edu.stanford.nlp.trees.DependencyScoring Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of stanford-parser Show documentation
Show all versions of stanford-parser Show documentation
Stanford Parser processes raw text in English, Chinese, German, Arabic, and French, and extracts constituency parse trees.
package edu.stanford.nlp.trees;
import edu.stanford.nlp.util.logging.Redwood;
import static edu.stanford.nlp.trees.GrammaticalRelation.DEPENDENT;
import java.io.LineNumberReader;
import java.io.FileReader;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Collection;
import java.util.List;
import java.util.Map;
import java.util.Properties;
import java.util.Set;
import edu.stanford.nlp.international.Language;
import edu.stanford.nlp.ling.IndexedWord;
import edu.stanford.nlp.util.Generics;
import edu.stanford.nlp.util.StringUtils;
import edu.stanford.nlp.stats.ClassicCounter;
import edu.stanford.nlp.stats.Counters;
/**
* Scoring of typed dependencies
*
* @author danielcer
*
*/
public class DependencyScoring {
/** A logger for this class */
private static Redwood.RedwoodChannels log = Redwood.channels(DependencyScoring.class);
public final static boolean VERBOSE = false;
public final List> goldDeps;
public final List> goldDepsUnlabeled;
public final boolean ignorePunc;
private static List> toSets(Collection depCollection) {
Set depSet = Generics.newHashSet();
Set unlabeledDepSet = Generics.newHashSet();
for (TypedDependency dep : depCollection) {
unlabeledDepSet.add(new TypedDependencyStringEquality(null, dep.gov(), dep.dep()));
depSet.add(new TypedDependencyStringEquality(dep.reln(), dep.gov(), dep.dep()));
}
List> l = new ArrayList<>(2);
l.add(depSet);
l.add(unlabeledDepSet);
return l;
}
public DependencyScoring(List> goldDeps, boolean ignorePunc) {
this.goldDeps = new ArrayList<>(goldDeps.size());
this.goldDepsUnlabeled = new ArrayList<>(goldDeps.size());
this.ignorePunc = ignorePunc;
for (Collection depCollection : goldDeps) {
List> sets = toSets(depCollection);
this.goldDepsUnlabeled.add(sets.get(1));
this.goldDeps.add(sets.get(0));
}
if (ignorePunc) {
removeHeadsAssignedToPunc(this.goldDeps);
removeHeadsAssignedToPunc(this.goldDepsUnlabeled);
}
}
private static void removeHeadsAssignedToPunc(Set depSet) {
List deps = new ArrayList<>(depSet);
for (TypedDependency dep : deps) {
if (langIndependentPuncCheck(dep.dep().word())) {
if (VERBOSE) {
System.err.printf("Dropping Punctuation Dependency: %s\n", dep);
}
depSet.remove(dep);
}
}
}
private static void removeHeadsAssignedToPunc(List> depSets) {
for (Set depSet : depSets) {
removeHeadsAssignedToPunc(depSet);
}
}
public static boolean langIndependentPuncCheck(String token) {
boolean isNotWord = true;
for (int offset = 0; offset < token.length(); ) {
final int codepoint = token.codePointAt(offset);
if (Character.isLetterOrDigit(codepoint)) {
isNotWord = false;
}
offset += Character.charCount(codepoint);
}
return isNotWord;
}
public static DependencyScoring newInstanceStringEquality(List> goldDeps, boolean ignorePunc) {
return new DependencyScoring(convertStringEquality(goldDeps), ignorePunc);
}
public DependencyScoring(String filename, boolean CoNLLX, boolean ignorePunc) throws IOException {
this((CoNLLX ? readDepsCoNLLX(filename) : readDeps(filename)), ignorePunc);
}
public DependencyScoring(String filename) throws IOException {
this(filename, false, false);
}
public static List> convertStringEquality(List> deps){
List> convertedDeps = new ArrayList<>();
for(Collection depSet : deps){
Collection converted = Generics.newHashSet();
for(TypedDependency dep : depSet){
converted.add(new TypedDependencyStringEquality(dep.reln(), dep.gov(), dep.dep()));
}
convertedDeps.add(converted);
}
return convertedDeps;
}
static private class TypedDependencyStringEquality extends TypedDependency {
private static final long serialVersionUID = 1L;
public TypedDependencyStringEquality(GrammaticalRelation reln, IndexedWord gov, IndexedWord dep) {
super(reln, gov, dep);
}
public boolean equals(Object o) {
// some parsers, like Relex, screw up the casing
return o.toString().toLowerCase().equals(this.toString().toLowerCase());
}
@Override
public int hashCode() {
return toString().toLowerCase().hashCode();
}
}
/** Normalize all number tokens to <num> in order to allow
* for proper scoring of MSTParser productions.
