@SuppressWarnings({"unchecked"})
@Override
protected void fillFeatures(
Pair<Mention, ClusteredMention> input,
Counter<Feature> inFeatures,
Boolean output,
Counter<Feature> outFeatures) {
// --Input Features
for (Object o : ACTIVE_FEATURES) {
if (o instanceof Class) {
// (case: singleton feature)
Option<Double> count = new Option<Double>(1.0);
Feature feat = feature((Class) o, input, count);
if (count.get() > 0.0) {
inFeatures.incrementCount(feat, count.get());
}
} else if (o instanceof Pair) {
// (case: pair of features)
Pair<Class, Class> pair = (Pair<Class, Class>) o;
Option<Double> countA = new Option<Double>(1.0);
Option<Double> countB = new Option<Double>(1.0);
Feature featA = feature(pair.getFirst(), input, countA);
Feature featB = feature(pair.getSecond(), input, countB);
if (countA.get() * countB.get() > 0.0) {
inFeatures.incrementCount(
new Feature.PairFeature(featA, featB), countA.get() * countB.get());
}
}
}
// --Output Features
if (output != null) {
outFeatures.incrementCount(new Feature.CoreferentIndicator(output), 1.0);
}
}
public List<ClusteredMention> runCoreference(Document doc) {
// --Overhead
startTrack("Testing " + doc.id);
// (variables)
List<ClusteredMention> rtn = new ArrayList<ClusteredMention>(doc.getMentions().size());
List<Mention> mentions = doc.getMentions();
int singletons = 0;
// --Run Classifier
for (int i = 0; i < mentions.size(); i++) {
// (variables)
Mention onPrix = mentions.get(i);
int coreferentWith = -1;
// (get mention it is coreferent with)
for (int j = i - 1; j >= 0; j--) {
ClusteredMention cand = rtn.get(j);
boolean coreferent =
classifier.classOf(
new RVFDatum<Boolean, Feature>(extractor.extractFeatures(Pair.make(onPrix, cand))));
if (coreferent) {
coreferentWith = j;
break;
}
}
if (coreferentWith < 0) {
singletons += 1;
rtn.add(onPrix.markSingleton());
} else {
// log("Mention " + onPrix + " coreferent with " + mentions.get(coreferentWith));
rtn.add(onPrix.markCoreferent(rtn.get(coreferentWith)));
}
}
// log("" + singletons + " singletons");
// --Return
endTrack("Testing " + doc.id);
return rtn;
}
private <E> Feature feature(
Class<E> clazz, Pair<Mention, ClusteredMention> input, Option<Double> count) {
// --Variables
Mention onPrix =
input.getFirst(); // the first mention (referred to as m_i in the handout)
Mention candidate =
input.getSecond().mention; // the second mention (referred to as m_j in the handout)
Entity candidateCluster =
input.getSecond().entity; // the cluster containing the second mention
// --Features:w
if (clazz.equals(Feature.ExactMatch.class)) {
// (exact string match)
return new Feature.ExactMatch(onPrix.gloss().equals(candidate.gloss()));
} else if (clazz.equals(Feature.SentenceDist.class)) {
return new Feature.SentenceDist(
Math.abs(
onPrix.doc.indexOfMention(onPrix) - candidate.doc.indexOfMention(candidate)));
} else if (clazz.equals(Feature.MentionDist.class)) {
return new Feature.MentionDist(
Math.abs(
onPrix.doc.indexOfSentence(onPrix.sentence)
- candidate.doc.indexOfSentence(candidate.sentence)));
} else if (clazz.equals(Feature.EitherHeadWordPronoun.class)) {
return new Feature.EitherHeadWordPronoun(
