/*
* Copyright 1999-2002 Carnegie Mellon University.
* Portions Copyright 2002 Sun Microsystems, Inc.
* Portions Copyright 2002 Mitsubishi Electric Research Laboratories.
* All Rights Reserved. Use is subject to license terms.
*
* See the file "license.terms" for information on usage and
* redistribution of this file, and for a DISCLAIMER OF ALL
* WARRANTIES.
*
*/
package edu.cmu.sphinx.linguist.lextree;
import edu.cmu.sphinx.linguist.WordSequence;
import edu.cmu.sphinx.linguist.acoustic.HMM;
import edu.cmu.sphinx.linguist.acoustic.HMMPool;
import edu.cmu.sphinx.linguist.acoustic.HMMPosition;
import edu.cmu.sphinx.linguist.acoustic.Unit;
import edu.cmu.sphinx.linguist.dictionary.Dictionary;
import edu.cmu.sphinx.linguist.dictionary.Pronunciation;
import edu.cmu.sphinx.linguist.dictionary.Word;
import edu.cmu.sphinx.linguist.language.ngram.LanguageModel;
import edu.cmu.sphinx.util.LogMath;
import edu.cmu.sphinx.util.TimerPool;
import edu.cmu.sphinx.util.Utilities;
import java.util.*;
import java.util.logging.Logger;
/** Represents a node in the HMM Tree */
// For large vocabularies we may create millions of these objects,
// therefore they are extremely space sensitive. So we want to make
// these objects as small as possible. The requirements for these
// objects when building the tree of nodes are very different from once
// we have built it. When building, we need to easily add successor
// nodes and quickly identify duplicate children nodes. After the tree
// is built we just need to quickly identify successors. We want the
// flexibility of a map to manage successors at startup, but we don't
// want the space penalty (at least 5 32 bit fields per map), instead
// we'd like an array. To support this dual mode, we manage the
// successors in an Object which can either be a Map or a List
// depending upon whether the node has been frozen or not.
class Node {
private static int nodeCount;
private static int successorCount;
private static final Map<Pronunciation, WordNode> wordNodeMap = new HashMap<Pronunciation, WordNode>();
/**
* This can be either Map during tree construction or Array after
* tree freeze. Conversion to array helps to save memory.
*/
private Object successors;
private float logUnigramProbability;
/**
* Creates a node
*
* @param probability the unigram probability for the node
*/
Node(float probability) {
logUnigramProbability = probability;
nodeCount++;
// if ((nodeCount % 10000) == 0) {
// System.out.println("NC " + nodeCount);
// }
}
/**
* Returns the unigram probability
*
* @return the unigram probability
*/
public float getUnigramProbability() {
return logUnigramProbability;
}
/**
* Sets the unigram probability
*
* @param probability the unigram probability
*/
public void setUnigramProbability(float probability) {
logUnigramProbability = probability;
}
/**
* Given an object get the set of successors for this object
*
* @param key the object key
* @return the node containing the successors
*/
private Node getSuccessor(Object key) {
Map<Object, Node> successors = getSuccessorMap();
return successors.get(key);
}
/**
* Add the child to the set of successors
*
* @param key the object key
* @param child the child to add
*/
void putSuccessor(Object key, Node child) {
Map<Object, Node> successors = getSuccessorMap();
successors.put(key, child);
}
/**
* Gets the successor map for this node
*
* @return the successor map
*/
@SuppressWarnings({"unchecked"})
public Map<Object, Node> getSuccessorMap() {
if (successors == null) {
successors = new HashMap<Object, Node>(4);
}
assert successors instanceof Map;
return (Map<Object, Node>) successors;
}
/** Freeze the node. Convert the successor map into an array list */
void freeze() {
if (successors instanceof Map<?,?>) {
Map<Object, Node> map = getSuccessorMap();
successors = map.values().toArray(new Node[map.size()]);
for (Node node : map.values()) {
node.freeze();
}
successorCount += map.size();
}
}
static void dumpNodeInfo() {
System.out.println("Nodes: " + nodeCount + " successors " +
successorCount + " avg " + (successorCount / nodeCount));
}
/**
* Adds a child node holding an hmm to the successor. If a node similar to the child has already been added, we use
* the previously added node, otherwise we add this. Also, we record the base unit of the child in the set of right
* context
*
* @param hmm the hmm to add\n * @return the node that holds the hmm (new or old)
*/
Node addSuccessor(HMM hmm, float probability) {
Node child = null;
Node matchingChild = getSuccessor(hmm);
if (matchingChild == null) {
child = new HMMNode(hmm, probability);
putSuccessor(hmm, child);
} else {
if (matchingChild.getUnigramProbability() < probability) {
matchingChild.setUnigramProbability(probability);
}
child = matchingChild;
}
return child;
}
/**
* Adds a child node holding a pronunciation to the successor. If a node similar to the child has already been
* added, we use the previously added node, otherwise we add this. Also, we record the base unit of the child in the
* set of right context
*
* @param pronunciation the pronunciation to add
* @return the node that holds the pronunciation (new or old)
*/
WordNode addSuccessor(Pronunciation pronunciation, float probability) {
WordNode child = null;
WordNode matchingChild = (WordNode) getSuccessor(pronunciation);
if (matchingChild == null) {
child = getWordNode(pronunciation, probability);
putSuccessor(pronunciation, child);
} else {
if (matchingChild.getUnigramProbability() < probability) {
matchingChild.setUnigramProbability(probability);
}
child = matchingChild;
}
return child;
}
void addSuccessor(WordNode wordNode) {
putSuccessor(wordNode, wordNode);
}
/**
* Adds an EndNode to the set of successors for this node If a node similar to the child has already been added, we
* use the previously added node, otherwise we add this.