*/
static protected String normalizeNumbers(String token) {
String norm = token.replaceFirst("^([0-9]+)-([0-9]+)$", "-$2");
if (!norm.equals(token)) {
System.err.printf("Normalized numbers in token: %s => %s\n", token, norm);
}
return token;
}
/**
* Read in typed dependencies in CoNLLX format.
*
* @param filename
* @throws IOException
*/
static protected List> readDepsCoNLLX(String filename) throws IOException {
List gss = GrammaticalStructure.readCoNLLXGrammaticalStructureCollection(filename,
new fakeShortNameToGRel(), new GraphLessGrammaticalStructureFactory());
List> readDeps = new ArrayList<>(gss.size());
for (GrammaticalStructure gs : gss) {
Collection deps = gs.typedDependencies();
readDeps.add(deps);
}
return readDeps;
}
/**
* Read in typed dependencies. Warning created typed dependencies are not
* backed by any sort of a tree structure.
*
* @param filename
* @throws IOException
*/
static protected List> readDeps(String filename) throws IOException {
LineNumberReader breader = new LineNumberReader(new FileReader(filename));
List> readDeps = new ArrayList<>();
Collection deps = new ArrayList<>();
for (String line = breader.readLine(); line != null; line = breader.readLine()) {
if (line.equals("null(-0,-0)") || line.equals("null(-1,-1)")) {
readDeps.add(deps);
deps = new ArrayList<>();
continue; // relex parse error
}
try {
if (line.equals("")) {
if (deps.size() != 0) {
//System.out.println(deps);
readDeps.add(deps);
deps = new ArrayList<>();
}
continue;
}
int firstParen = line.indexOf("(");
int commaSpace = line.indexOf(", ");
String depName = line.substring(0, firstParen);
String govName = line.substring(firstParen + 1, commaSpace);
String childName = line.substring(commaSpace+2, line.length() - 1);
GrammaticalRelation grel = GrammaticalRelation.valueOf(depName);
if (depName.startsWith("prep_")) {
String prep = depName.substring(5);
grel = EnglishGrammaticalRelations.getPrep(prep);
}
if (depName.startsWith("prepc_")) {
String prepc = depName.substring(6);
grel = EnglishGrammaticalRelations.getPrepC(prepc);
}
if (depName.startsWith("conj_")) {
String conj = depName.substring(5);
grel = EnglishGrammaticalRelations.getConj(conj);
}
if (grel == null) {
throw new RuntimeException("Unknown grammatical relation '" + depName+"'");
}
//Word govWord = new Word(govName.substring(0, govDash));
IndexedWord govWord = new IndexedWord();
govWord.setValue(normalizeNumbers(govName));
govWord.setWord(govWord.value());
//Word childWord = new Word(childName.substring(0, childDash));
IndexedWord childWord = new IndexedWord();
childWord.setValue(normalizeNumbers(childName));
childWord.setWord(childWord.value());
TypedDependency dep = new TypedDependencyStringEquality(grel, govWord, childWord);
deps.add(dep);
} catch (Exception e) {
breader.close();
throw new RuntimeException("Error on line "+breader.getLineNumber()+":\n\n"+e);
}
}
if (deps.size() != 0) {
readDeps.add(deps);
}
//log.info("last: "+readDeps.get(readDeps.size()-1));
breader.close();
return readDeps;
}
/**
* Score system typed dependencies
*
* @param system
* @return a triple consisting of (labeled attachment, unlabeled attachment,
* label accuracy)
*/
public Score score(List> system) {
int parserCnt = 0;
int goldCnt = 0;
int parserUnlabeledCnt = 0;
int goldUnlabeledCnt = 0;
int correctAttachment = 0;
int correctUnlabeledAttachment = 0;
int labelCnt = 0;
int labelCorrect = 0;
