// if survive, return true
// best-deduction is always kept
private boolean pruningEdge(HyperEdge dt, HGNode parent) {
/**
* TODO: theoretically, if an item is get called, then its best deduction should always be kept
* even just by the threshold-checling. In reality, due to precision of Double, the
* threshold-checking may not be perfect
*/
if (dt != parent.bestHyperedge) { // best deduction should always survive if the Item is get
// called
// ### prune?
if (shouldPruneHyperedge(dt, parent)) {
return false; // early stop
}
}
// ### still survive, recursive call all my ant-items
if (null != dt.getTailNodes()) {
for (HGNode ant_it : dt.getTailNodes()) {
pruningNode(ant_it); // recursive call on each ant item, note: the ant_it will not be pruned
// as I need it
}
}
// ### if get to here, then survive; remember: if I survive, then my upper-item must survive
numSurvivedEdges++;
return true; // survive
}
private boolean shouldPruneHyperedge(HyperEdge dt, HGNode parent) {
// ### get merit
double postLogProb = getEdgeUnormalizedPosteriorLogProb(dt, parent);
if (dt.getRule() != null
&& dt.getRule().getOwner() == glueGrammarOwner
&& dt.getRule().getArity() == 2) { // specicial rule: S->S X
// TODO
return (postLogProb - this.bestLogProb < THRESHOLD_GLUE);
} else {
return (postLogProb - this.bestLogProb < THRESHOLD_GENERAL);
}
}
private static int[] getNewSource(boolean isGlue, HyperEdge edge) {
BilingualRule rule = (BilingualRule) edge.getRule();
int[] english = rule.getEnglish();
// if this is a unary abstract rule, just return null
// TODO: except glue rules!
if (english.length == 1 && english[0] < 0 && !isGlue) return null;
int currNT = 1;
int[] result = new int[english.length];
for (int i = 0; i < english.length; i++) {
int curr = english[i];
if (!Vocabulary.nt(curr)) {
result[i] = curr;
} else {
int index = -curr - 1;
int label = getLabelWithSpan(edge.getAntNodes().get(index));
result[i] = label * 2 - currNT;
currNT++;
}
}
// System.err.printf("source: %s\n", result);
return result;
}
private BilingualRule getRuleWithSpans(HyperEdge edge, HGNode head) {
Rule edgeRule = edge.getRule();
// System.err.printf("EdgeRule: %s\n", edgeRule);
if (!(edgeRule instanceof BilingualRule)) {
// System.err.println("edge rule is not a bilingual rule");
return null;
}
int headLabel = getLabelWithSpan(head);
// System.err.printf("Head label: %s\n", headLabel);
// if (edge.getAntNodes() != null) {
// for (HGNode n : edge.getAntNodes())
// System.err.printf("> %s\n", getLabelWithSpan(n));
// }
int[] source = getNewSource(nodeHasGoalSymbol(head), edge);
// if this would be unary abstract, getNewSource will be null
if (source == null) return null;
int[] target = getNewTargetFromSource(source);
BilingualRule result =
new BilingualRule(
headLabel, source, target, edgeRule.getFeatureScores(), edgeRule.getArity());
// System.err.printf("new rule is %s\n", result);
return result;
}
public double finalTransitionLogP(HyperEdge edge, int spanStart, int spanEnd, int sentID) {
return finalTransitionLogP(
edge.getAntNodes().get(0), spanStart, spanEnd, edge.getSourcePath(), sentID);
}
public double transitionLogP(HyperEdge edge, int spanStart, int spanEnd, int sentID) {
return transitionLogP(
edge.getRule(), edge.getAntNodes(), spanStart, spanEnd, edge.getSourcePath(), sentID);
}