/**
* Creates annotations for each node in parse tree <CODE>node</NODE>.
* These annotations are added to the parse tree and to the document
* <CODE>doc</CODE>. In constrast to <CODE>setAnnotations</CODE>,
* the categories used for terminal nodes are Jet categories obtained by
* Jet tokenization and lexical look-up. This means that hyphenated
* items are split, and multi-word names are reduced to a single node.
*
* @param node the root of the parse tree
* @param treeSpan the span of the document matching the parse tree
* @param doc the document to which annotations will be added
*/
private void setJetAnnotations(ParseTreeNode node, Span treeSpan, Document doc) {
StatParser.buildParserInput(doc, treeSpan.start(), treeSpan.end(), false);
StatParser.fixHyphenatedItems(doc);
int nameConstitEnd = -1;
List<ParseTreeNode> terminals = getTerminalNodes(node);
for (ParseTreeNode terminal : terminals) {
int terminalEnd = terminal.end;
// is there a 'name' constituent or 'hyphword' constituent here?
Vector<Annotation> constits = doc.annotationsAt(terminal.start, "constit");
Annotation constit = null;
Annotation nameConstit = null;
Annotation hyphword = null;
if (constits != null) {
for (Annotation c : constits) {
if (c.get("cat") == "name") {
nameConstit = c;
} else if (c.get("cat") == "hyphword") {
hyphword = c;
}
if (constit == null) constit = c;
}
}
if (hyphword != null) {
nameConstit = null;
constit = hyphword;
}
// if there is a name which is not part of a hyphword, associate the
// name with this (first) terminal node, and mark any remaining terminal
// nodes which match tokens in the name as empty
if (nameConstit != null) {
terminal.end = nameConstit.end();
terminal.ann = nameConstit;
nameConstitEnd = nameConstit.end();
} else if (nameConstitEnd >= 0) {
terminal.word = null;
} else {
Span span = new Span(terminal.start, terminal.end);
String pennPOS = ((String) terminal.category).toUpperCase().intern();
String word = terminal.word;
terminal.ann = StatParser.buildWordDefn(doc, word, span, constit, pennPOS);
}
if (nameConstitEnd == terminalEnd) nameConstitEnd = -1;
}
// prune parse tree: remove a node if it has no word or children
pruneTree(node);
determineNonTerminalSpans(node, treeSpan.start());
// add head links
if (hr == null) hr = HeadRule.createDefaultRule();
hr.apply(node);
// add annotations for non-terminals:
Jet.Parser.ParseTreeNode.makeParseAnnotations(doc, node);
}
/**
* train the tagger using the DocumentCollection in file 'trainingCollection'.
* 'trainingCollection' should consist of documents which have been explicitly tagged with
* part-of-speech information.
*/
void train(String trainingCollection) {
for (int i = 0; i < posTable.length; i++)
tagTable[i] = new String[] {"constit", "cat", posTable[i], posTable[i]};
// build ergodic HMM with one state for each POS (plus start and end states)
HMMstate startState = new HMMstate("start", "", WordFeatureHMMemitter.class);
posh.addState(startState);
for (int j = 0; j < posTable.length; j++) startState.addArc(new HMMarc(posTable[j], 0));
HMMstate endState = new HMMstate("end", "", WordFeatureHMMemitter.class);
posh.addState(endState);
for (int i = 0; i < posTable.length; i++) {
String pos = posTable[i];
HMMstate state = new HMMstate(pos, pos, WordFeatureHMMemitter.class);
posh.addState(state);
for (int j = 0; j < posTable.length; j++) state.addArc(new HMMarc(posTable[j], 0));
state.addArc(new HMMarc("end", 0));
}
posh.resolveNames();
posh.resetForTraining();
annotator = new HMMannotator(posh);
annotator.setTagTable(tagTable);
annotator.setBItag(false);
DocumentCollection col = new DocumentCollection(trainingCollection);
col.open();
for (int i = 0; i < col.size(); i++) {
ExternalDocument doc = col.get(i);
doc.open();
System.out.println("Training from " + doc.fileName());
// divide at endmarks (constit cat="."), adding "S" marks
int posn = 0;
int start = posn;
Vector anns;
while ((anns = doc.annotationsAt(posn, "constit")) != null) {
Annotation ann = (Annotation) anns.get(0);
posn = ann.span().end();
String pos = (String) ann.get("cat");
if (pos.equals(".")) {
doc.annotate("S", new Span(start, posn), new FeatureSet());
start = posn;
}
}
annotator.train(doc);
// free up space taken by annotations on document
doc.clearAnnotations();
}
posh.computeProbabilities();
}