ClassicCounter unlabeledErrorCounts = new ClassicCounter<>();
ClassicCounter labeledErrorCounts = new ClassicCounter<>();
//System.out.println("Gold size: "+ goldDeps.size() + " System size: "+system.size());
for (int i = 0; i < system.size(); i++) {
List> l = toSets(system.get(i));
if (ignorePunc) {
removeHeadsAssignedToPunc(l.get(0));
removeHeadsAssignedToPunc(l.get(1));
}
parserCnt += l.get(0).size();
goldCnt += goldDeps.get(i).size();
parserUnlabeledCnt += l.get(1).size();
goldUnlabeledCnt += goldDepsUnlabeled.get(i).size();
l.get(0).retainAll(goldDeps.get(i));
l.get(1).retainAll(goldDepsUnlabeled.get(i));
correctAttachment += l.get(0).size();
correctUnlabeledAttachment += l.get(1).size();
labelCnt += l.get(1).size();
labelCorrect += l.get(0).size();
//System.out.println(""+i+" Acc: "+(l.get(0).size())/(double)localCnt+" "+l.get(0).size()+"/"+localCnt);
// identify errors
List> errl = toSets(system.get(i));
errl.get(0).removeAll(goldDeps.get(i));
errl.get(1).removeAll(goldDepsUnlabeled.get(i));
Map childCorrectWithLabel = Generics.newHashMap();
Map childCorrectWithOutLabel = Generics.newHashMap();
for (TypedDependency goldDep: goldDeps.get(i)) {
//System.out.print(goldDep);
String sChild = goldDep.dep().toString().replaceFirst("-[^-]*$", "");
String prefixLabeled = "";
String prefixUnlabeled = "";
if (childCorrectWithLabel.containsKey(sChild)) {
prefixLabeled = childCorrectWithLabel.get(sChild)+", ";
prefixUnlabeled = childCorrectWithOutLabel.get(sChild)+", ";
}
childCorrectWithLabel.put(sChild, prefixLabeled + goldDep.reln()+"("+goldDep.gov().toString().replaceFirst("-[^-]*$", "")+", "+sChild+")");
childCorrectWithOutLabel.put(sChild, prefixUnlabeled + "dep("+goldDep.gov().toString().replaceFirst("-[^-]*$", "")+", "+sChild+")");
}
for (TypedDependency labeledError: errl.get(0)) {
String sChild = labeledError.dep().toString().replaceFirst("-[^-]*$", "");
String sGov = labeledError.gov().toString().replaceFirst("-[^-]*$", "");
labeledErrorCounts.incrementCount(labeledError.reln().toString()+"("+sGov+", "+sChild+") <= "+childCorrectWithLabel.get(sChild));
}
for (TypedDependency unlabeledError: errl.get(1)) {
String sChild = unlabeledError.dep().toString().replaceFirst("-[^-]*$", "");
String sGov = unlabeledError.gov().toString().replaceFirst("-[^-]*$", "");
unlabeledErrorCounts.incrementCount("dep("+sGov+", "+sChild+") <= "+childCorrectWithOutLabel.get(sChild));
}
}
return new Score(parserCnt, goldCnt, parserUnlabeledCnt, goldUnlabeledCnt, correctAttachment, correctUnlabeledAttachment, labelCnt, labelCorrect, labeledErrorCounts, unlabeledErrorCounts);
}
public static class Score {
final int parserCnt;
final int goldCnt;
final int parserUnlabeledCnt;
final int goldUnlabeledCnt;
final int correctAttachment;
final int correctUnlabeledAttachment;
final int labelCnt;
final int labelCorrect;
final ClassicCounter unlabeledErrorCounts;
final ClassicCounter labeledErrorCounts;
public Score(int parserCnt, int goldCnt, int parserUnlabeledCnt, int goldUnlabeledCnt, int correctAttachment, int correctUnlabeledAttachment, int labelCnt, int labelCorrect, ClassicCounter labeledErrorCounts, ClassicCounter unlabeledErrorCounts) {
this.parserCnt = parserCnt;
this.goldCnt = goldCnt;
this.parserUnlabeledCnt = parserUnlabeledCnt;
this.goldUnlabeledCnt = goldUnlabeledCnt;
this.correctAttachment = correctAttachment;
this.correctUnlabeledAttachment = correctUnlabeledAttachment;
