public virtual void RestoreOriginalWords(Tree tree)
{
parserQuery.RestoreOriginalWords(tree);
}
public virtual void ProcessResults(IParserQuery pq, Tree goldTree, PrintWriter pwErr, PrintWriter pwOut, PrintWriter pwFileOut, PrintWriter pwStats, TreePrint treePrint)
{
if (pq.SaidMemMessage())
{
saidMemMessage = true;
}
Tree tree;
IList <IHasWord> sentence = pq.OriginalSentence();
try
{
tree = pq.GetBestParse();
}
catch (NoSuchParseException)
{
tree = null;
}
IList <ScoredObject <Tree> > kbestPCFGTrees = null;
if (tree != null && kbestPCFG > 0)
{
kbestPCFGTrees = pq.GetKBestPCFGParses(kbestPCFG);
}
//combo parse goes to pwOut (System.out)
if (op.testOptions.verbose)
{
pwOut.Println("ComboParser best");
Tree ot = tree;
if (ot != null && !op.tlpParams.TreebankLanguagePack().IsStartSymbol(ot.Value()))
{
ot = ot.TreeFactory().NewTreeNode(op.tlpParams.TreebankLanguagePack().StartSymbol(), Java.Util.Collections.SingletonList(ot));
}
treePrint.PrintTree(ot, pwOut);
}
else
{
treePrint.PrintTree(tree, pwOut);
}
// **OUTPUT**
// print various n-best like outputs (including 1-best)
// print various statistics
if (tree != null)
{
if (op.testOptions.printAllBestParses)
{
IList <ScoredObject <Tree> > parses = pq.GetBestPCFGParses();
int sz = parses.Count;
if (sz > 1)
{
pwOut.Println("There were " + sz + " best PCFG parses with score " + parses[0].Score() + '.');
Tree transGoldTree = collinizer.TransformTree(goldTree);
int iii = 0;
foreach (ScoredObject <Tree> sot in parses)
{
iii++;
Tree tb = sot.Object();
Tree tbd = debinarizer.TransformTree(tb);
tbd = subcategoryStripper.TransformTree(tbd);
pq.RestoreOriginalWords(tbd);
pwOut.Println("PCFG Parse #" + iii + " with score " + tbd.Score());
tbd.PennPrint(pwOut);
Tree tbtr = collinizer.TransformTree(tbd);
// pwOut.println("Tree size = " + tbtr.size() + "; depth = " + tbtr.depth());
kGoodLB.Evaluate(tbtr, transGoldTree, pwErr);
}
}
}
else
{
// Huang and Chiang (2006) Algorithm 3 output from the PCFG parser
if (op.testOptions.printPCFGkBest > 0 && op.testOptions.outputkBestEquivocation == null)
{
IList <ScoredObject <Tree> > trees = kbestPCFGTrees.SubList(0, op.testOptions.printPCFGkBest);
Tree transGoldTree = collinizer.TransformTree(goldTree);
int i = 0;
foreach (ScoredObject <Tree> tp in trees)
{
i++;
pwOut.Println("PCFG Parse #" + i + " with score " + tp.Score());
Tree tbd = tp.Object();
tbd.PennPrint(pwOut);
Tree tbtr = collinizer.TransformTree(tbd);
kGoodLB.Evaluate(tbtr, transGoldTree, pwErr);
}
}
else
{
// Chart parser (factored) n-best list
if (op.testOptions.printFactoredKGood > 0 && pq.HasFactoredParse())
{
// DZ: debug n best trees
IList <ScoredObject <Tree> > trees = pq.GetKGoodFactoredParses(op.testOptions.printFactoredKGood);
Tree transGoldTree = collinizer.TransformTree(goldTree);
int ii = 0;
foreach (ScoredObject <Tree> tp in trees)
{
ii++;
pwOut.Println("Factored Parse #" + ii + " with score " + tp.Score());
Tree tbd = tp.Object();
tbd.PennPrint(pwOut);
Tree tbtr = collinizer.TransformTree(tbd);
kGoodLB.Evaluate(tbtr, transGoldTree, pwOut);
}
}
else
{
//1-best output
if (pwFileOut != null)
{
pwFileOut.Println(tree.ToString());
}
}
}
}
//Print the derivational entropy
if (op.testOptions.outputkBestEquivocation != null && op.testOptions.printPCFGkBest > 0)
{
IList <ScoredObject <Tree> > trees = kbestPCFGTrees.SubList(0, op.testOptions.printPCFGkBest);
double[] logScores = new double[trees.Count];
int treeId = 0;
foreach (ScoredObject <Tree> kBestTree in trees)
{
logScores[treeId++] = kBestTree.Score();
}
//Re-normalize
double entropy = 0.0;
double denom = ArrayMath.LogSum(logScores);
foreach (double logScore in logScores)
{
double logPr = logScore - denom;
entropy += System.Math.Exp(logPr) * (logPr / System.Math.Log(2));
}
entropy *= -1;
//Convert to bits
pwStats.Printf("%f\t%d\t%d\n", entropy, trees.Count, sentence.Count);
}
}
// **EVALUATION**
// Perform various evaluations specified by the user
if (tree != null)
{
//Strip subcategories and remove punctuation for evaluation
tree = subcategoryStripper.TransformTree(tree);
Tree treeFact = collinizer.TransformTree(tree);
//Setup the gold tree
if (op.testOptions.verbose)
{
pwOut.Println("Correct parse");
treePrint.PrintTree(goldTree, pwOut);
}
Tree transGoldTree = collinizer.TransformTree(goldTree);
if (transGoldTree != null)
{
transGoldTree = subcategoryStripper.TransformTree(transGoldTree);
}
//Can't do evaluation in these two cases
if (transGoldTree == null)
{
pwErr.Println("Couldn't transform gold tree for evaluation, skipping eval. Gold tree was:");