Pronoun.isSomePronoun(onPrix.gloss()) || Pronoun.isSomePronoun(candidate.gloss()));
} else if (clazz.equals(Feature.CandidateNERTag.class)) {
return new Feature.CandidateNERTag(candidate.headToken().nerTag());
} else if (clazz.equals(Feature.CandidateSpeaker.class)) {
return new Feature.CandidateSpeaker(candidate.headToken().speaker());
} else if (clazz.equals(Feature.FixedSpeaker.class)) {
return new Feature.FixedSpeaker(onPrix.headToken().speaker());
} else if (clazz.equals(Feature.HeadWordMatch.class)) {
return new Feature.HeadWordMatch(onPrix.equals(candidate.headWord()));
} else if (clazz.equals(Feature.HeadWordLemmaMatch.class)) {
return new Feature.HeadWordLemmaMatch(
onPrix.headToken().lemma().equals(candidate.headToken().lemma()));
} else if (clazz.equals(Feature.FixedNERTag.class)) {
return new Feature.FixedNERTag(onPrix.headToken().nerTag());
} else if (clazz.equals(Feature.SpeakerMatch.class)) {
return new Feature.SpeakerMatch(
candidate.headToken().speaker().equals(onPrix.headToken().speaker()));
} else if (clazz.equals(Feature.NERTagMatch.class)) {
return new Feature.NERTagMatch(
candidate.headToken().nerTag().equals(onPrix.headToken().nerTag()));
} else if (clazz.equals(Feature.CandidatePOSTag.class)) {
return new Feature.CandidatePOSTag(candidate.headToken().posTag());
} else if (clazz.equals(Feature.FixedPOSTag.class)) {
return new Feature.FixedPOSTag(onPrix.headToken().posTag());
} else if (clazz.equals(Feature.GenderMatch.class)) {
Pair<Boolean, Boolean> match = Util.haveGenderAndAreSameGender(onPrix, candidate);
boolean finalMatch = (!match.getFirst() || match.getSecond());
return new Feature.GenderMatch(finalMatch);
} else if (clazz.equals(Feature.NumberMatch.class)) {
Pair<Boolean, Boolean> match = Util.haveNumberAndAreSameNumber(onPrix, candidate);
boolean finalMatch = (!match.getFirst() || match.getSecond());
return new Feature.NumberMatch(finalMatch);
}
// } else if(clazz.equals(Feature.NewFeature.class) {
/*
* TODO: Add features to return for specific classes. Implement calculating values of features here.
*/
else {
throw new IllegalArgumentException("Unregistered feature: " + clazz);
}
}
public void train(Collection<Pair<Document, List<Entity>>> trainingData) {
startTrack("Training");
// --Variables
RVFDataset<Boolean, Feature> dataset = new RVFDataset<Boolean, Feature>();
LinearClassifierFactory<Boolean, Feature> fact =
new LinearClassifierFactory<Boolean, Feature>();
// --Feature Extraction
startTrack("Feature Extraction");
for (Pair<Document, List<Entity>> datum : trainingData) {
// (document variables)
Document doc = datum.getFirst();
List<Entity> goldClusters = datum.getSecond();
List<Mention> mentions = doc.getMentions();
Map<Mention, Entity> goldEntities = Entity.mentionToEntityMap(goldClusters);
startTrack("Document " + doc.id);
// (for each mention...)
for (int i = 0; i < mentions.size(); i++) {
// (get the mention and its cluster)
Mention onPrix = mentions.get(i);
Entity source = goldEntities.get(onPrix);
if (source == null) {
throw new IllegalArgumentException("Mention has no gold entity: " + onPrix);
}
// (for each previous mention...)
