private static boolean isLeaf(Constituent constituent) {
return (constituent.getChildren() == null || constituent.getChildren().size() == 0);
}
public static Tree createStanfordTree(Annotation root, TreeFactory tFact) {
JCas aJCas;
try {
aJCas = root.getCAS().getJCas();
} catch (CASException e) {
throw new IllegalStateException("Unable to get JCas from JCas wrapper");
}
// define the new (root) node
Tree rootNode;
// before we can create a node, we must check if we have any children (we have to know
// whether to create a node or a leaf - not very dynamic)
if (root instanceof Constituent && !isLeaf((Constituent) root)) {
Constituent node = (Constituent) root;
List<Tree> childNodes = new ArrayList<Tree>();
// get childNodes from child annotations
FSArray children = node.getChildren();
for (int i = 0; i < children.size(); i++) {
childNodes.add(createStanfordTree(node.getChildren(i), tFact));
}
// now create the node with its children
rootNode = tFact.newTreeNode(node.getConstituentType(), childNodes);
} else {
// Handle leaf annotations
// Leafs are always Token-annotations
// We also have to insert a Preterminal node with the value of the
// POS-Annotation on the token
// because the POS is not directly stored within the treee
Token wordAnnotation = (Token) root;
// create leaf-node for the tree
Tree wordNode = tFact.newLeaf(wordAnnotation.getCoveredText());
// create information about preceding and trailing whitespaces in the leaf node
StringBuilder preWhitespaces = new StringBuilder();
StringBuilder trailWhitespaces = new StringBuilder();
List<Token> precedingTokenList = selectPreceding(aJCas, Token.class, wordAnnotation, 1);
List<Token> followingTokenList = selectFollowing(aJCas, Token.class, wordAnnotation, 1);
if (precedingTokenList.size() > 0) {
Token precedingToken = precedingTokenList.get(0);
int precedingWhitespaces = wordAnnotation.getBegin() - precedingToken.getEnd();
for (int i = 0; i < precedingWhitespaces; i++) {
preWhitespaces.append(" ");
}
}
if (followingTokenList.size() > 0) {
Token followingToken = followingTokenList.get(0);
int trailingWhitespaces = followingToken.getBegin() - wordAnnotation.getEnd();
for (int i = 0; i < trailingWhitespaces; i++) {
trailWhitespaces.append(" ");
}
}
// write whitespace information as CoreAnnotation.BeforeAnnotation and
// CoreAnnotation.AfterAnnotation to the node add annotation to list and write back to
// node label
((CoreLabel) wordNode.label())
.set(CoreAnnotations.BeforeAnnotation.class, preWhitespaces.toString());
((CoreLabel) wordNode.label())
.set(CoreAnnotations.AfterAnnotation.class, trailWhitespaces.toString());
// get POS-annotation
// get the token that is covered by the POS
List<POS> coveredPos = JCasUtil.selectCovered(aJCas, POS.class, wordAnnotation);
// the POS should only cover one token
assert coveredPos.size() == 1;
POS pos = coveredPos.get(0);
// create POS-Node in the tree and attach word-node to it
rootNode = tFact.newTreeNode(pos.getPosValue(), Arrays.asList((new Tree[] {wordNode})));
}
return rootNode;
}
public Set<Feature> extract(JCas jcas) {
double nrOfNPs = 0.0;
double nrOfVPs = 0.0;
double nrOfPPs = 0.0;
int nrOfSbars = 0;
int nrOfVerbphrases = 0;
int nrOfComplexNominals = 0;
double nrOfClauses = 0.0;
int nrOfDependentClauses = 0;
double nrOfTunits = 0.0;
int nrOfComplexTunits = 0;
int nrOfCoords = 0;
int lengthSumNPs = 0;
int lengthSumVPs = 0;
int lengthSumPPs = 0;
int lengthSumClauses = 0;
int lengthSumTunits = 0;
int parseTreeDepthSum = 0;
Set<Feature> featSet = new HashSet<Feature>();
double nrOfSentences = JCasUtil.select(jcas, Sentence.class).size() * 1.0;
for (Sentence s : JCasUtil.select(jcas, Sentence.class)) {