*
* @param child the endNode to add
* @param probability probability for this transition
* @return the node that holds the endNode (new or old)
*/
EndNode addSuccessor(EndNode child, float probability) {
Unit baseUnit = child.getBaseUnit();
EndNode matchingChild = (EndNode) getSuccessor(baseUnit);
if (matchingChild == null) {
putSuccessor(baseUnit, child);
} else {
if (matchingChild.getUnigramProbability() < probability) {
matchingChild.setUnigramProbability(probability);
}
child = matchingChild;
}
return child;
}
/**
* Gets a word node associated with the pronunciation.
*
* @param p the pronunciation
* @return the word node
*/
private WordNode getWordNode(Pronunciation p, float probability) {
WordNode node = wordNodeMap.get(p);
if (node == null) {
node = new WordNode(p, probability);
wordNodeMap.put(p, node);
}
return node;
}
/**
* Adds a child node to the successor. If a node similar to the child has already been added, we use the previously
* added node, otherwise we add this. Also, we record the base unit of the child in the set of right context
*
* @param child the child to add
* @return the node (may be different than child if there was already a node attached holding the hmm held by
* child)
*/
UnitNode addSuccessor(UnitNode child) {
UnitNode matchingChild = (UnitNode) getSuccessor(child.getKey());
if (matchingChild == null) {
putSuccessor(child.getKey(), child);
} else {
child = matchingChild;
}
return child;
}
/**
* Returns the successors for this node
*
* @return the set of successor nodes
*/
Node[] getSuccessors() {
if (successors instanceof Map<?, ?>) {
freeze();
}
return (Node[])successors;
}
/**
* Returns the string representation for this object
*
* @return the string representation of the object
*/
@Override
public String toString() {
return "Node ";
}
}
/** A node representing a word in the HMM tree */
class WordNode extends Node {
private final Pronunciation pronunciation;
private final boolean isFinal;
/**
* Creates a word node
*
* @param pronunciation the pronunciation to wrap in this node
* @param probability the word unigram probability
*/
WordNode(Pronunciation pronunciation, float probability) {
super(probability);
this.pronunciation = pronunciation;
this.isFinal = pronunciation.getWord().isSentenceEndWord();
}
/**
* Gets the word associated with this node
*
* @return the word
*/
Word getWord() {
return pronunciation.getWord();
}
/**
* Gets the pronunciation associated with this node
*
* @return the pronunciation
*/
Pronunciation getPronunciation() {
return pronunciation;
}
/**
* Gets the last unit for this word
*
* @return the last unit
*/
Unit getLastUnit() {
Unit[] units = pronunciation.getUnits();
return units[units.length - 1];
}
/**
* Returns the successors for this node
*
* @return the set of successor nodes
*/
@Override
Node[] getSuccessors() {
throw new Error("Not supported");
}
/**
* Returns a string representation for this object
*
* @return a string representation
*/
@Override
public String toString() {
return "WordNode " + pronunciation + " p " +
getUnigramProbability();
}
public boolean isFinal() {
return isFinal;
}
}
/**
* A class that represents the initial word in the search space. It is treated specially because we need to keep track
* of the context as well. The context is embodied in the parent node
*/
class InitialWordNode extends WordNode {
final HMMNode parent;
/**
* Creates an InitialWordNode
*
* @param pronunciation the pronunciation
* @param parent the parent node
*/
InitialWordNode(Pronunciation pronunciation, HMMNode parent) {
super(pronunciation, LogMath.LOG_ONE);
this.parent = parent;
}
/**
* Gets the parent for this word node
*
* @return the parent
*/
HMMNode getParent() {
return parent;
}
}
abstract class UnitNode extends Node {
public final static int SIMPLE_UNIT = 1;
public final static int WORD_BEGINNING_UNIT = 2;
public final static int SILENCE_UNIT = 3;
public final static int FILLER_UNIT = 4;
private int type;
/**
* Creates the UnitNode
*
* @param probablilty the probability for the node
*/
UnitNode(float probablilty) {
super(probablilty);
}
/**
* Returns the base unit for this hmm node
*
* @return the base unit
*/
abstract Unit getBaseUnit();
abstract Object getKey();
abstract HMMPosition getPosition();
/**
* Gets the unit type (one of SIMPLE_UNIT, WORD_BEGINNING_UNIT, SIMPLE_UNIT or FILLER_UNIT
*
* @return the unit type
*/
int getType() {
return type;
}
/**
* Sets the unit type
*
* @param type the unit type
*/
void setType(int type) {
this.type = type;
}
}
/** A node that represents an HMM in the hmm tree */
class HMMNode extends UnitNode {
private final HMM hmm;
// There can potentially be a large number of nodes (millions),
// therefore it is important to conserve space as much as
// possible. While building the HMMNodes, we keep right contexts
// in a set to allow easy pruning of duplicates. Once the tree is
// entirely built, we no longer need to manage the right contexts
// as a set, a simple array will do. The freeze method converts
// the set to the array of units. This rcSet object holds the set
// during construction and the array after the freeze.