this.labelCnt = labelCnt;
this.labelCorrect = labelCorrect;
this.unlabeledErrorCounts = new ClassicCounter<>(unlabeledErrorCounts);
this.labeledErrorCounts = new ClassicCounter<>(labeledErrorCounts);
}
public String toString() {
return toStringFScore(false, false);
}
public String toStringAttachmentScore(boolean json) {
if (parserCnt != goldCnt) {
throw new RuntimeException(
String.format("AttachmentScore cannot be used when count(gold deps:%d) != count(system deps:%d)", parserCnt, goldCnt));
}
double las = correctAttachment/(double)goldCnt;
double uas = correctUnlabeledAttachment/(double)goldCnt;
StringBuilder sbuild = new StringBuilder();
if (json) {
sbuild.append("{");
sbuild.append(String.format("'LAS' : %.3f, ", las));
sbuild.append(String.format("'UAS' : %.3f, ", uas));
sbuild.append("}");
} else {
sbuild.append(String.format("|| Labeled Attachment Score ||"));
sbuild.append(String.format(" %.3f (%d/%d) ||\n", las, correctAttachment, goldCnt));
sbuild.append(String.format("|| Unlabeled Attachment Score ||"));
sbuild.append(String.format(" %.3f (%d/%d) ||\n", uas, correctUnlabeledAttachment, goldCnt));
}
return sbuild.toString();
}
public String toStringFScore(boolean verbose, boolean json) {
double lp = correctAttachment/(double)parserCnt;
double lr = correctAttachment/(double)goldCnt;
double lf = 2.0*(lp*lr)/(lp+lr);
/*sbuild.append(String.format("Labeled Attachment P: %.3f (%d/%d)\n", correctAttachment/(double)parserCnt, correctAttachment, parserCnt));
sbuild.append(String.format("Labeled Attachment R: %.3f (%d/%d)\n", correctAttachment/(double)goldCnt, correctAttachment, goldCnt));
*/
double ulp = correctUnlabeledAttachment/(double)parserUnlabeledCnt;
double ulr = correctUnlabeledAttachment/(double)goldUnlabeledCnt;
double ulf = 2.0*(ulp*ulr)/(ulp+ulr);
/*
sbuild.append(String.format("Unlabeled Attachment P: %.3f (%d/%d)\n", correctUnlabeledAttachment/(double)parserCnt, correctUnlabeledAttachment, parserCnt));
sbuild.append(String.format("Unlabeled Attachment R: %.3f (%d/%d)\n", correctUnlabeledAttachment/(double)goldCnt, correctUnlabeledAttachment, goldCnt));
sbuild.append(String.format("LabelAccuracy: %.3f (%d/%d)\n", labelCorrect/(double)labelCnt, labelCorrect, labelCnt));
*/
StringBuilder sbuild = new StringBuilder();
if (json) {
sbuild.append("{");
sbuild.append(String.format("'LF1' : %.3f, ", lf));
sbuild.append(String.format("'LP' : %.3f, ", lp));
sbuild.append(String.format("'LR' : %.3f, ", lr));
sbuild.append(String.format("'UF1' : %.3f, ", ulf));
sbuild.append(String.format("'UP' : %.3f, ", ulp));
sbuild.append(String.format("'UR' : %.3f, ", ulr));
sbuild.append("}");
} else {
sbuild.append(String.format("|| Labeled Attachment || F || P || R ||\n"));
sbuild.append(String.format("|| || %.3f || %.3f (%d/%d) || %.3f (%d/%d)||\n",
lf, lp, correctAttachment, parserCnt, lr, correctAttachment, goldCnt));
sbuild.append(String.format("|| Unlabeled Attachment || F || P || R ||\n"));
sbuild.append(String.format("|| || %.3f || %.3f (%d/%d) || %.3f (%d/%d)||\n",
ulf, ulp, correctUnlabeledAttachment, parserCnt, ulr, correctUnlabeledAttachment, goldCnt));
if (verbose) {
sbuild.append("\nLabeled Attachment Error Counts\n");
sbuild.append(Counters.toSortedString(labeledErrorCounts, Integer.MAX_VALUE, "\t%2$f\t%1$s", "\n"));
sbuild.append("\n");
sbuild.append("\nUnlabeled Attachment Error Counts\n");
sbuild.append(Counters.toSortedString(unlabeledErrorCounts, Integer.MAX_VALUE, "\t%2$f\t%1$s", "\n"));
