/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package opennlp.tools.parser.treeinsert;
import java.io.IOException;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.LinkedList;
import java.util.List;
import java.util.Map;
import java.util.Set;
import opennlp.tools.chunker.Chunker;
import opennlp.tools.chunker.ChunkerME;
import opennlp.tools.chunker.ChunkerModel;
import opennlp.tools.dictionary.Dictionary;
import opennlp.tools.ml.EventTrainer;
import opennlp.tools.ml.TrainerFactory;
import opennlp.tools.ml.model.Event;
import opennlp.tools.ml.model.MaxentModel;
import opennlp.tools.parser.AbstractBottomUpParser;
import opennlp.tools.parser.ChunkSampleStream;
import opennlp.tools.parser.HeadRules;
import opennlp.tools.parser.Parse;
import opennlp.tools.parser.ParserChunkerFactory;
import opennlp.tools.parser.ParserEventTypeEnum;
import opennlp.tools.parser.ParserModel;
import opennlp.tools.parser.ParserType;
import opennlp.tools.parser.PosSampleStream;
import opennlp.tools.postag.POSModel;
import opennlp.tools.postag.POSTagger;
import opennlp.tools.postag.POSTaggerFactory;
import opennlp.tools.postag.POSTaggerME;
import opennlp.tools.util.ObjectStream;
import opennlp.tools.util.TrainingParameters;
/**
* Built/attach parser. Nodes are built when their left-most
* child is encountered. Subsequent children are attached as
* daughters. Attachment is based on node in the right-frontier
* of the tree. After each attachment or building, nodes are
* assesed as either complete or incomplete. Complete nodes
* are no longer elligable for daughter attachment.
* Complex modifiers which produce additional node
* levels of the same type are attached with sister-adjunction.
* Attachment can not take place higher in the right-frontier
* than an incomplete node.
*/
public class Parser extends AbstractBottomUpParser {
/** Outcome used when a constituent needs an no additional parent node/building. */
public static final String DONE = "d";
/** Outcome used when a node should be attached as a sister to another node. */
public static final String ATTACH_SISTER = "s";
/** Outcome used when a node should be attached as a daughter to another node. */
public static final String ATTACH_DAUGHTER = "d";
/** Outcome used when a node should not be attached to another node. */
public static final String NON_ATTACH = "n";
/** Label used to distinguish build nodes from non-built nodes. */
public static final String BUILT = "built";
private MaxentModel buildModel;
private MaxentModel attachModel;
private MaxentModel checkModel;
static boolean checkComplete = false;
private BuildContextGenerator buildContextGenerator;
private AttachContextGenerator attachContextGenerator;
private CheckContextGenerator checkContextGenerator;
private double[] bprobs;
private double[] aprobs;
private double[] cprobs;
private int doneIndex;
private int sisterAttachIndex;
private int daughterAttachIndex;
private int nonAttachIndex;
private int completeIndex;
private int[] attachments;
public Parser(ParserModel model, int beamSize, double advancePercentage) {
this(model.getBuildModel(), model.getAttachModel(), model.getCheckModel(),
new POSTaggerME(model.getParserTaggerModel()),
new ChunkerME(model.getParserChunkerModel()),
model.getHeadRules(),
beamSize, advancePercentage);
}
public Parser(ParserModel model) {
this(model, defaultBeamSize, defaultAdvancePercentage);
}
private Parser(MaxentModel buildModel, MaxentModel attachModel, MaxentModel checkModel,
POSTagger tagger, Chunker chunker, HeadRules headRules, int beamSize,
double advancePercentage) {
super(tagger,chunker,headRules,beamSize,advancePercentage);
this.buildModel = buildModel;
this.attachModel = attachModel;
this.checkModel = checkModel;
this.buildContextGenerator = new BuildContextGenerator();
this.attachContextGenerator = new AttachContextGenerator(punctSet);
this.checkContextGenerator = new CheckContextGenerator(punctSet);
this.bprobs = new double[buildModel.getNumOutcomes()];
this.aprobs = new double[attachModel.getNumOutcomes()];
this.cprobs = new double[checkModel.getNumOutcomes()];
this.doneIndex = buildModel.getIndex(DONE);
this.sisterAttachIndex = attachModel.getIndex(ATTACH_SISTER);
this.daughterAttachIndex = attachModel.getIndex(ATTACH_DAUGHTER);
this.nonAttachIndex = attachModel.getIndex(NON_ATTACH);
attachments = new int[] {daughterAttachIndex,sisterAttachIndex};
this.completeIndex = checkModel.getIndex(Parser.COMPLETE);
}
/**
* Returns the right frontier of the specified parse tree with nodes ordered from deepest
* to shallowest.
