(
new State(
outgoingEdge,
subjectOrNull == null ? source.subjectOrNull : subjectOrNull,
subjectOrNull == null ? (source.distanceFromSubj + 1) : 0,
objectOrNull == null ? source.objectOrNull : objectOrNull,
source.thunk.andThen(
toModify -> {
assert Util.isTree(toModify);
simpleClause(toModify, outgoingEdge);
if (outgoingEdge.getRelation().toString().endsWith("comp")) {
stripAuxMark(toModify);
}
assert Util.isTree(toModify);
}),
false));
}
});
// CLONE ROOT
actionSpace.add(
new Action() {
@Override
public String signature() {
return "clone_root_as_nsubjpass";
}
@Override
public boolean prerequisitesMet(SemanticGraph originalTree, SemanticGraphEdge edge) {
// Only valid if there's a single nontrivial outgoing edge from a node. Otherwise it's a
// whole can of worms.
Iterator<SemanticGraphEdge> iter =
originalTree.outgoingEdgeIterable(edge.getGovernor()).iterator();
if (!iter.hasNext()) {
return false; // what?
}
boolean nontrivialEdge = false;
while (iter.hasNext()) {
SemanticGraphEdge outEdge = iter.next();
switch (outEdge.getRelation().toString()) {
case "nn":
case "amod":
break;
default:
if (nontrivialEdge) {
return false;
}
nontrivialEdge = true;
}
}
return true;
}
@Override
public Optional<State> applyTo(
SemanticGraph tree,
State source,
SemanticGraphEdge outgoingEdge,
SemanticGraphEdge subjectOrNull,
SemanticGraphEdge objectOrNull) {
return Optional.of(
new State(
outgoingEdge,
subjectOrNull == null ? source.subjectOrNull : subjectOrNull,
subjectOrNull == null ? (source.distanceFromSubj + 1) : 0,
objectOrNull == null ? source.objectOrNull : objectOrNull,
source.thunk.andThen(
toModify -> {
assert Util.isTree(toModify);
simpleClause(toModify, outgoingEdge);
addSubtree(
toModify,
outgoingEdge.getDependent(),
"nsubjpass",
tree,
outgoingEdge.getGovernor(),
Collections.singleton(outgoingEdge));
// addWord(toModify, outgoingEdge.getDependent(), "auxpass",
// mockNode(outgoingEdge.getDependent().backingLabel(), "is", "VBZ"));
assert Util.isTree(toModify);
}),
true));
}
});
// COPY SUBJECT
actionSpace.add(
new Action() {
@Override
public String signature() {
return "clone_nsubj";
}
@Override
public boolean prerequisitesMet(SemanticGraph originalTree, SemanticGraphEdge edge) {
// Don't split into anything but verbs or nouns
char tag = edge.getDependent().tag().charAt(0);
if (tag != 'V' && tag != 'N') {
return false;
}
for (SemanticGraphEdge grandchild :
originalTree.outgoingEdgeIterable(edge.getDependent())) {
if (grandchild.getRelation().toString().contains("subj")) {
return false;
}
}
return true;
}
@Override
public Optional<State> applyTo(
SemanticGraph tree,
State source,
SemanticGraphEdge outgoingEdge,
SemanticGraphEdge subjectOrNull,
SemanticGraphEdge objectOrNull) {
if (subjectOrNull != null && !outgoingEdge.equals(subjectOrNull)) {
return Optional.of(
new State(
outgoingEdge,
subjectOrNull,
0,
objectOrNull == null ? source.objectOrNull : objectOrNull,
source.thunk.andThen(
toModify -> {
assert Util.isTree(toModify);
simpleClause(toModify, outgoingEdge);
addSubtree(
toModify,
outgoingEdge.getDependent(),
"nsubj",
tree,
subjectOrNull.getDependent(),
Collections.singleton(outgoingEdge));
assert Util.isTree(toModify);
stripAuxMark(toModify);
assert Util.isTree(toModify);
}),
false));
} else {
return Optional.empty();
}
}
});
// COPY OBJECT
actionSpace.add(
new Action() {
@Override
public String signature() {
return "clone_dobj";
}
@Override
public boolean prerequisitesMet(SemanticGraph originalTree, SemanticGraphEdge edge) {
// Don't split into anything but verbs or nouns
