@Override
public Pair<DeepTree, DeepTree> process(Tree tree) {
// For each tree, move in the direction of the gold tree, and
// move away from the direction of the best scoring hypothesis
IdentityHashMap<Tree, SimpleMatrix> goldVectors = new IdentityHashMap<>();
double scoreGold = score(tree, goldVectors);
DeepTree bestTree = getHighestScoringTree(tree, TRAIN_LAMBDA);
DeepTree goldTree = new DeepTree(tree, goldVectors, scoreGold);
return Pair.makePair(goldTree, bestTree);
}
// Update incompatibles for two clusters that are about to be merged
public void mergeIncompatibles(CorefCluster to, CorefCluster from) {
List<Pair<Pair<Integer, Integer>, Pair<Integer, Integer>>> replacements =
new ArrayList<Pair<Pair<Integer, Integer>, Pair<Integer, Integer>>>();
for (Pair<Integer, Integer> p : incompatibleClusters) {
Integer other = null;
if (p.first == from.clusterID) {
other = p.second;
} else if (p.second == from.clusterID) {
other = p.first;
}
if (other != null && other != to.clusterID) {
int cid1 = Math.min(other, to.clusterID);
int cid2 = Math.max(other, to.clusterID);
replacements.add(Pair.makePair(p, Pair.makePair(cid1, cid2)));
}
}
for (Pair<Pair<Integer, Integer>, Pair<Integer, Integer>> r : replacements) {
incompatibleClusters.remove(r.first.first(), r.first.second());
incompatibleClusters.add(r.second.first(), r.second.second());
}
}
/**
* Parse a CoNLL formatted tree into a SemanticGraph object (along with a list of tokens).
*
* @param conll The CoNLL formatted tree.
* @return A pair of a SemanticGraph and a token list, corresponding to the parse of the sentence
* and to tokens in the sentence.
*/
protected Pair<SemanticGraph, List<CoreLabel>> mkTree(String conll) {
List<CoreLabel> sentence = new ArrayList<>();
SemanticGraph tree = new SemanticGraph();
for (String line : conll.split("\n")) {
if (line.trim().equals("")) {
continue;
}
String[] fields = line.trim().split("\\s+");
int index = Integer.parseInt(fields[0]);
String word = fields[1];
CoreLabel label = IETestUtils.mkWord(word, index);
sentence.add(label);
if (fields[2].equals("0")) {
tree.addRoot(new IndexedWord(label));
} else {
tree.addVertex(new IndexedWord(label));
}
if (fields.length > 4) {
label.setTag(fields[4]);
}
if (fields.length > 5) {
label.setNER(fields[5]);
}
if (fields.length > 6) {
label.setLemma(fields[6]);
}
}
int i = 0;
for (String line : conll.split("\n")) {
if (line.trim().equals("")) {
continue;
}
String[] fields = line.trim().split("\\s+");
int parent = Integer.parseInt(fields[2]);
String reln = fields[3];
if (parent > 0) {
tree.addEdge(
new IndexedWord(sentence.get(parent - 1)),
new IndexedWord(sentence.get(i)),
new GrammaticalRelation(Language.UniversalEnglish, reln, null, null),
1.0,
false);
}
i += 1;
}
return Pair.makePair(tree, sentence);
}
/**
* The core implementation of the search.
*
* @param root The root word to search from. Traditionally, this is the root of the sentence.
* @param 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
* <ol>
* <li>The log probability of the sentence fragment, according to the featurizer and the
* weights
* <li>The features along the path to this fragment. The last element of this is the
* features from the most recent step.
* <li>The sentence fragment. Because it is relatively expensive to compute the resulting
* tree, this is returned as a lazy {@link Supplier}.
* </ol>
*
* @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. Make sure this matches the weights!
* @param actionSpace The action space we are allowed to take. Each action defines a means of
* splitting a clause on a dependency boundary.