* @param root The root of the parse tree.
* @return The right frontier of the specified parse tree.
*/
public static List<Parse> getRightFrontier(Parse root,Set<String> punctSet) {
List<Parse> rf = new LinkedList<>();
Parse top;
if (AbstractBottomUpParser.TOP_NODE.equals(root.getType()) ||
AbstractBottomUpParser.INC_NODE.equals(root.getType())) {
top = collapsePunctuation(root.getChildren(),punctSet)[0];
}
else {
top = root;
}
while (!top.isPosTag()) {
rf.add(0,top);
Parse[] kids = top.getChildren();
top = kids[kids.length - 1];
}
return new ArrayList<>(rf);
}
private void setBuilt(Parse p) {
String l = p.getLabel();
if (l == null) {
p.setLabel(Parser.BUILT);
}
else {
if (isComplete(p)) {
p.setLabel(Parser.BUILT + "." + Parser.COMPLETE);
}
else {
p.setLabel(Parser.BUILT + "." + Parser.INCOMPLETE);
}
}
}
private void setComplete(Parse p) {
String l = p.getLabel();
if (!isBuilt(p)) {
p.setLabel(Parser.COMPLETE);
}
else {
p.setLabel(Parser.BUILT + "." + Parser.COMPLETE);
}
}
private void setIncomplete(Parse p) {
if (!isBuilt(p)) {
p.setLabel(Parser.INCOMPLETE);
}
else {
p.setLabel(Parser.BUILT + "." + Parser.INCOMPLETE);
}
}
private boolean isBuilt(Parse p) {
String l = p.getLabel();
return l != null && l.startsWith(Parser.BUILT);
}
private boolean isComplete(Parse p) {
String l = p.getLabel();
return l != null && l.endsWith(Parser.COMPLETE);
}
@Override
protected Parse[] advanceChunks(Parse p, double minChunkScore) {
Parse[] parses = super.advanceChunks(p, minChunkScore);
for (Parse parse : parses) {
Parse[] chunks = parse.getChildren();
for (int ci = 0; ci < chunks.length; ci++) {
setComplete(chunks[ci]);
}
}
return parses;
}
@Override
protected Parse[] advanceParses(Parse p, double probMass) {
double q = 1 - probMass;
/* The index of the node which will be labeled in this iteration of advancing the parse. */
int advanceNodeIndex;
/* The node which will be labeled in this iteration of advancing the parse. */
Parse advanceNode = null;
Parse[] originalChildren = p.getChildren();
Parse[] children = collapsePunctuation(originalChildren,punctSet);
int numNodes = children.length;
if (numNodes == 0) {
return null;
}
else if (numNodes == 1) { //put sentence initial and final punct in top node
if (children[0].isPosTag()) {
return null;
}
else {
p.expandTopNode(children[0]);
return new Parse[] { p };
}
}
//determines which node needs to adanced.
for (advanceNodeIndex = 0; advanceNodeIndex < numNodes; advanceNodeIndex++) {
advanceNode = children[advanceNodeIndex];
if (!isBuilt(advanceNode)) {
break;
}
}
int originalZeroIndex = mapParseIndex(0,children,originalChildren);
int originalAdvanceIndex = mapParseIndex(advanceNodeIndex,children,originalChildren);
List<Parse> newParsesList = new ArrayList<>();
//call build model
buildModel.eval(buildContextGenerator.getContext(children, advanceNodeIndex), bprobs);
double doneProb = bprobs[doneIndex];
if (debugOn)
System.out.println("adi=" + advanceNodeIndex + " " + advanceNode.getType() + "."