parseTreeDepthSum += ParsePatternUtils.getParseDepth(s);
for (Constituent c : JCasUtil.selectCovered(Constituent.class, s)) {
if (c instanceof NP) {
nrOfNPs++;
lengthSumNPs += c.getCoveredText().length();
} else if (c instanceof VP) {
nrOfVPs++;
lengthSumVPs += c.getCoveredText().length();
} else if (c instanceof PP) {
nrOfPPs++;
lengthSumPPs += c.getCoveredText().length();
} else if (c instanceof SBAR) {
nrOfSbars++;
if (ParsePatternUtils.isDependentClause(c)) {
nrOfDependentClauses++;
}
} else if (ParsePatternUtils.isClause(c)) {
nrOfClauses++;
lengthSumClauses += c.getCoveredText().length();
}
if (ParsePatternUtils.isTunit(c)) {
nrOfTunits++;
lengthSumTunits += c.getCoveredText().length();
if (ParsePatternUtils.isComplexTunit(c)) {
nrOfComplexTunits++;
}
}
if (ParsePatternUtils.isCoordinate(c)) {
nrOfCoords++;
}
if (ParsePatternUtils.isComplexNominal(c)) {
nrOfComplexNominals++;
}
if (ParsePatternUtils.isVerbPhrase(c)) {
nrOfVerbphrases++;
}
}
}
// avoid division by zero, there should be at least one sentence in the cas
nrOfSentences = Math.max(1, nrOfSentences);
featSet.addAll(Arrays.asList(new Feature(NPS_PER_SENTENCE, nrOfNPs / nrOfSentences)));
featSet.addAll(Arrays.asList(new Feature(VPS_PER_SENTENCE, nrOfVPs / nrOfSentences)));
featSet.addAll(Arrays.asList(new Feature(PPS_PER_SENTENCE, nrOfPPs / nrOfSentences)));
featSet.addAll(Arrays.asList(new Feature(SBARS_PER_SENTENCE, nrOfSbars / nrOfSentences)));
featSet.addAll(Arrays.asList(new Feature(CLAUSES_PER_SENTENCE, nrOfClauses / nrOfSentences)));
featSet.addAll(
Arrays.asList(new Feature(DEP_CLAUSES_PER_SENTENCE, nrOfDependentClauses / nrOfSentences)));
featSet.addAll(Arrays.asList(new Feature(TUNITS_PER_SENTENCE, nrOfTunits / nrOfSentences)));
featSet.addAll(
Arrays.asList(new Feature(COMPLEX_TUNITS_PER_SENTENCE, nrOfComplexTunits / nrOfSentences)));
featSet.addAll(Arrays.asList(new Feature(COORDS_PER_SENTENCE, nrOfCoords / nrOfSentences)));
// avoid division by 0,
// if we don't have any NPs, the lengthSum is 0, division by 1 will yield 0 as average
// length
nrOfNPs = Math.max(1, nrOfNPs);
nrOfVPs = Math.max(1, nrOfVPs);
nrOfPPs = Math.max(1, nrOfPPs);
nrOfTunits = Math.max(1, nrOfTunits);
featSet.addAll(Arrays.asList(new Feature(AVG_NP_LENGTH, lengthSumNPs / nrOfNPs)));
featSet.addAll(Arrays.asList(new Feature(AVG_VP_LENGTH, lengthSumVPs / nrOfVPs)));
featSet.addAll(Arrays.asList(new Feature(AVG_PP_LENGTH, lengthSumPPs / nrOfPPs)));
featSet.addAll(Arrays.asList(new Feature(AVG_TUNIT_LENGTH, lengthSumTunits / nrOfTunits)));
featSet.addAll(Arrays.asList(new Feature(AVG_TREE_DEPTH, parseTreeDepthSum / nrOfSentences)));
featSet.addAll(Arrays.asList(new Feature(CLAUSES_PER_TUNIT, nrOfClauses / nrOfTunits)));
nrOfClauses = Math.max(1, nrOfClauses);
featSet.addAll(Arrays.asList(new Feature(AVG_CLAUSE_LENGTH, lengthSumClauses / nrOfClauses)));
featSet.addAll(
Arrays.asList(new Feature(COMPLEX_TUNITS_PER_TUNIT, nrOfComplexTunits / nrOfTunits)));
featSet.addAll(Arrays.asList(new Feature(COORDS_PER_TUNIT, nrOfCoords / nrOfTunits)));
featSet.addAll(
Arrays.asList(new Feature(COMPLEXNOMINALS_PER_TUNIT, nrOfComplexNominals / nrOfTunits)));
featSet.addAll(Arrays.asList(new Feature(VERBPHRASES_PER_TUNIT, nrOfVerbphrases / nrOfTunits)));
featSet.addAll(
Arrays.asList(new Feature(DEPCLAUSE_TUNIT_RATIO, nrOfDependentClauses / nrOfTunits)));
;
featSet.addAll(
Arrays.asList(new Feature(DEPCLAUSE_CLAUSE_RATIO, nrOfDependentClauses / nrOfClauses)));
featSet.addAll(Arrays.asList(new Feature(COORDS_PER_CLAUSE, nrOfCoords / nrOfClauses)));
;
featSet.addAll(
Arrays.asList(new Feature(COMPLEXNOMINALS_PER_CLAUSE, nrOfComplexNominals / nrOfClauses)));
;
return featSet;
}