*/
protected void search(
// The root to search from
IndexedWord root,
// The output specs
final Predicate<Triple<Double, List<Counter<String>>, Supplier<SentenceFragment>>>
candidateFragments,
// The learning specs
final Classifier<ClauseSplitter.ClauseClassifierLabel, String> classifier,
Map<String, ? extends List<String>> hardCodedSplits,
final Function<Triple<State, Action, State>, Counter<String>> featurizer,
final Collection<Action> actionSpace,
final int maxTicks) {
// (the fringe)
PriorityQueue<Pair<State, List<Counter<String>>>> fringe = new FixedPrioritiesPriorityQueue<>();
// (avoid duplicate work)
Set<IndexedWord> seenWords = new HashSet<>();
State firstState =
new State(null, null, -9000, null, x -> {}, true); // First state is implicitly "done"
fringe.add(Pair.makePair(firstState, new ArrayList<>(0)), -0.0);
int ticks = 0;
while (!fringe.isEmpty()) {
if (++ticks > maxTicks) {
// System.err.println("WARNING! Timed out on search with " + ticks + " ticks");
return;
}
// Useful variables
double logProbSoFar = fringe.getPriority();
assert logProbSoFar <= 0.0;
Pair<State, List<Counter<String>>> lastStatePair = fringe.removeFirst();
State lastState = lastStatePair.first;
List<Counter<String>> featuresSoFar = lastStatePair.second;
IndexedWord rootWord = lastState.edge == null ? root : lastState.edge.getDependent();
// Register thunk
if (lastState.isDone) {
if (!candidateFragments.test(
Triple.makeTriple(
logProbSoFar,
featuresSoFar,
() -> {
SemanticGraph copy = new SemanticGraph(tree);
lastState
.thunk
.andThen(
x -> {
// Add the extra edges back in, if they don't break the tree-ness of the
// extraction
for (IndexedWord newTreeRoot : x.getRoots()) {
if (newTreeRoot != null) { // what a strange thing to have happen...
for (SemanticGraphEdge extraEdge :
extraEdgesByGovernor.get(newTreeRoot)) {
assert Util.isTree(x);
//noinspection unchecked
addSubtree(
x,
newTreeRoot,
extraEdge.getRelation().toString(),
tree,
extraEdge.getDependent(),
tree.getIncomingEdgesSorted(newTreeRoot));
assert Util.isTree(x);
}
}
}
})
.accept(copy);
return new SentenceFragment(copy, assumedTruth, false);
}))) {
break;
}
}
// Find relevant auxilliary terms
SemanticGraphEdge subjOrNull = null;
SemanticGraphEdge objOrNull = null;
for (SemanticGraphEdge auxEdge : tree.outgoingEdgeIterable(rootWord)) {
String relString = auxEdge.getRelation().toString();
if (relString.contains("obj")) {
objOrNull = auxEdge;
} else if (relString.contains("subj")) {
subjOrNull = auxEdge;
}
}
// Iterate over children
// For each outgoing edge...
for (SemanticGraphEdge outgoingEdge : tree.outgoingEdgeIterable(rootWord)) {
// Prohibit indirect speech verbs from splitting off clauses
// (e.g., 'said', 'think')
// This fires if the governor is an indirect speech verb, and the outgoing edge is a ccomp
if (outgoingEdge.getRelation().toString().equals("ccomp")
&& ((outgoingEdge.getGovernor().lemma() != null
&& INDIRECT_SPEECH_LEMMAS.contains(outgoingEdge.getGovernor().lemma()))
|| INDIRECT_SPEECH_LEMMAS.contains(outgoingEdge.getGovernor().word()))) {
continue;
}
// Get some variables
String outgoingEdgeRelation = outgoingEdge.getRelation().toString();
List<String> forcedArcOrder = hardCodedSplits.get(outgoingEdgeRelation);
if (forcedArcOrder == null && outgoingEdgeRelation.contains(":")) {
forcedArcOrder =
hardCodedSplits.get(
outgoingEdgeRelation.substring(0, outgoingEdgeRelation.indexOf(":")) + ":*");
}
boolean doneForcedArc = false;
// For each action...