+ advanceNode.getLabel() + " " + advanceNode + " choose build=" + (1 - doneProb)
+ " attach=" + doneProb);
if (1 - doneProb > q) {
double bprobSum = 0;
while (bprobSum < probMass) {
/* The largest unadvanced labeling. */
int max = 0;
for (int pi = 1; pi < bprobs.length; pi++) { //for each build outcome
if (bprobs[pi] > bprobs[max]) {
max = pi;
}
}
if (bprobs[max] == 0) {
break;
}
double bprob = bprobs[max];
bprobs[max] = 0; //zero out so new max can be found
bprobSum += bprob;
String tag = buildModel.getOutcome(max);
if (!tag.equals(DONE)) {
Parse newParse1 = (Parse) p.clone();
Parse newNode = new Parse(p.getText(),advanceNode.getSpan(),tag,bprob,advanceNode.getHead());
newParse1.insert(newNode);
newParse1.addProb(Math.log(bprob));
newParsesList.add(newParse1);
if (checkComplete) {
cprobs = checkModel.eval(checkContextGenerator.getContext(newNode, children,
advanceNodeIndex,false));
if (debugOn) System.out.println("building " + tag + " " + bprob + " c=" + cprobs[completeIndex]);
if (cprobs[completeIndex] > probMass) { //just complete advances
setComplete(newNode);
newParse1.addProb(Math.log(cprobs[completeIndex]));
if (debugOn) System.out.println("Only advancing complete node");
}
else if (1 - cprobs[completeIndex] > probMass) { //just incomplete advances
setIncomplete(newNode);
newParse1.addProb(Math.log(1 - cprobs[completeIndex]));
if (debugOn) System.out.println("Only advancing incomplete node");
}
else { //both complete and incomplete advance
if (debugOn) System.out.println("Advancing both complete and incomplete nodes");
setComplete(newNode);
newParse1.addProb(Math.log(cprobs[completeIndex]));
Parse newParse2 = (Parse) p.clone();
Parse newNode2 = new Parse(p.getText(),advanceNode.getSpan(),tag,bprob,advanceNode.getHead());
newParse2.insert(newNode2);
newParse2.addProb(Math.log(bprob));
newParsesList.add(newParse2);
newParse2.addProb(Math.log(1 - cprobs[completeIndex]));
setIncomplete(newNode2); //set incomplete for non-clone
}
}
else {
if (debugOn) System.out.println("building " + tag + " " + bprob);
}
}
}
}
//advance attaches
if (doneProb > q) {
Parse newParse1 = (Parse) p.clone(); //clone parse
//mark nodes as built
if (checkComplete) {
if (isComplete(advanceNode)) {
//replace constituent being labeled to create new derivation
newParse1.setChild(originalAdvanceIndex,Parser.BUILT + "." + Parser.COMPLETE);
}
else {
//replace constituent being labeled to create new derivation
newParse1.setChild(originalAdvanceIndex,Parser.BUILT + "." + Parser.INCOMPLETE);
}
}
else {
//replace constituent being labeled to create new derivation
newParse1.setChild(originalAdvanceIndex,Parser.BUILT);
}
newParse1.addProb(Math.log(doneProb));
if (advanceNodeIndex == 0) { //no attach if first node.
