private List<String> listBasicCategories(List<String> l) {
List<String> l1 = new ArrayList<String>();
for (String s : l) {
l1.add(ctlp.basicCategory(s));
}
return l1;
}
// TODO: Rewrite this as general matching predicate
private static boolean hasV(List tags) {
for (int i = 0, tsize = tags.size(); i < tsize; i++) {
String str = tags.get(i).toString();
if (str.startsWith("V")) {
return true;
}
}
return false;
}
private List<Tree> helper(List<Tree> treeList, int start) {
List<Tree> newTreeList = new ArrayList<Tree>(treeList.size());
for (Tree tree : treeList) {
int end = start + tree.yield().size();
newTreeList.add(prune(tree, start));
start = end;
}
return newTreeList;
}
private static List<TaggedWord> cleanTags(List twList, TreebankLanguagePack tlp) {
int sz = twList.size();
List<TaggedWord> l = new ArrayList<TaggedWord>(sz);
for (int i = 0; i < sz; i++) {
TaggedWord tw = (TaggedWord) twList.get(i);
TaggedWord tw2 = new TaggedWord(tw.word(), tlp.basicCategory(tw.tag()));
l.add(tw2);
}
return l;
}
List<Tree> prune(List<Tree> treeList, Label label, int start, int end) {
// get reference tree
if (treeList.size() == 1) {
return treeList;
}
Tree testTree = treeList.get(0).treeFactory().newTreeNode(label, treeList);
int goal = Numberer.getGlobalNumberer("states").number(label.value());
Tree tempTree = parser.extractBestParse(goal, start, end);
// parser.restoreUnaries(tempTree);
Tree pcfgTree = debinarizer.transformTree(tempTree);
Set<Constituent> pcfgConstituents =
pcfgTree.constituents(new LabeledScoredConstituentFactory());
// delete child labels that are not in reference but do not cross reference
List<Tree> prunedChildren = new ArrayList<Tree>();
int childStart = 0;
for (int c = 0, numCh = testTree.numChildren(); c < numCh; c++) {
Tree child = testTree.getChild(c);
boolean isExtra = true;
int childEnd = childStart + child.yield().size();
Constituent childConstituent =
new LabeledScoredConstituent(childStart, childEnd, child.label(), 0);
if (pcfgConstituents.contains(childConstituent)) {
isExtra = false;
}
if (childConstituent.crosses(pcfgConstituents)) {
isExtra = false;
}
if (child.isLeaf() || child.isPreTerminal()) {
isExtra = false;
}
if (pcfgTree.yield().size() != testTree.yield().size()) {
isExtra = false;
}
if (!label.value().startsWith("NP^NP")) {
isExtra = false;
}
if (isExtra) {
System.err.println(
"Pruning: "
+ child.label()
+ " from "
+ (childStart + start)
+ " to "
+ (childEnd + start));
System.err.println("Was: " + testTree + " vs " + pcfgTree);
prunedChildren.addAll(child.getChildrenAsList());
} else {
prunedChildren.add(child);
}
childStart = childEnd;
}
return prunedChildren;
}
protected String historyToString(List history) {
String str = (String) historyToString.get(history);
if (str == null) {
StringBuilder sb = new StringBuilder();
for (int i = 0; i < history.size(); i++) {
sb.append('^');
sb.append(history.get(i));
}
str = sb.toString();
historyToString.put(history, str);
}
return str;
}
public Object formResult() {
Set brs = new HashSet();
Set urs = new HashSet();
// scan each rule / history pair
int ruleCount = 0;
for (Iterator pairI = rulePairs.keySet().iterator(); pairI.hasNext(); ) {
if (ruleCount % 100 == 0) {
System.err.println("Rules multiplied: " + ruleCount);
}
ruleCount++;
Pair rulePair = (Pair) pairI.next();
Rule baseRule = (Rule) rulePair.first;
String baseLabel = (String) ruleToLabel.get(baseRule);
List history = (List) rulePair.second;
double totalProb = 0;
for (int depth = 1; depth <= HISTORY_DEPTH() && depth <= history.size(); depth++) {
List subHistory = history.subList(0, depth);
double c_label = labelPairs.getCount(new Pair(baseLabel, subHistory));
double c_rule = rulePairs.getCount(new Pair(baseRule, subHistory));
// System.out.println("Multiplying out "+baseRule+" with history "+subHistory);
// System.out.println("Count of "+baseLabel+" with "+subHistory+" is "+c_label);
// System.out.println("Count of "+baseRule+" with "+subHistory+" is "+c_rule );
double prob = (1.0 / HISTORY_DEPTH()) * (c_rule) / (c_label);
totalProb += prob;
for (int childDepth = 0; childDepth <= Math.min(HISTORY_DEPTH() - 1, depth); childDepth++) {
Rule rule = specifyRule(baseRule, subHistory, childDepth);
rule.score = (float) Math.log(totalProb);