goldTree.PennPrint(pwErr);
numSkippedEvals++;
return;
}
else
{
if (treeFact == null)
{
pwErr.Println("Couldn't transform hypothesis tree for evaluation, skipping eval. Tree was:");
tree.PennPrint(pwErr);
numSkippedEvals++;
return;
}
else
{
if (treeFact.Yield().Count != transGoldTree.Yield().Count)
{
IList <ILabel> fYield = treeFact.Yield();
IList <ILabel> gYield = transGoldTree.Yield();
pwErr.Println("WARNING: Evaluation could not be performed due to gold/parsed yield mismatch.");
pwErr.Printf(" sizes: gold: %d (transf) %d (orig); parsed: %d (transf) %d (orig).%n", gYield.Count, goldTree.Yield().Count, fYield.Count, tree.Yield().Count);
pwErr.Println(" gold: " + SentenceUtils.ListToString(gYield, true));
pwErr.Println(" pars: " + SentenceUtils.ListToString(fYield, true));
numSkippedEvals++;
return;
}
}
}
if (topKEvals.Count > 0)
{
IList <Tree> transGuesses = new List <Tree>();
int kbest = System.Math.Min(op.testOptions.evalPCFGkBest, kbestPCFGTrees.Count);
foreach (ScoredObject <Tree> guess in kbestPCFGTrees.SubList(0, kbest))
{
transGuesses.Add(collinizer.TransformTree(guess.Object()));
}
foreach (BestOfTopKEval eval in topKEvals)
{
eval.Evaluate(transGuesses, transGoldTree, pwErr);
}
}
//PCFG eval
Tree treePCFG = pq.GetBestPCFGParse();
if (treePCFG != null)
{
Tree treePCFGeval = collinizer.TransformTree(treePCFG);
if (pcfgLB != null)
{
pcfgLB.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgChildSpecific != null)
{
pcfgChildSpecific.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgLA != null)
{
pcfgLA.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgCB != null)
{
pcfgCB.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgDA != null)
{
// Re-index the leaves after Collinization, stripping traces, etc.
treePCFGeval.IndexLeaves(true);
transGoldTree.IndexLeaves(true);
pcfgDA.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgTA != null)
{
pcfgTA.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgLL != null && pq.GetPCFGParser() != null)
{
pcfgLL.RecordScore(pq.GetPCFGParser(), pwErr);
}
if (pcfgRUO != null)
{
pcfgRUO.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgCUO != null)
{
pcfgCUO.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
if (pcfgCatE != null)
{
pcfgCatE.Evaluate(treePCFGeval, transGoldTree, pwErr);
}
}
//Dependency eval
// todo: is treeDep really useful here, or should we really use depDAEval tree (debinarized) throughout? We use it for parse, and it sure seems like we could use it for tag eval, but maybe not factDA?
Tree treeDep = pq.GetBestDependencyParse(false);
if (treeDep != null)
{
Tree goldTreeB = binarizerOnly.TransformTree(goldTree);
Tree goldTreeEval = goldTree.DeepCopy();
goldTreeEval.IndexLeaves(true);
goldTreeEval.PercolateHeads(op.Langpack().HeadFinder());
Tree depDAEval = pq.GetBestDependencyParse(true);
depDAEval.IndexLeaves(true);
depDAEval.PercolateHeadIndices();
if (depDA != null)
{
depDA.Evaluate(depDAEval, goldTreeEval, pwErr);
}
if (depTA != null)
{
Tree undoneTree = debinarizer.TransformTree(treeDep);
undoneTree = subcategoryStripper.TransformTree(undoneTree);
pq.RestoreOriginalWords(undoneTree);
// pwErr.println("subcategoryStripped tree: " + undoneTree.toStructureDebugString());
depTA.Evaluate(undoneTree, goldTree, pwErr);
}
if (depLL != null && pq.GetDependencyParser() != null)
{
depLL.RecordScore(pq.GetDependencyParser(), pwErr);
}
Tree factTreeB;
if (pq.HasFactoredParse())
{
factTreeB = pq.GetBestFactoredParse();
}
else
{
factTreeB = treeDep;
}
if (factDA != null)
{
factDA.Evaluate(factTreeB, goldTreeB, pwErr);
}
}
//Factored parser (1best) eval
if (factLB != null)
{
factLB.Evaluate(treeFact, transGoldTree, pwErr);
}
if (factChildSpecific != null)
{
factChildSpecific.Evaluate(treeFact, transGoldTree, pwErr);
}
if (factLA != null)
{
factLA.Evaluate(treeFact, transGoldTree, pwErr);
}
if (factTA != null)
{
factTA.Evaluate(tree, boundaryRemover.TransformTree(goldTree), pwErr);
}
if (factLL != null && pq.GetFactoredParser() != null)
{
factLL.RecordScore(pq.GetFactoredParser(), pwErr);
}
if (factCB != null)
{
factCB.Evaluate(treeFact, transGoldTree, pwErr);
}
foreach (IEval eval_1 in evals)
{
eval_1.Evaluate(treeFact, transGoldTree, pwErr);
}
if (parserQueryEvals != null)
{
foreach (IParserQueryEval eval in parserQueryEvals)
{
eval_1.Evaluate(pq, transGoldTree, pwErr);
}
}
if (op.testOptions.evalb)
{
// empty out scores just in case
NanScores(tree);
EvalbFormatWriter.WriteEVALBline(treeFact, transGoldTree);
}
}
pwErr.Println();
}