newParsesList.add(newParse1);
}
else {
List<Parse> rf = getRightFrontier(p,punctSet);
for (int fi = 0,fs = rf.size(); fi < fs; fi++) {
Parse fn = rf.get(fi);
attachModel.eval(attachContextGenerator.getContext(children, advanceNodeIndex, rf, fi), aprobs);
if (debugOn) {
// List cs = java.util.Arrays.asList(attachContextGenerator.getContext(children,
// advanceNodeIndex,rf,fi,punctSet));
System.out.println("Frontier node(" + fi + "): " + fn.getType() + "." + fn.getLabel()
+ " " + fn + " <- " + advanceNode.getType() + " " + advanceNode + " d="
+ aprobs[daughterAttachIndex] + " s=" + aprobs[sisterAttachIndex] + " ");
}
for (int ai = 0; ai < attachments.length; ai++) {
double prob = aprobs[attachments[ai]];
//should we try an attach if p > threshold and
// if !checkComplete then prevent daughter attaching to chunk
// if checkComplete then prevent daughter attacing to complete node or
// sister attaching to an incomplete node
if (prob > q && (
(!checkComplete && (attachments[ai] != daughterAttachIndex || !isComplete(fn)))
||
(checkComplete && ((attachments[ai] == daughterAttachIndex && !isComplete(fn))
|| (attachments[ai] == sisterAttachIndex && isComplete(fn)))))) {
Parse newParse2 = newParse1.cloneRoot(fn,originalZeroIndex);
Parse[] newKids = Parser.collapsePunctuation(newParse2.getChildren(),punctSet);
//remove node from top level since were going to attach it (including punct)
for (int ri = originalZeroIndex + 1; ri <= originalAdvanceIndex; ri++) {
//System.out.println(at"-removing "+(originalZeroIndex+1)+" "
// +newParse2.getChildren()[originalZeroIndex+1]);
newParse2.remove(originalZeroIndex + 1);
}
List<Parse> crf = getRightFrontier(newParse2,punctSet);
Parse updatedNode;
if (attachments[ai] == daughterAttachIndex) { //attach daughter
updatedNode = crf.get(fi);
updatedNode.add(advanceNode,headRules);
}
else { //attach sister
Parse psite;
if (fi + 1 < crf.size()) {
psite = crf.get(fi + 1);
updatedNode = psite.adjoin(advanceNode,headRules);
}
else {
psite = newParse2;
updatedNode = psite.adjoinRoot(advanceNode,headRules,originalZeroIndex);
newKids[0] = updatedNode;
}
}
//update spans affected by attachment
for (int ni = fi + 1; ni < crf.size(); ni++) {
Parse node = crf.get(ni);
node.updateSpan();
}
//if (debugOn) {System.out.print(ai+"-result: ");newParse2.show();System.out.println();}
newParse2.addProb(Math.log(prob));
newParsesList.add(newParse2);
if (checkComplete) {
cprobs = checkModel.eval(
checkContextGenerator.getContext(updatedNode,newKids,advanceNodeIndex,true));
if (cprobs[completeIndex] > probMass) {
setComplete(updatedNode);
newParse2.addProb(Math.log(cprobs[completeIndex]));
if (debugOn) System.out.println("Only advancing complete node");
}
else if (1 - cprobs[completeIndex] > probMass) {
setIncomplete(updatedNode);
newParse2.addProb(Math.log(1 - cprobs[completeIndex]));
if (debugOn) System.out.println("Only advancing incomplete node");
}
else {
setComplete(updatedNode);
Parse newParse3 = newParse2.cloneRoot(updatedNode,originalZeroIndex);
newParse3.addProb(Math.log(cprobs[completeIndex]));
newParsesList.add(newParse3);
setIncomplete(updatedNode);
newParse2.addProb(Math.log(1 - cprobs[completeIndex]));
if (debugOn)
System.out.println("Advancing both complete and incomplete nodes; c="
+ cprobs[completeIndex]);
}
}
}
else {
if (debugOn)
System.out.println("Skipping " + fn.getType() + "." + fn.getLabel() + " "
+ fn + " daughter=" + (attachments[ai] == daughterAttachIndex)
+ " complete=" + isComplete(fn) + " prob=" + prob);
}
}
if (checkComplete && !isComplete(fn)) {