}
}
return sbuild.toString();
}
} // end static class Score
public static void main(String[] args) throws IOException {
Properties props = StringUtils.argsToProperties(args);
boolean verbose = Boolean.parseBoolean(props.getProperty("v", "False"));
boolean conllx = Boolean.parseBoolean(props.getProperty("conllx", "False"));
boolean jsonOutput = Boolean.parseBoolean(props.getProperty("jsonOutput", "False"));
boolean ignorePunc = Boolean.parseBoolean(props.getProperty("nopunc", "False"));
String goldFilename = props.getProperty("g");
String systemFilename = props.getProperty("s");
if (goldFilename == null || systemFilename == null) {
log.info("Usage:\n\tjava ...DependencyScoring [-v True/False] [-conllx True/False] [-jsonOutput True/False] [-ignorePunc True/False] -g goldFile -s systemFile\n");
log.info("\nOptions:\n\t-v verbose output");
System.exit(-1);
}
DependencyScoring goldScorer = new DependencyScoring(goldFilename, conllx, ignorePunc);
List> systemDeps;
if (conllx) {
systemDeps = DependencyScoring.readDepsCoNLLX(systemFilename);
} else {
systemDeps = DependencyScoring.readDeps(systemFilename);
}
Score score = goldScorer.score(systemDeps);
if (conllx) {
System.out.println(score.toStringAttachmentScore(jsonOutput));
} else {
System.out.println(score.toStringFScore(verbose,jsonOutput));
}
}
}
class GraphLessGrammaticalStructureFactory implements GrammaticalStructureFromDependenciesFactory {
public GrammaticalStructure build(
List projectiveDependencies, TreeGraphNode root) {
return new GraphLessGrammaticalStructure(projectiveDependencies, root);
}
}
class GraphLessGrammaticalStructure extends GrammaticalStructure {
private static final long serialVersionUID = 1L;
public GraphLessGrammaticalStructure(
List projectiveDependencies, TreeGraphNode root) {
super(projectiveDependencies, root);
}
}
class fakeShortNameToGRel implements Map{
public void clear() {
throw new UnsupportedOperationException();
}
public boolean containsKey(Object o) {
// since we generate grammatical relations dynamically, this "map" technically contains any String key
if (o instanceof String) {
return true;
} else {
return false;
}
}
public boolean containsValue(Object o) {
throw new UnsupportedOperationException();
}
public Set> entrySet() {
throw new UnsupportedOperationException();
}
public GrammaticalRelation get(Object key) {
if (!(key instanceof String)) {
throw new UnsupportedOperationException();
}
String strkey = (String)key;
return new GrammaticalRelation(Language.Any, strkey, null, DEPENDENT) {
private static final long serialVersionUID = 1L;
@Override
public boolean equals(Object o) {
if (o instanceof GrammaticalRelation) {
return this.getShortName().equals(((GrammaticalRelation)o).getShortName());
}
return false;
}
@Override
public int hashCode() {
return this.getShortName().hashCode();
}
};
}
public boolean isEmpty() {
return false;
}
public Set keySet() {
throw new UnsupportedOperationException();
}
public GrammaticalRelation put(String key, GrammaticalRelation value) {
throw new UnsupportedOperationException();
}
public void putAll(Map m) {
throw new UnsupportedOperationException();
}
public GrammaticalRelation remove(Object key) {
throw new UnsupportedOperationException();
}
public int size() {
throw new UnsupportedOperationException();
}
public Collection values() {
throw new UnsupportedOperationException();
}
}