// System.out.println("Created "+rule+" with score "+rule.score);
if (rule instanceof UnaryRule) {
urs.add(rule);
} else {
brs.add(rule);
}
}
}
}
System.out.println("Total states: " + stateNumberer.total());
BinaryGrammar bg = new BinaryGrammar(stateNumberer.total());
UnaryGrammar ug = new UnaryGrammar(stateNumberer.total());
for (Iterator brI = brs.iterator(); brI.hasNext(); ) {
BinaryRule br = (BinaryRule) brI.next();
bg.addRule(br);
}
for (Iterator urI = urs.iterator(); urI.hasNext(); ) {
UnaryRule ur = (UnaryRule) urI.next();
ug.addRule(ur);
}
return new Pair(ug, bg);
}
protected void tallyInternalNode(Tree lt, List parents) {
// form base rule
String label = lt.label().value();
Rule baseR = ltToRule(lt);
ruleToLabel.put(baseR, label);
// act on each history depth
for (int depth = 0, maxDepth = Math.min(HISTORY_DEPTH(), parents.size());
depth <= maxDepth;
depth++) {
List history = new ArrayList(parents.subList(0, depth));
// tally each history level / rewrite pair
rulePairs.incrementCount(new Pair(baseR, history), 1);
labelPairs.incrementCount(new Pair(label, history), 1);
}
}
public static ArrayList<ArrayList<TaggedWord>> getPhrases(Tree parse, int phraseSizeLimit) {
ArrayList<ArrayList<TaggedWord>> newList = new ArrayList<ArrayList<TaggedWord>>();
List<Tree> leaves = parse.getLeaves();
if (leaves.size() <= phraseSizeLimit) {
// ArrayList<TaggedWord> phraseElements = PreprocessPhrase(parse.taggedYield());
ArrayList<TaggedWord> phraseElements = Preprocess(parse.taggedYield());
if (phraseElements.size() > 0) newList.add(phraseElements);
} else {
Tree[] childrenNodes = parse.children();
for (int i = 0; i < childrenNodes.length; i++) {
Tree currentParse = childrenNodes[i];
newList.addAll(getPhrases(currentParse, phraseSizeLimit));
}
}
return newList;
}
/**
* Assumes the given reader has only tsurgeon operations (not a tregex pattern), and parses these
* out, collecting them into one operation. Stops on a whitespace line.
*
* @throws IOException
*/
public static TsurgeonPattern getTsurgeonOperationsFromReader(BufferedReader reader)
throws IOException {
List<TsurgeonPattern> operations = new ArrayList<TsurgeonPattern>();
for (String thisLine; (thisLine = reader.readLine()) != null; ) {
if (emptyLinePattern.matcher(thisLine).matches()) {
break;
}
thisLine = removeComments(thisLine);
if (emptyLinePattern.matcher(thisLine).matches()) {
continue;
}
// System.err.println("Read tsurgeon op: " + thisLine);
operations.add(parseOperation(thisLine));
}
if (operations.size() == 0) throw new TsurgeonParseException("No Tsurgeon operation provided.");
return collectOperations(operations);
}
protected Rule specifyRule(Rule rule, List history, int childDepth) {
Rule r;
String topHistoryStr = historyToString(history.subList(1, history.size()));
String bottomHistoryStr = historyToString(history.subList(0, childDepth));
if (rule instanceof UnaryRule) {
UnaryRule ur = new UnaryRule();
UnaryRule urule = (UnaryRule) rule;
ur.parent = stateNumberer.number(stateNumberer.object(urule.parent) + topHistoryStr);
if (isSynthetic(urule.child)) {
ur.child = stateNumberer.number(stateNumberer.object(urule.child) + topHistoryStr);
} else if (isTag(urule.child)) {
ur.child = urule.child;
} else {
ur.child = stateNumberer.number(stateNumberer.object(urule.child) + bottomHistoryStr);
}
r = ur;
} else {
BinaryRule br = new BinaryRule();
BinaryRule brule = (BinaryRule) rule;
br.parent = stateNumberer.number(stateNumberer.object(brule.parent) + topHistoryStr);
if (isSynthetic(brule.leftChild)) {
br.leftChild = stateNumberer.number(stateNumberer.object(brule.leftChild) + topHistoryStr);
} else if (isTag(brule.leftChild)) {
br.leftChild = brule.leftChild;
} else {
br.leftChild =
stateNumberer.number(stateNumberer.object(brule.leftChild) + bottomHistoryStr);
}
if (isSynthetic(brule.rightChild)) {
br.rightChild =
stateNumberer.number(stateNumberer.object(brule.rightChild) + topHistoryStr);
} else if (isTag(brule.rightChild)) {
br.rightChild = brule.rightChild;
} else {
br.rightChild =
stateNumberer.number(stateNumberer.object(brule.rightChild) + bottomHistoryStr);
}
r = br;
}
return r;
}
/**
* transformTree does all language-specific tree transformations. Any parameterizations should be
* inside the specific TreebankLangParserParams class.
