public Object formResult() {
Set brs = new HashSet();
Set urs = new HashSet();
// scan each rule / history pair
int ruleCount = 0;
for (Iterator pairI = rulePairs.keySet().iterator(); pairI.hasNext(); ) {
if (ruleCount % 100 == 0) {
System.err.println("Rules multiplied: " + ruleCount);
}
ruleCount++;
Pair rulePair = (Pair) pairI.next();
Rule baseRule = (Rule) rulePair.first;
String baseLabel = (String) ruleToLabel.get(baseRule);
List history = (List) rulePair.second;
double totalProb = 0;
for (int depth = 1; depth <= HISTORY_DEPTH() && depth <= history.size(); depth++) {
List subHistory = history.subList(0, depth);
double c_label = labelPairs.getCount(new Pair(baseLabel, subHistory));
double c_rule = rulePairs.getCount(new Pair(baseRule, subHistory));
// System.out.println("Multiplying out "+baseRule+" with history "+subHistory);
// System.out.println("Count of "+baseLabel+" with "+subHistory+" is "+c_label);
// System.out.println("Count of "+baseRule+" with "+subHistory+" is "+c_rule );
double prob = (1.0 / HISTORY_DEPTH()) * (c_rule) / (c_label);
totalProb += prob;
for (int childDepth = 0; childDepth <= Math.min(HISTORY_DEPTH() - 1, depth); childDepth++) {
Rule rule = specifyRule(baseRule, subHistory, childDepth);
rule.score = (float) Math.log(totalProb);
// System.out.println("Created "+rule+" with score "+rule.score);
if (rule instanceof UnaryRule) {
urs.add(rule);
} else {
brs.add(rule);
}
}
}
}
System.out.println("Total states: " + stateNumberer.total());
BinaryGrammar bg = new BinaryGrammar(stateNumberer.total());
UnaryGrammar ug = new UnaryGrammar(stateNumberer.total());
for (Iterator brI = brs.iterator(); brI.hasNext(); ) {
BinaryRule br = (BinaryRule) brI.next();
bg.addRule(br);
}
for (Iterator urI = urs.iterator(); urI.hasNext(); ) {
UnaryRule ur = (UnaryRule) urI.next();
ug.addRule(ur);
}
return new Pair(ug, bg);
}
/** Run some sanity checks on the training statistics, to make sure they look valid. */
public void validate() {
for (Map<SentenceKey, EnsembleStatistics> map : impl) {
for (EnsembleStatistics stats : map.values()) {
for (SentenceStatistics component : stats.statisticsForClassifiers) {
assert !Counters.isUniformDistribution(component.relationDistribution, 1e-5);
Counters.normalize(
component.relationDistribution); // TODO(gabor) this shouldn't be necessary
assert (Math.abs(component.relationDistribution.totalCount() - 1.0)) < 1e-5;
}
assert (Math.abs(stats.mean().relationDistribution.totalCount() - 1.0)) < 1e-5;
assert !Counters.isUniformDistribution(stats.mean().relationDistribution, 1e-5);
}
}
}
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);
}
protected void tallyInternalNode(Tree lt, List parents) {
// form base rule
String label = lt.label().value();
Rule baseR = ltToRule(lt);
ruleToLabel.put(baseR, label);
// act on each history depth
for (int depth = 0, maxDepth = Math.min(HISTORY_DEPTH(), parents.size());
depth <= maxDepth;
depth++) {
List history = new ArrayList(parents.subList(0, depth));
// tally each history level / rewrite pair
rulePairs.incrementCount(new Pair(baseR, history), 1);
labelPairs.incrementCount(new Pair(label, history), 1);
}
}
/**
* Get the top few clauses from this searcher, cutting off at the given minimum probability.
*
* @param thresholdProbability The threshold under which to stop returning clauses. This should be
* between 0 and 1.
* @return The resulting {@link edu.stanford.nlp.naturalli.SentenceFragment} objects, representing
* the top clauses of the sentence.
*/
public List<SentenceFragment> topClauses(double thresholdProbability) {
List<SentenceFragment> results = new ArrayList<>();
search(
triple -> {
assert triple.first <= 0.0;
double prob = Math.exp(triple.first);
assert prob <= 1.0;
assert prob >= 0.0;
assert !Double.isNaN(prob);
if (prob >= thresholdProbability) {
SentenceFragment fragment = triple.third.get();
fragment.score = prob;
results.add(fragment);
return true;
} else {
return false;
}
});
return results;
}