if (debugOn)
System.out.println("Stopping at incomplete node(" + fi + "): "
+ fn.getType() + "." + fn.getLabel() + " " + fn);
break;
}
}
}
}
Parse[] newParses = new Parse[newParsesList.size()];
newParsesList.toArray(newParses);
return newParses;
}
@Override
protected void advanceTop(Parse p) {
p.setType(TOP_NODE);
}
public static ParserModel train(String languageCode,
ObjectStream<Parse> parseSamples, HeadRules rules, TrainingParameters mlParams)
throws IOException {
Map<String, String> manifestInfoEntries = new HashMap<>();
System.err.println("Building dictionary");
Dictionary mdict = buildDictionary(parseSamples, rules, mlParams);
parseSamples.reset();
// tag
POSModel posModel = POSTaggerME.train(languageCode, new PosSampleStream(
parseSamples), mlParams.getParameters("tagger"), new POSTaggerFactory());
parseSamples.reset();
// chunk
ChunkerModel chunkModel = ChunkerME.train(languageCode, new ChunkSampleStream(
parseSamples), mlParams.getParameters("chunker"), new ParserChunkerFactory());
parseSamples.reset();
// build
System.err.println("Training builder");
ObjectStream<Event> bes = new ParserEventStream(parseSamples, rules,
ParserEventTypeEnum.BUILD, mdict);
Map<String, String> buildReportMap = new HashMap<>();
EventTrainer buildTrainer = TrainerFactory.getEventTrainer(
mlParams.getParameters("build"), buildReportMap);
MaxentModel buildModel = buildTrainer.train(bes);
opennlp.tools.parser.chunking.Parser.mergeReportIntoManifest(
manifestInfoEntries, buildReportMap, "build");
parseSamples.reset();
// check
System.err.println("Training checker");
ObjectStream<Event> kes = new ParserEventStream(parseSamples, rules,
ParserEventTypeEnum.CHECK);
Map<String, String> checkReportMap = new HashMap<>();
EventTrainer checkTrainer = TrainerFactory.getEventTrainer(
mlParams.getParameters("check"), checkReportMap);
MaxentModel checkModel = checkTrainer.train(kes);
opennlp.tools.parser.chunking.Parser.mergeReportIntoManifest(
manifestInfoEntries, checkReportMap, "check");
parseSamples.reset();
// attach
System.err.println("Training attacher");
ObjectStream<Event> attachEvents = new ParserEventStream(parseSamples, rules,
ParserEventTypeEnum.ATTACH);
Map<String, String> attachReportMap = new HashMap<>();
EventTrainer attachTrainer = TrainerFactory.getEventTrainer(
mlParams.getParameters("attach"), attachReportMap);
MaxentModel attachModel = attachTrainer.train(attachEvents);
opennlp.tools.parser.chunking.Parser.mergeReportIntoManifest(
manifestInfoEntries, attachReportMap, "attach");
// TODO: Remove cast for HeadRules
return new ParserModel(languageCode, buildModel, checkModel,
attachModel, posModel, chunkModel,
rules, ParserType.TREEINSERT, manifestInfoEntries);
}
public static ParserModel train(String languageCode,
ObjectStream<Parse> parseSamples, HeadRules rules, int iterations, int cut)
throws IOException {
TrainingParameters params = new TrainingParameters();
params.put("dict", TrainingParameters.CUTOFF_PARAM, cut);
params.put("tagger", TrainingParameters.CUTOFF_PARAM, cut);
params.put("tagger", TrainingParameters.ITERATIONS_PARAM, iterations);
params.put("chunker", TrainingParameters.CUTOFF_PARAM, cut);
params.put("chunker", TrainingParameters.ITERATIONS_PARAM, iterations);
params.put("check", TrainingParameters.CUTOFF_PARAM, cut);
params.put("check", TrainingParameters.ITERATIONS_PARAM, iterations);
params.put("build", TrainingParameters.CUTOFF_PARAM, cut);
params.put("build", TrainingParameters.ITERATIONS_PARAM, iterations);
return train(languageCode, parseSamples, rules, params);
}
}