*/
@Override
public Tree transformTree(Tree t, Tree root) {
if (t == null || t.isLeaf()) {
return t;
}
String parentStr;
String grandParentStr;
Tree parent;
Tree grandParent;
if (root == null || t.equals(root)) {
parent = null;
parentStr = "";
} else {
parent = t.parent(root);
parentStr = parent.label().value();
}
if (parent == null || parent.equals(root)) {
grandParent = null;
grandParentStr = "";
} else {
grandParent = parent.parent(root);
grandParentStr = grandParent.label().value();
}
String baseParentStr = ctlp.basicCategory(parentStr);
String baseGrandParentStr = ctlp.basicCategory(grandParentStr);
CoreLabel lab = (CoreLabel) t.label();
String word = lab.word();
String tag = lab.tag();
String baseTag = ctlp.basicCategory(tag);
String category = lab.value();
String baseCategory = ctlp.basicCategory(category);
if (t.isPreTerminal()) { // it's a POS tag
List<String> leftAunts =
listBasicCategories(SisterAnnotationStats.leftSisterLabels(parent, grandParent));
List<String> rightAunts =
listBasicCategories(SisterAnnotationStats.rightSisterLabels(parent, grandParent));
// Chinese-specific punctuation splits
if (chineseSplitPunct && baseTag.equals("PU")) {
if (ChineseTreebankLanguagePack.chineseDouHaoAcceptFilter().accept(word)) {
tag = tag + "-DOU";
// System.out.println("Punct: Split dou hao"); // debugging
} else if (ChineseTreebankLanguagePack.chineseCommaAcceptFilter().accept(word)) {
tag = tag + "-COMMA";
// System.out.println("Punct: Split comma"); // debugging
} else if (ChineseTreebankLanguagePack.chineseColonAcceptFilter().accept(word)) {
tag = tag + "-COLON";
// System.out.println("Punct: Split colon"); // debugging
} else if (ChineseTreebankLanguagePack.chineseQuoteMarkAcceptFilter().accept(word)) {
if (chineseSplitPunctLR) {
if (ChineseTreebankLanguagePack.chineseLeftQuoteMarkAcceptFilter().accept(word)) {
tag += "-LQUOTE";
} else {
tag += "-RQUOTE";
}
} else {
tag = tag + "-QUOTE";
}
// System.out.println("Punct: Split quote"); // debugging
} else if (ChineseTreebankLanguagePack.chineseEndSentenceAcceptFilter().accept(word)) {
tag = tag + "-ENDSENT";
// System.out.println("Punct: Split end sent"); // debugging
} else if (ChineseTreebankLanguagePack.chineseParenthesisAcceptFilter().accept(word)) {
if (chineseSplitPunctLR) {
if (ChineseTreebankLanguagePack.chineseLeftParenthesisAcceptFilter().accept(word)) {
tag += "-LPAREN";
} else {
tag += "-RPAREN";
}
} else {
tag += "-PAREN";
// printlnErr("Just used -PAREN annotation");
// printlnErr(word);
// throw new RuntimeException();
}
// System.out.println("Punct: Split paren"); // debugging
} else if (ChineseTreebankLanguagePack.chineseDashAcceptFilter().accept(word)) {
tag = tag + "-DASH";
// System.out.println("Punct: Split dash"); // debugging
} else if (ChineseTreebankLanguagePack.chineseOtherAcceptFilter().accept(word)) {
tag = tag + "-OTHER";
} else {
printlnErr("Unknown punct (you should add it to CTLP): " + tag + " |" + word + "|");
}
} else if (chineseSplitDouHao) { // only split DouHao
if (ChineseTreebankLanguagePack.chineseDouHaoAcceptFilter().accept(word)
&& baseTag.equals("PU")) {
tag = tag + "-DOU";
}
}
// Chinese-specific POS tag splits (non-punctuation)
if (tagWordSize) {
int l = word.length();
tag += "-" + l + "CHARS";
}
if (mergeNNVV && baseTag.equals("NN")) {
tag = "VV";
}
if ((chineseSelectiveTagPA || chineseVerySelectiveTagPA)
&& (baseTag.equals("CC") || baseTag.equals("P"))) {
tag += "-" + baseParentStr;
}
if (chineseSelectiveTagPA && (baseTag.equals("VV"))) {
tag += "-" + baseParentStr;
}
if (markMultiNtag && tag.startsWith("N")) {
for (int i = 0; i < parent.numChildren(); i++) {
if (parent.children()[i].label().value().startsWith("N") && parent.children()[i] != t) {
tag += "=N";
// System.out.println("Found multi=N rewrite");
}
}
}
if (markVVsisterIP && baseTag.equals("VV")) {
boolean seenIP = false;
for (int i = 0; i < parent.numChildren(); i++) {
if (parent.children()[i].label().value().startsWith("IP")) {
seenIP = true;
}
}
if (seenIP) {
tag += "-IP";
// System.out.println("Found VV with IP sister"); // testing
}
}
if (markPsisterIP && baseTag.equals("P")) {
boolean seenIP = false;
for (int i = 0; i < parent.numChildren(); i++) {
if (parent.children()[i].label().value().startsWith("IP")) {
seenIP = true;
}
}
if (seenIP) {
tag += "-IP";
}
}
if (markADgrandchildOfIP && baseTag.equals("AD") && baseGrandParentStr.equals("IP")) {
tag += "~IP";
// System.out.println("Found AD with IP grandparent"); // testing
}
if (gpaAD && baseTag.equals("AD")) {
tag += "~" + baseGrandParentStr;
// System.out.println("Found AD with grandparent " + grandParentStr); // testing
}
if (markPostverbalP && leftAunts.contains("VV") && baseTag.equals("P")) {
// System.out.println("Found post-verbal P");
tag += "^=lVV";
}
// end Chinese-specific tag splits
Label label = new CategoryWordTag(tag, word, tag);
t.setLabel(label);
} else {
// it's a phrasal category
Tree[] kids = t.children();
// Chinese-specific category splits
List<String> leftSis = listBasicCategories(SisterAnnotationStats.leftSisterLabels(t, parent));
List<String> rightSis =
listBasicCategories(SisterAnnotationStats.rightSisterLabels(t, parent));
if (paRootDtr && baseParentStr.equals("ROOT")) {
category += "^ROOT";
}
if (markIPsisterBA && baseCategory.equals("IP")) {
if (leftSis.contains("BA")) {
category += "=BA";
// System.out.println("Found IP sister of BA");
}
}
if (dominatesV && hasV(t.preTerminalYield())) {
// mark categories containing a verb
category += "-v";
}
if (markIPsisterVVorP && baseCategory.equals("IP")) {
// todo: cdm: is just looking for "P" here selective enough??