int oldSize = dataset.size();
for (int j = i - 1; j >= 0; j--) {
// (get previous mention and its cluster)
Mention cand = mentions.get(j);
Entity target = goldEntities.get(cand);
if (target == null) {
throw new IllegalArgumentException("Mention has no gold entity: " + cand);
}
// (extract features)
Counter<Feature> feats =
extractor.extractFeatures(Pair.make(onPrix, cand.markCoreferent(target)));
// (add datum)
dataset.add(new RVFDatum<Boolean, Feature>(feats, target == source));
// (stop if
if (target == source) {
break;
}
}
// logf("Mention %s (%d datums)", onPrix.toString(), dataset.size() - oldSize);
}
endTrack("Document " + doc.id);
}
endTrack("Feature Extraction");
// --Train Classifier
startTrack("Minimizer");
this.classifier = fact.trainClassifier(dataset);
endTrack("Minimizer");
// --Dump Weights
startTrack("Features");
// (get labels to print)
Set<Boolean> labels = new HashSet<Boolean>();
labels.add(true);
// (print features)
for (Triple<Feature, Boolean, Double> featureInfo :
this.classifier.getTopFeatures(labels, 0.0, true, 100, true)) {
Feature feature = featureInfo.first();
Boolean label = featureInfo.second();
Double magnitude = featureInfo.third();
// log(FORCE,new DecimalFormat("0.000").format(magnitude) + " [" + label + "] " + feature);
}
end_Track("Features");
endTrack("Training");
}
public void train(List<SentencePair> trainingPairs) {
sourceTargetCounts = new CounterMap<String, String>();
sourceTargetDistortions = new CounterMap<Pair<Integer, Integer>, Pair<Integer, Integer>>();
for (SentencePair pair : trainingPairs) {
List<String> sourceSentence = pair.getSourceWords();
List<String> targetSentence = pair.getTargetWords();
targetSentence.add(WordAligner.NULL_WORD);
int m = sourceSentence.size();
int l = targetSentence.size();
for (int i = 0; i < m; i++) {
String sourceWord = sourceSentence.get(i);
for (int j = 0; j < l; j++) {
String targetWord = targetSentence.get(j);
sourceTargetCounts.setCount(sourceWord, targetWord, 1.0);
Pair<Integer, Integer> lmPair = new Pair<Integer, Integer>(l, m);
Pair<Integer, Integer> jiPair = new Pair<Integer, Integer>(j, i);
sourceTargetDistortions.setCount(jiPair, lmPair, 1.0);
}
}
}
// Use Model 1 to train params
double delta = Double.POSITIVE_INFINITY;
for (int i = 0; i < MAX_ITERS && delta > CONVERGENCE; i++) {
CounterMap<String, String> tempSourceTargetCounts = new CounterMap<String, String>();
Counter<String> targetCounts = new Counter<String>();
delta = 0.0;
for (SentencePair pair : trainingPairs) {
List<String> sourceSentence = pair.getSourceWords();
List<String> targetSentence = pair.getTargetWords();
Counter<String> sourceTotals = new Counter<String>();
for (String sourceWord : sourceSentence) {
for (String targetWord : targetSentence) {
sourceTotals.incrementCount(
sourceWord, sourceTargetCounts.getCount(sourceWord, targetWord));
}
}
for (String sourceWord : sourceSentence) {
for (String targetWord : targetSentence) {
double transProb = sourceTargetCounts.getCount(sourceWord, targetWord);
double sourceTotal = sourceTotals.getCount(sourceWord);
tempSourceTargetCounts.incrementCount(sourceWord, targetWord, transProb / sourceTotal);
targetCounts.incrementCount(targetWord, transProb / sourceTotal);
}
}
}
// update t(s|t) values
for (String sourceWord : tempSourceTargetCounts.keySet()) {
for (String targetWord : tempSourceTargetCounts.getCounter(sourceWord).keySet()) {
double oldProb = sourceTargetCounts.getCount(sourceWord, targetWord);
double newProb =
tempSourceTargetCounts.getCount(sourceWord, targetWord)
/ targetCounts.getCount(targetWord);