for (Action action :
(forcedArcOrder == null ? actionSpace : orderActions(actionSpace, forcedArcOrder))) {
// Check the prerequisite
if (!action.prerequisitesMet(tree, outgoingEdge)) {
continue;
}
if (forcedArcOrder != null && doneForcedArc) {
break;
}
// 1. Compute the child state
Optional<State> candidate =
action.applyTo(tree, lastState, outgoingEdge, subjOrNull, objOrNull);
if (candidate.isPresent()) {
double logProbability;
ClauseClassifierLabel bestLabel;
Counter<String> features =
featurizer.apply(Triple.makeTriple(lastState, action, candidate.get()));
if (forcedArcOrder != null && !doneForcedArc) {
logProbability = 0.0;
bestLabel = ClauseClassifierLabel.CLAUSE_SPLIT;
doneForcedArc = true;
} else if (features.containsKey("__undocumented_junit_no_classifier")) {
logProbability = Double.NEGATIVE_INFINITY;
bestLabel = ClauseClassifierLabel.CLAUSE_INTERM;
} else {
Counter<ClauseClassifierLabel> scores = classifier.scoresOf(new RVFDatum<>(features));
if (scores.size() > 0) {
Counters.logNormalizeInPlace(scores);
}
String rel = outgoingEdge.getRelation().toString();
if ("nsubj".equals(rel) || "dobj".equals(rel)) {
scores.remove(
ClauseClassifierLabel.NOT_A_CLAUSE); // Always at least yield on nsubj and dobj
}
logProbability = Counters.max(scores, Double.NEGATIVE_INFINITY);
bestLabel = Counters.argmax(scores, (x, y) -> 0, ClauseClassifierLabel.CLAUSE_SPLIT);
}
if (bestLabel != ClauseClassifierLabel.NOT_A_CLAUSE) {
Pair<State, List<Counter<String>>> childState =
Pair.makePair(
candidate.get().withIsDone(bestLabel),
new ArrayList<Counter<String>>(featuresSoFar) {
{
add(features);
}
});
// 2. Register the child state
if (!seenWords.contains(childState.first.edge.getDependent())) {
// System.err.println(" pushing " + action.signature() + " with " +
// argmax.first.edge);
fringe.add(childState, logProbability);
}
}
}
}
}
seenWords.add(rootWord);
}
// System.err.println("Search finished in " + ticks + " ticks and " + classifierEvals + "
// classifier evaluations.");
}
public List<Pair<String, Double>> selectWeightedKeysWithSampling(
ActiveLearningSelectionCriterion criterion, int numSamples, int seed) {
List<Pair<String, Double>> result = new ArrayList<>();
forceTrack("Sampling Keys");
log("" + numSamples + " to collect");
// Get uncertainty
forceTrack("Computing Uncertainties");
Counter<String> weightCounter = uncertainty(criterion);
assert weightCounter.equals(uncertainty(criterion));
endTrack("Computing Uncertainties");
// Compute some statistics
startTrack("Uncertainty Histogram");
// log(new Histogram(weightCounter, 50).toString()); // removed to make the release easier
// (Histogram isn't in CoreNLP)
endTrack("Uncertainty Histogram");
double totalCount = weightCounter.totalCount();
Random random = new Random(seed);
// Flatten counter
List<String> keys = new LinkedList<>();
List<Double> weights = new LinkedList<>();
List<String> zeroUncertaintyKeys = new LinkedList<>();
for (Pair<String, Double> elem :
Counters.toSortedListWithCounts(
weightCounter,
(o1, o2) -> {
int value = o1.compareTo(o2);
if (value == 0) {
return o1.first.compareTo(o2.first);
} else {
return value;
}
})) {
if (elem.second != 0.0
|| weightCounter.totalCount() == 0.0
|| weightCounter.size() <= numSamples) { // ignore 0 probability weights
keys.add(elem.first);
weights.add(elem.second);
} else {
zeroUncertaintyKeys.add(elem.first);
}
}
// Error check
if (Utils.assertionsEnabled()) {
for (Double elem : weights) {
if (!(elem >= 0 && !Double.isInfinite(elem) && !Double.isNaN(elem))) {
throw new IllegalArgumentException("Invalid weight: " + elem);
}
}
}
// Sample
SAMPLE_ITER:
for (int i = 1; i <= numSamples; ++i) { // For each sample
if (i % 1000 == 0) {
// Debug log
log("sampled " + (i / 1000) + "k keys");
// Recompute total count to mitigate floating point errors
totalCount = 0.0;
for (double val : weights) {
totalCount += val;
}
}
if (weights.size() == 0) {
continue;
}
assert totalCount >= 0.0;
assert weights.size() == keys.size();
double target = random.nextDouble() * totalCount;
Iterator<String> keyIter = keys.iterator();
Iterator<Double> weightIter = weights.iterator();
double runningTotal = 0.0;
while (keyIter.hasNext()) { // For each candidate
String key = keyIter.next();
double weight = weightIter.next();
runningTotal += weight;
if (target <= runningTotal) { // Select that sample
result.add(Pair.makePair(key, weight));
keyIter.remove();
weightIter.remove();
totalCount -= weight;
continue SAMPLE_ITER; // continue sampling
}
}
// We should get here only if the keys list is empty
warn(
"No more uncertain samples left to draw from! (target="
+ target
+ " totalCount="
+ totalCount
+ " size="
+ keys.size());
assert keys.size() == 0;
if (zeroUncertaintyKeys.size() > 0) {
result.add(Pair.makePair(zeroUncertaintyKeys.remove(0), 0.0));
} else {
break;
}
}
endTrack("Sampling Keys");
return result;
}