if (leftSis.contains("VV") || leftSis.contains("P")) {
category += "=VVP";
}
}
if (markIPsisDEC && baseCategory.equals("IP")) {
if (rightSis.contains("DEC")) {
category += "=DEC";
// System.out.println("Found prenominal IP");
}
}
if (baseCategory.equals("VP")) {
// cdm 2008: this used to just check that it startsWith("VP"), but
// I think that was bad because it also matched VPT verb compounds
if (chineseSplitVP == 3) {
boolean hasCC = false;
boolean hasPU = false;
boolean hasLexV = false;
for (Tree kid : kids) {
if (kid.label().value().startsWith("CC")) {
hasCC = true;
} else if (kid.label().value().startsWith("PU")) {
hasPU = true;
} else if (StringUtils.lookingAt(
kid.label().value(), "(V[ACEV]|VCD|VCP|VNV|VPT|VRD|VSB)")) {
hasLexV = true;
}
}
if (hasCC || (hasPU && !hasLexV)) {
category += "-CRD";
// System.out.println("Found coordinate VP"); // testing
} else if (hasLexV) {
category += "-COMP";
// System.out.println("Found complementing VP"); // testing
} else {
category += "-ADJT";
// System.out.println("Found adjoining VP"); // testing
}
} else if (chineseSplitVP >= 1) {
boolean hasBA = false;
for (Tree kid : kids) {
if (kid.label().value().startsWith("BA")) {
hasBA = true;
} else if (chineseSplitVP == 2 && tlp.basicCategory(kid.label().value()).equals("VP")) {
for (Tree kidkid : kid.children()) {
if (kidkid.label().value().startsWith("BA")) {
hasBA = true;
}
}
}
}
if (hasBA) {
category += "-BA";
}
}
}
if (markVPadjunct && baseParentStr.equals("VP")) {
// cdm 2008: This used to use startsWith("VP") but changed to baseCat
Tree[] sisters = parent.children();
boolean hasVPsister = false;
boolean hasCC = false;
boolean hasPU = false;
boolean hasLexV = false;
for (Tree sister : sisters) {
if (tlp.basicCategory(sister.label().value()).equals("VP")) {
hasVPsister = true;
}
if (sister.label().value().startsWith("CC")) {
hasCC = true;
}
if (sister.label().value().startsWith("PU")) {
hasPU = true;
}
if (StringUtils.lookingAt(sister.label().value(), "(V[ACEV]|VCD|VCP|VNV|VPT|VRD|VSB)")) {
hasLexV = true;
}
}
if (hasVPsister && !(hasCC || hasPU || hasLexV)) {
category += "-VPADJ";
// System.out.println("Found adjunct of VP"); // testing
}
}
if (markNPmodNP && baseCategory.equals("NP") && baseParentStr.equals("NP")) {
if (rightSis.contains("NP")) {
category += "=MODIFIERNP";
// System.out.println("Found NP modifier of NP"); // testing
}
}
if (markModifiedNP && baseCategory.equals("NP") && baseParentStr.equals("NP")) {
if (rightSis.isEmpty()
&& (leftSis.contains("ADJP")
|| leftSis.contains("NP")
|| leftSis.contains("DNP")
|| leftSis.contains("QP")
|| leftSis.contains("CP")
|| leftSis.contains("PP"))) {
category += "=MODIFIEDNP";
// System.out.println("Found modified NP"); // testing
}
}
if (markNPconj && baseCategory.equals("NP") && baseParentStr.equals("NP")) {
if (rightSis.contains("CC")
|| rightSis.contains("PU")
|| leftSis.contains("CC")
|| leftSis.contains("PU")) {
category += "=CONJ";
// System.out.println("Found NP conjunct"); // testing
}
}
if (markIPconj && baseCategory.equals("IP") && baseParentStr.equals("IP")) {
Tree[] sisters = parent.children();
boolean hasCommaSis = false;
boolean hasIPSis = false;
for (Tree sister : sisters) {
if (ctlp.basicCategory(sister.label().value()).equals("PU")
&& ChineseTreebankLanguagePack.chineseCommaAcceptFilter()
.accept(sister.children()[0].label().toString())) {
hasCommaSis = true;
// System.out.println("Found CommaSis"); // testing
}
if (ctlp.basicCategory(sister.label().value()).equals("IP") && sister != t) {
hasIPSis = true;
}
}
if (hasCommaSis && hasIPSis) {
category += "-CONJ";
// System.out.println("Found IP conjunct"); // testing
}
}
if (unaryIP && baseCategory.equals("IP") && t.numChildren() == 1) {
category += "-U";
// System.out.println("Found unary IP"); //testing
}
if (unaryCP && baseCategory.equals("CP") && t.numChildren() == 1) {
category += "-U";
// System.out.println("Found unary CP"); //testing
}
if (splitBaseNP && baseCategory.equals("NP")) {
if (t.isPrePreTerminal()) {
category = category + "-B";
}
}
// if (Test.verbose) printlnErr(baseCategory + " " + leftSis.toString()); //debugging
if (markPostverbalPP && leftSis.contains("VV") && baseCategory.equals("PP")) {
// System.out.println("Found post-verbal PP");
category += "=lVV";
}
if ((markADgrandchildOfIP || gpaAD)
&& listBasicCategories(SisterAnnotationStats.kidLabels(t)).contains("AD")) {
category += "^ADVP";
}
if (markCC) {
// was: for (int i = 0; i < kids.length; i++) {
// This second version takes an idea from Collins: don't count
// marginal conjunctions which don't conjoin 2 things.
for (int i = 1; i < kids.length - 1; i++) {
String cat2 = kids[i].label().value();
if (cat2.startsWith("CC")) {
category += "-CC";
}
}
}
Label label = new CategoryWordTag(category, word, tag);
t.setLabel(label);
}
return t;
}
/**
* Collects a list of operation patterns into a sequence of operations to be applied. Required to
* keep track of global properties across a sequence of operations. For example, if you want to
* insert a named node and then coindex it with another node, you will need to collect the
* insertion and coindexation operations into a single TsurgeonPattern so that tsurgeon is aware
* of the name of the new node and coindexation becomes possible.