sourceTargetCounts.setCount(sourceWord, targetWord, newProb);
delta += Math.pow(oldProb - newProb, 2.0);
}
}
delta /= sourceTargetCounts.totalSize();
}
// Maximizing for ibm model 2
delta = Double.POSITIVE_INFINITY;
for (int iter = 0; iter < MAX_ITERS && delta > CONVERGENCE; iter++) {
CounterMap<String, String> tempSourceTargetCounts = new CounterMap<String, String>();
CounterMap<Pair<Integer, Integer>, Pair<Integer, Integer>> tempSourceTargetDistortions =
new CounterMap<Pair<Integer, Integer>, Pair<Integer, Integer>>();
Counter<String> targetCounts = new Counter<String>();
CounterMap<Pair<Integer, Integer>, Integer> targetDistorts =
new CounterMap<Pair<Integer, Integer>, Integer>();
delta = 0.0;
for (SentencePair pair : trainingPairs) {
List<String> sourceSentence = pair.getSourceWords();
List<String> targetSentence = pair.getTargetWords();
CounterMap<Pair<Integer, Integer>, Integer> distortSourceTotals =
new CounterMap<Pair<Integer, Integer>, Integer>();
Pair<Integer, Integer> lmPair =
new Pair<Integer, Integer>(targetSentence.size(), sourceSentence.size());
for (int i = 0; i < sourceSentence.size(); i++) {
String sourceWord = sourceSentence.get(i);
for (int j = 0; j < targetSentence.size(); j++) {
String targetWord = targetSentence.get(j);
Pair<Integer, Integer> jiPair = new Pair<Integer, Integer>(j, i);
double currTransProb = sourceTargetCounts.getCount(sourceWord, targetWord);
double currAlignProb = sourceTargetDistortions.getCount(jiPair, lmPair);
distortSourceTotals.incrementCount(lmPair, i, currTransProb * currAlignProb);
}
}
for (int i = 0; i < sourceSentence.size(); i++) {
String sourceWord = sourceSentence.get(i);
double distortTransSourceTotal = distortSourceTotals.getCount(lmPair, i);
for (int j = 0; j < targetSentence.size(); j++) {
String targetWord = targetSentence.get(j);
Pair<Integer, Integer> jiPair = new Pair<Integer, Integer>(j, i);
double transProb = sourceTargetCounts.getCount(sourceWord, targetWord);
double distortProb = sourceTargetDistortions.getCount(jiPair, lmPair);
double update =
(transProb * distortProb) / (distortTransSourceTotal); // q(j|ilm)t(f|e)/totals
tempSourceTargetCounts.incrementCount(sourceWord, targetWord, update);
tempSourceTargetDistortions.incrementCount(jiPair, lmPair, update);
targetCounts.incrementCount(targetWord, update);
targetDistorts.incrementCount(lmPair, i, update);
}
}
}
// update t(s|t) values
double delta_trans = 0.0;
for (String sourceWord : tempSourceTargetCounts.keySet()) {
for (String targetWord : tempSourceTargetCounts.getCounter(sourceWord).keySet()) {
double oldProb = sourceTargetCounts.getCount(sourceWord, targetWord);
double newProb =
tempSourceTargetCounts.getCount(sourceWord, targetWord)
/ targetCounts.getCount(targetWord);
sourceTargetCounts.setCount(sourceWord, targetWord, newProb);
delta += Math.pow(oldProb - newProb, 2.0);
}
}
// update q(j|ilm) values
double delta_dist = 0.0;
for (Pair<Integer, Integer> jiPair : tempSourceTargetDistortions.keySet()) {
for (Pair<Integer, Integer> lmPair :
tempSourceTargetDistortions.getCounter(jiPair).keySet()) {
double oldProb = sourceTargetDistortions.getCount(jiPair, lmPair);
double tempAlignProb = tempSourceTargetDistortions.getCount(jiPair, lmPair);
double tempTargetDist = targetDistorts.getCount(lmPair, jiPair.getSecond());
double newProb = tempAlignProb / tempTargetDist;
sourceTargetDistortions.setCount(jiPair, lmPair, newProb);
delta_dist += Math.pow(oldProb - newProb, 2.0);
}
}
delta =
(delta_trans / sourceTargetCounts.totalSize()
+ delta_dist / sourceTargetDistortions.totalSize())
/ 2.0;
}
}