char tag = edge.getDependent().tag().charAt(0);
if (tag != 'V' && tag != 'N') {
return false;
}
for (SemanticGraphEdge grandchild :
originalTree.outgoingEdgeIterable(edge.getDependent())) {
if (grandchild.getRelation().toString().contains("subj")) {
return false;
}
}
return true;
}
@Override
public Optional<State> applyTo(
SemanticGraph tree,
State source,
SemanticGraphEdge outgoingEdge,
SemanticGraphEdge subjectOrNull,
SemanticGraphEdge objectOrNull) {
if (objectOrNull != null && !outgoingEdge.equals(objectOrNull)) {
return Optional.of(
new State(
outgoingEdge,
subjectOrNull == null ? source.subjectOrNull : subjectOrNull,
subjectOrNull == null ? (source.distanceFromSubj + 1) : 0,
objectOrNull,
source.thunk.andThen(
toModify -> {
assert Util.isTree(toModify);
// Split the clause
simpleClause(toModify, outgoingEdge);
// Attach the new subject
addSubtree(
toModify,
outgoingEdge.getDependent(),
"nsubj",
tree,
objectOrNull.getDependent(),
Collections.singleton(outgoingEdge));
// Strip bits we don't want
assert Util.isTree(toModify);
stripAuxMark(toModify);
assert Util.isTree(toModify);
}),
false));
} else {
return Optional.empty();
}
}
});
for (IndexedWord root : tree.getRoots()) {
search(
root, candidateFragments, classifier, hardCodedSplits, featurizer, actionSpace, maxTicks);
}
}
/**
* A helper to add an entire subtree to a given dependency tree.
*
* @param toModify The tree to add the subtree to.
* @param root The root of the tree where we should be adding the subtree.
* @param rel The relation to add the subtree with.
* @param originalTree The orignal tree (i.e., {@link ClauseSplitterSearchProblem#tree}).
* @param subject The root of the clause to add.
* @param ignoredEdges The edges to ignore adding when adding this subtree.
*/
private static void addSubtree(
SemanticGraph toModify,
IndexedWord root,
String rel,
SemanticGraph originalTree,
IndexedWord subject,
Collection<SemanticGraphEdge> ignoredEdges) {
if (toModify.containsVertex(subject)) {
return; // This subtree already exists.
}
Queue<IndexedWord> fringe = new LinkedList<>();
Collection<IndexedWord> wordsToAdd = new ArrayList<>();
Collection<SemanticGraphEdge> edgesToAdd = new ArrayList<>();
// Search for subtree to add
for (SemanticGraphEdge edge : originalTree.outgoingEdgeIterable(subject)) {
if (!ignoredEdges.contains(edge)) {
if (toModify.containsVertex(edge.getDependent())) {
// Case: we're adding a subtree that's not disjoint from toModify. This is bad news.
return;
}
edgesToAdd.add(edge);
fringe.add(edge.getDependent());
}
}
while (!fringe.isEmpty()) {
IndexedWord node = fringe.poll();
wordsToAdd.add(node);
for (SemanticGraphEdge edge : originalTree.outgoingEdgeIterable(node)) {
if (!ignoredEdges.contains(edge)) {
if (toModify.containsVertex(edge.getDependent())) {
// Case: we're adding a subtree that's not disjoint from toModify. This is bad news.
return;
}
edgesToAdd.add(edge);
fringe.add(edge.getDependent());
}
}
}
// Add subtree
// (add subject)
toModify.addVertex(subject);
toModify.addEdge(
root,
subject,
GrammaticalRelation.valueOf(Language.English, rel),
Double.NEGATIVE_INFINITY,
false);
// (add nodes)
wordsToAdd.forEach(toModify::addVertex);
// (add edges)
for (SemanticGraphEdge edge : edgesToAdd) {
assert !toModify.incomingEdgeIterator(edge.getDependent()).hasNext();
toModify.addEdge(
edge.getGovernor(),
edge.getDependent(),
edge.getRelation(),
edge.getWeight(),
edge.isExtra());
}
}
public void addInPlace(SentenceStatistics stats) {
statisticsForClassifiers.add(stats);
}