*
* @param patterns a list of {@link TsurgeonPattern} operations that you want to collect together
* into a single compound operation
* @return a new {@link TsurgeonPattern} that performs all the operations in the sequence of the
* <code>patterns</code> argument
*/
public static TsurgeonPattern collectOperations(List<TsurgeonPattern> patterns) {
return new TsurgeonPatternRoot(patterns.toArray(new TsurgeonPattern[patterns.size()]));
}
/**
* Applies {#processPattern} to a collection of trees.
*
* @param matchPattern A {@link TregexPattern} to be matched against a {@link Tree}.
* @param p A {@link TsurgeonPattern} to apply.
* @param inputTrees The input trees to be processed
* @return A List of the transformed trees
*/
public static List<Tree> processPatternOnTrees(
TregexPattern matchPattern, TsurgeonPattern p, Collection<Tree> inputTrees) {
List<Tree> result = new ArrayList<Tree>();
for (Tree tree : inputTrees) result.add(processPattern(matchPattern, p, tree));
return result;
}
public static void main(String[] args) {
Options op = new Options(new EnglishTreebankParserParams());
// op.tlpParams may be changed to something else later, so don't use it till
// after options are parsed.
System.out.println("Currently " + new Date());
System.out.print("Invoked with arguments:");
for (String arg : args) {
System.out.print(" " + arg);
}
System.out.println();
String path = "/u/nlp/stuff/corpora/Treebank3/parsed/mrg/wsj";
int trainLow = 200, trainHigh = 2199, testLow = 2200, testHigh = 2219;
String serializeFile = null;
int i = 0;
while (i < args.length && args[i].startsWith("-")) {
if (args[i].equalsIgnoreCase("-path") && (i + 1 < args.length)) {
path = args[i + 1];
i += 2;
} else if (args[i].equalsIgnoreCase("-train") && (i + 2 < args.length)) {
trainLow = Integer.parseInt(args[i + 1]);
trainHigh = Integer.parseInt(args[i + 2]);
i += 3;
} else if (args[i].equalsIgnoreCase("-test") && (i + 2 < args.length)) {
testLow = Integer.parseInt(args[i + 1]);
testHigh = Integer.parseInt(args[i + 2]);
i += 3;
} else if (args[i].equalsIgnoreCase("-serialize") && (i + 1 < args.length)) {
serializeFile = args[i + 1];
i += 2;
} else if (args[i].equalsIgnoreCase("-tLPP") && (i + 1 < args.length)) {
try {
op.tlpParams = (TreebankLangParserParams) Class.forName(args[i + 1]).newInstance();
} catch (ClassNotFoundException e) {
System.err.println("Class not found: " + args[i + 1]);
} catch (InstantiationException e) {
System.err.println("Couldn't instantiate: " + args[i + 1] + ": " + e.toString());
} catch (IllegalAccessException e) {
System.err.println("illegal access" + e);
}
i += 2;
} else if (args[i].equals("-encoding")) {
// sets encoding for TreebankLangParserParams
op.tlpParams.setInputEncoding(args[i + 1]);
op.tlpParams.setOutputEncoding(args[i + 1]);
i += 2;
} else {
i = op.setOptionOrWarn(args, i);
}
}
// System.out.println(tlpParams.getClass());
TreebankLanguagePack tlp = op.tlpParams.treebankLanguagePack();
Train.sisterSplitters = new HashSet(Arrays.asList(op.tlpParams.sisterSplitters()));
// BinarizerFactory.TreeAnnotator.setTreebankLang(tlpParams);
PrintWriter pw = op.tlpParams.pw();
Test.display();
Train.display();
op.display();
op.tlpParams.display();
// setup tree transforms
Treebank trainTreebank = op.tlpParams.memoryTreebank();
MemoryTreebank testTreebank = op.tlpParams.testMemoryTreebank();
// Treebank blippTreebank = ((EnglishTreebankParserParams) tlpParams).diskTreebank();
// String blippPath = "/afs/ir.stanford.edu/data/linguistic-data/BLLIP-WSJ/";
// blippTreebank.loadPath(blippPath, "", true);
Timing.startTime();
System.err.print("Reading trees...");
testTreebank.loadPath(path, new NumberRangeFileFilter(testLow, testHigh, true));
if (Test.increasingLength) {
Collections.sort(testTreebank, new TreeLengthComparator());
}
trainTreebank.loadPath(path, new NumberRangeFileFilter(trainLow, trainHigh, true));
Timing.tick("done.");
System.err.print("Binarizing trees...");
TreeAnnotatorAndBinarizer binarizer = null;
if (!Train.leftToRight) {
binarizer =
new TreeAnnotatorAndBinarizer(op.tlpParams, op.forceCNF, !Train.outsideFactor(), true);
} else {
binarizer =
new TreeAnnotatorAndBinarizer(
op.tlpParams.headFinder(),
new LeftHeadFinder(),
op.tlpParams,
op.forceCNF,
!Train.outsideFactor(),
true);
}
