protected internal virtual Tree FindSyntacticHead(Mention m, Tree root, IList <CoreLabel> tokens)
{
// mention ends with 's
int endIdx = m.endIndex;
if (m.originalSpan.Count > 0)
{
string lastWord = m.originalSpan[m.originalSpan.Count - 1].Get(typeof(CoreAnnotations.TextAnnotation));
if ((lastWord.Equals("'s") || lastWord.Equals("'")) && m.originalSpan.Count != 1)
{
endIdx--;
}
}
Tree exactMatch = FindTreeWithSpan(root, m.startIndex, endIdx);
//
// found an exact match
//
if (exactMatch != null)
{
return(SafeHead(exactMatch, endIdx));
}
// no exact match found
// in this case, we parse the actual extent of the mention, embedded in a sentence
// context, so as to make the parser work better :-)
if (allowReparsing)
{
int approximateness = 0;
IList <CoreLabel> extentTokens = new List <CoreLabel>();
extentTokens.Add(InitCoreLabel("It"));
extentTokens.Add(InitCoreLabel("was"));
int AddedWords = 2;
for (int i = m.startIndex; i < endIdx; i++)
{
// Add everything except separated dashes! The separated dashes mess with the parser too badly.
CoreLabel label = tokens[i];
if (!"-".Equals(label.Word()))
{
// necessary to copy tokens in case the parser does things like
// put new indices on the tokens
extentTokens.Add((CoreLabel)label.LabelFactory().NewLabel(label));
}
else
{
approximateness++;
}
}
extentTokens.Add(InitCoreLabel("."));
// constrain the parse to the part we're interested in.
// Starting from ADDED_WORDS comes from skipping "It was".
// -1 to exclude the period.
// We now let it be any kind of nominal constituent, since there
// are VP and S ones
ParserConstraint constraint = new ParserConstraint(AddedWords, extentTokens.Count - 1, Pattern.Compile(".*"));
IList <ParserConstraint> constraints = Java.Util.Collections.SingletonList(constraint);
Tree tree = Parse(extentTokens, constraints);
ConvertToCoreLabels(tree);
// now unnecessary, as parser uses CoreLabels?
tree.IndexSpans(m.startIndex - AddedWords);
// remember it has ADDED_WORDS extra words at the beginning
Tree subtree = FindPartialSpan(tree, m.startIndex);
// There was a possible problem that with a crazy parse, extentHead could be one of the added words, not a real word!
// Now we make sure in findPartialSpan that it can't be before the real start, and in safeHead, we disallow something
// passed the right end (that is, just that final period).
Tree extentHead = SafeHead(subtree, endIdx);
System.Diagnostics.Debug.Assert((extentHead != null));
// extentHead is a child in the local extent parse tree. we need to find the corresponding node in the main tree
// Because we deleted dashes, it's index will be >= the index in the extent parse tree
CoreLabel l = (CoreLabel)extentHead.Label();
Tree realHead = FunkyFindLeafWithApproximateSpan(root, l.Value(), l.Get(typeof(CoreAnnotations.BeginIndexAnnotation)), approximateness);
System.Diagnostics.Debug.Assert((realHead != null));
return(realHead);
}
// If reparsing wasn't allowed, try to find a span in the tree
// which happens to have the head
Tree wordMatch = FindTreeWithSmallestSpan(root, m.startIndex, endIdx);
if (wordMatch != null)
{
Tree head = SafeHead(wordMatch, endIdx);
if (head != null)
{
int index = ((CoreLabel)head.Label()).Get(typeof(CoreAnnotations.IndexAnnotation)) - 1;
if (index >= m.startIndex && index < endIdx)
{
return(head);
}
}
}
// If that didn't work, guess that it's the last word
int lastNounIdx = endIdx - 1;
for (int i_1 = m.startIndex; i_1 < m.endIndex; i_1++)
{
if (tokens[i_1].Tag().StartsWith("N"))
{
lastNounIdx = i_1;
}
else
{
if (tokens[i_1].Tag().StartsWith("W"))
{
break;
}
}
}
IList <Tree> leaves = root.GetLeaves();
Tree endLeaf = leaves[lastNounIdx];
return(endLeaf);
}
