/**
* If this tree contains max-marginals, recover the best parse subtree for a given symbol with the
* specified span.
*/
public CfgParseTree getBestParseTreeWithSpan(Object root, int spanStart, int spanEnd) {
Preconditions.checkState(!sumProduct);
Assignment rootAssignment = parentVar.outcomeArrayToAssignment(root);
int rootNonterminalNum = parentVar.assignmentToIntArray(rootAssignment)[0];
double prob =
insideChart[spanStart][spanEnd][rootNonterminalNum]
* outsideChart[spanStart][spanEnd][rootNonterminalNum];
if (prob == 0.0) {
return null;
}
int splitInd = splitBackpointers[spanStart][spanEnd][rootNonterminalNum];
if (splitInd < 0) {
long terminalKey = backpointers[spanStart][spanEnd][rootNonterminalNum];
int positiveSplitInd = (-1 * splitInd) - 1;
int terminalSpanStart = positiveSplitInd / numTerminals;
int terminalSpanEnd = positiveSplitInd % numTerminals;
// This is a really sucky way to transform the keys back to objects.
VariableNumMap vars = parentVar.union(ruleTypeVar);
int[] dimKey = TableFactor.zero(vars).getWeights().keyNumToDimKey(terminalKey);
Assignment a = vars.intArrayToAssignment(dimKey);
Object ruleType = a.getValue(ruleTypeVar.getOnlyVariableNum());
List<Object> terminalList = Lists.newArrayList();
terminalList.addAll(terminals.subList(terminalSpanStart, terminalSpanEnd + 1));
return new CfgParseTree(root, ruleType, terminalList, prob, spanStart, spanEnd);
} else {
long binaryRuleKey = backpointers[spanStart][spanEnd][rootNonterminalNum];
int[] binaryRuleComponents =
binaryRuleDistribution.coerceToDiscrete().getWeights().keyNumToDimKey(binaryRuleKey);
Assignment best = binaryRuleDistribution.getVars().intArrayToAssignment(binaryRuleComponents);
Object leftRoot = best.getValue(leftVar.getOnlyVariableNum());
Object rightRoot = best.getValue(rightVar.getOnlyVariableNum());
Object ruleType = best.getValue(ruleTypeVar.getOnlyVariableNum());
Preconditions.checkArgument(
spanStart + splitInd != spanEnd,
"CFG parse decoding error: %s %s %s",
spanStart,
spanEnd,
splitInd);
CfgParseTree leftTree = getBestParseTreeWithSpan(leftRoot, spanStart, spanStart + splitInd);
CfgParseTree rightTree =
getBestParseTreeWithSpan(rightRoot, spanStart + splitInd + 1, spanEnd);
Preconditions.checkState(leftTree != null);
Preconditions.checkState(rightTree != null);
return new CfgParseTree(root, ruleType, leftTree, rightTree, prob);
}
}
/**
* Update an entry of the inside chart with a new production. Depending on the type of the chart,
* this performs either a sum or max over productions of the same type in the same entry.
*/
public void updateInsideEntry(
int spanStart, int spanEnd, int splitInd, double[] values, Factor binaryRuleProbabilities) {
Preconditions.checkArgument(binaryRuleProbabilities.getVars().size() == 4);
if (sumProduct) {
updateEntrySumProduct(
insideChart[spanStart][spanEnd],
values,
binaryRuleProbabilities.coerceToDiscrete().getWeights(),
parentVar.getOnlyVariableNum());
} else {
Tensor weights = binaryRuleProbabilities.coerceToDiscrete().getWeights();
updateInsideEntryMaxProduct(spanStart, spanEnd, values, weights, splitInd);
}
}
/**
* Update an entry of the inside chart with a new production. Depending on the type of the chart,
* this performs either a sum or max over productions of the same type in the same entry.
*/
public void updateInsideEntryTerminal(
int spanStart, int spanEnd, int terminalSpanStart, int terminalSpanEnd, Factor factor) {
Preconditions.checkArgument(factor.getVars().size() == 2);
// The first entry initializes the chart at this span.
if (sumProduct) {
updateEntrySumProduct(
insideChart[spanStart][spanEnd],
factor.coerceToDiscrete().getWeights().getValues(),
factor.coerceToDiscrete().getWeights(),
parentVar.getOnlyVariableNum());
} else {
Tensor weights = factor.coerceToDiscrete().getWeights();
// Negative split indexes are used to represent terminal rules.
int splitInd = -1 * (terminalSpanStart * numTerminals + terminalSpanEnd + 1);
updateInsideEntryMaxProduct(spanStart, spanEnd, weights.getValues(), weights, splitInd);
}
}
/** Compute the expected *unnormalized* probability of every rule. */
public Factor getBinaryRuleExpectations() {
Tensor binaryRuleWeights = binaryRuleDistribution.coerceToDiscrete().getWeights();
SparseTensor tensor = SparseTensor.copyRemovingZeros(binaryRuleWeights, binaryRuleExpectations);
return new TableFactor(binaryRuleDistribution.getVars(), tensor);
}