CollinsPuncTransformer collinsPuncTransformer = null;
if (Train.collinsPunc) {
collinsPuncTransformer = new CollinsPuncTransformer(tlp);
}
TreeTransformer debinarizer = new Debinarizer(op.forceCNF);
List<Tree> binaryTrainTrees = new ArrayList<Tree>();
if (Train.selectiveSplit) {
Train.splitters =
ParentAnnotationStats.getSplitCategories(
trainTreebank,
Train.tagSelectiveSplit,
0,
Train.selectiveSplitCutOff,
Train.tagSelectiveSplitCutOff,
op.tlpParams.treebankLanguagePack());
if (Train.deleteSplitters != null) {
List<String> deleted = new ArrayList<String>();
for (String del : Train.deleteSplitters) {
String baseDel = tlp.basicCategory(del);
boolean checkBasic = del.equals(baseDel);
for (Iterator<String> it = Train.splitters.iterator(); it.hasNext(); ) {
String elem = it.next();
String baseElem = tlp.basicCategory(elem);
boolean delStr = checkBasic && baseElem.equals(baseDel) || elem.equals(del);
if (delStr) {
it.remove();
deleted.add(elem);
}
}
}
System.err.println("Removed from vertical splitters: " + deleted);
}
}
if (Train.selectivePostSplit) {
TreeTransformer myTransformer = new TreeAnnotator(op.tlpParams.headFinder(), op.tlpParams);
Treebank annotatedTB = trainTreebank.transform(myTransformer);
Train.postSplitters =
ParentAnnotationStats.getSplitCategories(
annotatedTB,
true,
0,
Train.selectivePostSplitCutOff,
Train.tagSelectivePostSplitCutOff,
op.tlpParams.treebankLanguagePack());
}
if (Train.hSelSplit) {
binarizer.setDoSelectiveSplit(false);
for (Tree tree : trainTreebank) {
if (Train.collinsPunc) {
tree = collinsPuncTransformer.transformTree(tree);
}
// tree.pennPrint(tlpParams.pw());
tree = binarizer.transformTree(tree);
// binaryTrainTrees.add(tree);
}
binarizer.setDoSelectiveSplit(true);
}
for (Tree tree : trainTreebank) {
if (Train.collinsPunc) {
tree = collinsPuncTransformer.transformTree(tree);
}
tree = binarizer.transformTree(tree);
binaryTrainTrees.add(tree);
}
if (Test.verbose) {
binarizer.dumpStats();
}
List<Tree> binaryTestTrees = new ArrayList<Tree>();
for (Tree tree : testTreebank) {
if (Train.collinsPunc) {
tree = collinsPuncTransformer.transformTree(tree);
}
tree = binarizer.transformTree(tree);
binaryTestTrees.add(tree);
}
Timing.tick("done."); // binarization
BinaryGrammar bg = null;
UnaryGrammar ug = null;
DependencyGrammar dg = null;
// DependencyGrammar dgBLIPP = null;
Lexicon lex = null;
// extract grammars
Extractor bgExtractor = new BinaryGrammarExtractor();
// Extractor bgExtractor = new SmoothedBinaryGrammarExtractor();//new BinaryGrammarExtractor();
// Extractor lexExtractor = new LexiconExtractor();
// Extractor dgExtractor = new DependencyMemGrammarExtractor();
Extractor dgExtractor = new MLEDependencyGrammarExtractor(op);
if (op.doPCFG) {
System.err.print("Extracting PCFG...");
Pair bgug = null;
if (Train.cheatPCFG) {
List allTrees = new ArrayList(binaryTrainTrees);
allTrees.addAll(binaryTestTrees);
bgug = (Pair) bgExtractor.extract(allTrees);
} else {
bgug = (Pair) bgExtractor.extract(binaryTrainTrees);
}
bg = (BinaryGrammar) bgug.second;
bg.splitRules();
ug = (UnaryGrammar) bgug.first;
ug.purgeRules();
Timing.tick("done.");
}
System.err.print("Extracting Lexicon...");
lex = op.tlpParams.lex(op.lexOptions);
lex.train(binaryTrainTrees);
Timing.tick("done.");
if (op.doDep) {
System.err.print("Extracting Dependencies...");
binaryTrainTrees.clear();
// dgBLIPP = (DependencyGrammar) dgExtractor.extract(new
// ConcatenationIterator(trainTreebank.iterator(),blippTreebank.iterator()),new
// TransformTreeDependency(tlpParams,true));
DependencyGrammar dg1 =
(DependencyGrammar)
dgExtractor.extract(
trainTreebank.iterator(), new TransformTreeDependency(op.tlpParams, true));
// dgBLIPP=(DependencyGrammar)dgExtractor.extract(blippTreebank.iterator(),new
// TransformTreeDependency(tlpParams));
// dg = (DependencyGrammar) dgExtractor.extract(new
// ConcatenationIterator(trainTreebank.iterator(),blippTreebank.iterator()),new
// TransformTreeDependency(tlpParams));
// dg=new DependencyGrammarCombination(dg1,dgBLIPP,2);
// dg = (DependencyGrammar) dgExtractor.extract(binaryTrainTrees); //uses information whether
// the words are known or not, discards unknown words
Timing.tick("done.");
// System.out.print("Extracting Unknown Word Model...");
// UnknownWordModel uwm = (UnknownWordModel)uwmExtractor.extract(binaryTrainTrees);
