private SimpleMatrix concatenateContextWords(
SimpleMatrix childVec, IntPair span, List<String> words) {
// TODO: factor out getting the words
SimpleMatrix left =
(span.getSource() < 0)
? dvModel.getStartWordVector()
: dvModel.getWordVector(words.get(span.getSource()));
SimpleMatrix right =
(span.getTarget() >= words.size())
? dvModel.getEndWordVector()
: dvModel.getWordVector(words.get(span.getTarget()));
return NeuralUtils.concatenate(childVec, left, right);
}
/**
* Recreates a Stanford Tree from the StanfordParser annotations and saves all
* non-StanfordParser-Annotations within the scope of the sentence in the label of the best
* fitting node.
*
* <p><strong>CAUTION: </strong><i>This method is intended for the use by CAS Multipliers, which
* create new CASes from this tree. The annotation-spans in the source-CAS will be changed!!!!!!
* You do NOT want to use the source CAS after this method has been called. The
* createStanfordTree()-method does not change the CAS, so use this instead, if the annotations do
* not have to be recovered or accessed in the tree.</i>
*
* <p>TODO: This behavior could be changed by making COPIES of the annotations and changing the
* copied instead of the originals. However, in order to being able to make copies, a dummy CAS
* must be introduced to which the annotations can be copied. When they are recovered, they will
* be copied to the new destination CAS anyway.
*
* @param root the ROOT annotation
* @return an {@link Tree} object representing the syntax structure of the sentence
* @throws CASException if the JCas cannot be accessed.
*/
public static Tree createStanfordTreeWithAnnotations(Annotation root) throws CASException {
JCas aJCas = root.getCAS().getJCas();
// Create tree
Tree tree = createStanfordTree(root);
// Get all non-parser related annotations
// and all tokens (needed for span-calculations later on)
List<Annotation> nonParserAnnotations = new ArrayList<Annotation>();
List<Token> tokens = new ArrayList<Token>();
// Using getCoveredAnnotations instead of iterate, because subiterators did not work in all
// cases
List<Annotation> annosWithinRoot = JCasUtil.selectCovered(aJCas, Annotation.class, root);
for (Annotation curAnno : annosWithinRoot) {
if (!(curAnno instanceof POS)
&& !(curAnno instanceof Constituent)
&& !(curAnno instanceof Dependency)
&& !(curAnno instanceof PennTree)
&& !(curAnno instanceof Lemma)
&& !(curAnno instanceof Token)
&& !(curAnno instanceof DocumentMetaData)) {
nonParserAnnotations.add(curAnno);
} else if (curAnno instanceof Token) {
tokens.add((Token) curAnno);
}
}
// create wrapper for tree and its tokens
TreeWithTokens annoTree = new TreeWithTokens(tree, tokens);
/*
* Add annotations to the best-fitting nodes. The best-fitting node for an annotation is the
* deepest node in the tree that still completely contains the annotation.
*/
for (Annotation curAnno : nonParserAnnotations) {
// get best fitting node
Tree bestFittingNode = annoTree.getBestFit(curAnno);
// Add annotation to node
if (bestFittingNode != null) {
// translate annotation span to a value relative to the
// node-span
IntPair span = annoTree.getSpan(bestFittingNode);
curAnno.setBegin(curAnno.getBegin() - span.getSource());
curAnno.setEnd(curAnno.getEnd() - span.getSource());
// get the collection from the label of the best-fitting node in which we store UIMA
// annotations or create it, if it does not exist
Collection<Annotation> annotations =
((CoreLabel) bestFittingNode.label()).get(UIMAAnnotations.class);
if (annotations == null) {
annotations = new ArrayList<Annotation>();
}
// add annotation + checksum of annotated text to list and write it back to node
// label
annotations.add(curAnno);
((CoreLabel) bestFittingNode.label()).set(UIMAAnnotations.class, annotations);
}
}
return tree;
}