// Timing.tick("done.");
System.out.print("Tuning Dependency Model...");
dg.tune(binaryTestTrees);
// System.out.println("TUNE DEPS: "+tuneDeps);
Timing.tick("done.");
}
BinaryGrammar boundBG = bg;
UnaryGrammar boundUG = ug;
GrammarProjection gp = new NullGrammarProjection(bg, ug);
// serialization
if (serializeFile != null) {
System.err.print("Serializing parser...");
LexicalizedParser.saveParserDataToSerialized(
new ParserData(lex, bg, ug, dg, Numberer.getNumberers(), op), serializeFile);
Timing.tick("done.");
}
// test: pcfg-parse and output
ExhaustivePCFGParser parser = null;
if (op.doPCFG) {
parser = new ExhaustivePCFGParser(boundBG, boundUG, lex, op);
}
ExhaustiveDependencyParser dparser =
((op.doDep && !Test.useFastFactored) ? new ExhaustiveDependencyParser(dg, lex, op) : null);
Scorer scorer = (op.doPCFG ? new TwinScorer(new ProjectionScorer(parser, gp), dparser) : null);
// Scorer scorer = parser;
BiLexPCFGParser bparser = null;
if (op.doPCFG && op.doDep) {
bparser =
(Test.useN5)
? new BiLexPCFGParser.N5BiLexPCFGParser(
scorer, parser, dparser, bg, ug, dg, lex, op, gp)
: new BiLexPCFGParser(scorer, parser, dparser, bg, ug, dg, lex, op, gp);
}
LabeledConstituentEval pcfgPE = new LabeledConstituentEval("pcfg PE", true, tlp);
LabeledConstituentEval comboPE = new LabeledConstituentEval("combo PE", true, tlp);
AbstractEval pcfgCB = new LabeledConstituentEval.CBEval("pcfg CB", true, tlp);
AbstractEval pcfgTE = new AbstractEval.TaggingEval("pcfg TE");
AbstractEval comboTE = new AbstractEval.TaggingEval("combo TE");
AbstractEval pcfgTEnoPunct = new AbstractEval.TaggingEval("pcfg nopunct TE");
AbstractEval comboTEnoPunct = new AbstractEval.TaggingEval("combo nopunct TE");
AbstractEval depTE = new AbstractEval.TaggingEval("depnd TE");
AbstractEval depDE =
new AbstractEval.DependencyEval("depnd DE", true, tlp.punctuationWordAcceptFilter());
AbstractEval comboDE =
new AbstractEval.DependencyEval("combo DE", true, tlp.punctuationWordAcceptFilter());
if (Test.evalb) {
EvalB.initEVALBfiles(op.tlpParams);
}
// int[] countByLength = new int[Test.maxLength+1];
// use a reflection ruse, so one can run this without needing the tagger
// edu.stanford.nlp.process.SentenceTagger tagger = (Test.preTag ? new
// edu.stanford.nlp.process.SentenceTagger("/u/nlp/data/tagger.params/wsj0-21.holder") : null);
SentenceProcessor tagger = null;
if (Test.preTag) {
try {
Class[] argsClass = new Class[] {String.class};
Object[] arguments =
new Object[] {"/u/nlp/data/pos-tagger/wsj3t0-18-bidirectional/train-wsj-0-18.holder"};
tagger =
(SentenceProcessor)
Class.forName("edu.stanford.nlp.tagger.maxent.MaxentTagger")
.getConstructor(argsClass)
.newInstance(arguments);
} catch (Exception e) {
System.err.println(e);
System.err.println("Warning: No pretagging of sentences will be done.");
}
}
for (int tNum = 0, ttSize = testTreebank.size(); tNum < ttSize; tNum++) {
Tree tree = testTreebank.get(tNum);
int testTreeLen = tree.yield().size();
if (testTreeLen > Test.maxLength) {
continue;
}
Tree binaryTree = binaryTestTrees.get(tNum);
// countByLength[testTreeLen]++;
System.out.println("-------------------------------------");
System.out.println("Number: " + (tNum + 1));
System.out.println("Length: " + testTreeLen);
// tree.pennPrint(pw);
// System.out.println("XXXX The binary tree is");
// binaryTree.pennPrint(pw);
// System.out.println("Here are the tags in the lexicon:");
// System.out.println(lex.showTags());
// System.out.println("Here's the tagnumberer:");
// System.out.println(Numberer.getGlobalNumberer("tags").toString());
long timeMil1 = System.currentTimeMillis();
Timing.tick("Starting parse.");
if (op.doPCFG) {
// System.err.println(Test.forceTags);
if (Test.forceTags) {
if (tagger != null) {
// System.out.println("Using a tagger to set tags");
// System.out.println("Tagged sentence as: " +
// tagger.processSentence(cutLast(wordify(binaryTree.yield()))).toString(false));
parser.parse(addLast(tagger.processSentence(cutLast(wordify(binaryTree.yield())))));
} else {
// System.out.println("Forcing tags to match input.");
parser.parse(cleanTags(binaryTree.taggedYield(), tlp));
}
} else {
// System.out.println("XXXX Parsing " + binaryTree.yield());
parser.parse(binaryTree.yield());
}
// Timing.tick("Done with pcfg phase.");
}
if (op.doDep) {
dparser.parse(binaryTree.yield());
// Timing.tick("Done with dependency phase.");
}
boolean bothPassed = false;
if (op.doPCFG && op.doDep) {
bothPassed = bparser.parse(binaryTree.yield());
// Timing.tick("Done with combination phase.");
}
long timeMil2 = System.currentTimeMillis();
long elapsed = timeMil2 - timeMil1;
System.err.println("Time: " + ((int) (elapsed / 100)) / 10.00 + " sec.");
// System.out.println("PCFG Best Parse:");
Tree tree2b = null;
Tree tree2 = null;
// System.out.println("Got full best parse...");
if (op.doPCFG) {
tree2b = parser.getBestParse();
tree2 = debinarizer.transformTree(tree2b);
}
// System.out.println("Debinarized parse...");
// tree2.pennPrint();
// System.out.println("DepG Best Parse:");
Tree tree3 = null;
Tree tree3db = null;
if (op.doDep) {
tree3 = dparser.getBestParse();
// was: but wrong Tree tree3db = debinarizer.transformTree(tree2);
tree3db = debinarizer.transformTree(tree3);
tree3.pennPrint(pw);
}
// tree.pennPrint();
// ((Tree)binaryTrainTrees.get(tNum)).pennPrint();
// System.out.println("Combo Best Parse:");
Tree tree4 = null;
if (op.doPCFG && op.doDep) {
try {
tree4 = bparser.getBestParse();
if (tree4 == null) {
tree4 = tree2b;
}
} catch (NullPointerException e) {
System.err.println("Blocked, using PCFG parse!");
tree4 = tree2b;
}
}
if (op.doPCFG && !bothPassed) {
tree4 = tree2b;
}
// tree4.pennPrint();
if (op.doDep) {
depDE.evaluate(tree3, binaryTree, pw);
depTE.evaluate(tree3db, tree, pw);
}
TreeTransformer tc = op.tlpParams.collinizer();
TreeTransformer tcEvalb = op.tlpParams.collinizerEvalb();
Tree tree4b = null;
if (op.doPCFG) {
// System.out.println("XXXX Best PCFG was: ");
// tree2.pennPrint();
// System.out.println("XXXX Transformed best PCFG is: ");
// tc.transformTree(tree2).pennPrint();
// System.out.println("True Best Parse:");
// tree.pennPrint();
// tc.transformTree(tree).pennPrint();
pcfgPE.evaluate(tc.transformTree(tree2), tc.transformTree(tree), pw);
pcfgCB.evaluate(tc.transformTree(tree2), tc.transformTree(tree), pw);
if (op.doDep) {
comboDE.evaluate((bothPassed ? tree4 : tree3), binaryTree, pw);
tree4b = tree4;
tree4 = debinarizer.transformTree(tree4);
if (op.nodePrune) {
NodePruner np = new NodePruner(parser, debinarizer);
tree4 = np.prune(tree4);
}
// tree4.pennPrint();
comboPE.evaluate(tc.transformTree(tree4), tc.transformTree(tree), pw);
}
// pcfgTE.evaluate(tree2, tree);
pcfgTE.evaluate(tcEvalb.transformTree(tree2), tcEvalb.transformTree(tree), pw);
pcfgTEnoPunct.evaluate(tc.transformTree(tree2), tc.transformTree(tree), pw);
if (op.doDep) {
comboTE.evaluate(tcEvalb.transformTree(tree4), tcEvalb.transformTree(tree), pw);
comboTEnoPunct.evaluate(tc.transformTree(tree4), tc.transformTree(tree), pw);
}
System.out.println("PCFG only: " + parser.scoreBinarizedTree(tree2b, 0));
// tc.transformTree(tree2).pennPrint();
tree2.pennPrint(pw);
if (op.doDep) {
System.out.println("Combo: " + parser.scoreBinarizedTree(tree4b, 0));
// tc.transformTree(tree4).pennPrint(pw);
tree4.pennPrint(pw);
}
System.out.println("Correct:" + parser.scoreBinarizedTree(binaryTree, 0));
/*
if (parser.scoreBinarizedTree(tree2b,true) < parser.scoreBinarizedTree(binaryTree,true)) {
System.out.println("SCORE INVERSION");
parser.validateBinarizedTree(binaryTree,0);
}
*/
tree.pennPrint(pw);
} // end if doPCFG
if (Test.evalb) {
if (op.doPCFG && op.doDep) {
EvalB.writeEVALBline(tcEvalb.transformTree(tree), tcEvalb.transformTree(tree4));
} else if (op.doPCFG) {
EvalB.writeEVALBline(tcEvalb.transformTree(tree), tcEvalb.transformTree(tree2));
} else if (op.doDep) {
EvalB.writeEVALBline(tcEvalb.transformTree(tree), tcEvalb.transformTree(tree3db));
}
}
} // end for each tree in test treebank
if (Test.evalb) {
EvalB.closeEVALBfiles();
}
// Test.display();
if (op.doPCFG) {
pcfgPE.display(false, pw);
System.out.println("Grammar size: " + Numberer.getGlobalNumberer("states").total());
pcfgCB.display(false, pw);
if (op.doDep) {
comboPE.display(false, pw);
}
pcfgTE.display(false, pw);
pcfgTEnoPunct.display(false, pw);
if (op.doDep) {
comboTE.display(false, pw);
comboTEnoPunct.display(false, pw);
}
}
if (op.doDep) {
depTE.display(false, pw);
depDE.display(false, pw);
}
if (op.doPCFG && op.doDep) {
comboDE.display(false, pw);
}
// pcfgPE.printGoodBad();
}