private Object rcSet;
/**
* Creates the node, wrapping the given hmm
*
* @param hmm the hmm to hold
*/
HMMNode(HMM hmm, float probablilty) {
super(probablilty);
this.hmm = hmm;
Unit base = getBaseUnit();
int type = SIMPLE_UNIT;
if (base.isSilence()) {
type = SILENCE_UNIT;
} else if (base.isFiller()) {
type = FILLER_UNIT;
} else if (hmm.getPosition().isWordBeginning()) {
type = WORD_BEGINNING_UNIT;
}
setType(type);
}
/**
* Returns the base unit for this hmm node
*
* @return the base unit
*/
@Override
Unit getBaseUnit() {
// return hmm.getUnit().getBaseUnit();
return hmm.getBaseUnit();
}
/**
* Returns the hmm for this node
*
* @return the hmm
*/
HMM getHMM() {
return hmm;
}
@Override
HMMPosition getPosition() {
return hmm.getPosition();
}
@Override
HMM getKey() {
return getHMM();
}
/**
* Returns a string representation for this object
*
* @return a string representation
*/
@Override
public String toString() {
return "HMMNode " + hmm + " p " + getUnigramProbability();
}
/**
* Adds a right context to the set of possible right contexts for this node. This is typically only needed for hmms
* at the ends of words.
*
* @param rc the right context.
*/
void addRC(Unit rc) {
getRCSet().add(rc);
}
/** Freeze this node. Convert the set into an array to reduce memory overhead */
@Override
@SuppressWarnings({"unchecked"})
void freeze() {
super.freeze();
if (rcSet instanceof Set) {
Set<Unit> set = (Set<Unit>) rcSet;
rcSet = set.toArray(new Unit[set.size()]);
}
}
/**
* Gets the rc as a set. If we've already been frozen it is an error
*
* @return the set of right contexts
*/
@SuppressWarnings({"unchecked"})
private Set<Unit> getRCSet() {
if (rcSet == null) {
rcSet = new HashSet<Unit>();
}
assert rcSet instanceof HashSet;
return (Set<Unit>) rcSet;
}
/**
* returns the set of right contexts for this node
*
* @return the set of right contexts
*/
Unit[] getRC() {
if (rcSet instanceof HashSet<?>) {
freeze();
}
return (Unit[]) rcSet;
}
}
class EndNode extends UnitNode {
final Unit baseUnit;
final Unit leftContext;
final Integer key;
/**
* Creates the node, wrapping the given hmm
*
* @param baseUnit the base unit for this node
* @param lc the left context
* @param probablilty the probability for the transition to this node
*/
EndNode(Unit baseUnit, Unit lc, float probablilty) {
super(probablilty);
this.baseUnit = baseUnit;
this.leftContext = lc;
key = baseUnit.getBaseID() * 121 + leftContext.getBaseID();
}
/**
* Returns the base unit for this hmm node
*
* @return the base unit
*/
@Override
Unit getBaseUnit() {
return baseUnit;
}
/**
* Returns the base unit for this hmm node
*
* @return the base unit
*/
Unit getLeftContext() {
return leftContext;
}
@Override
Integer getKey() {
return key;
}
@Override
HMMPosition getPosition() {
return HMMPosition.END;
}
/**
* Returns a string representation for this object
*
* @return a string representation
*/
@Override
public String toString() {
return "EndNode base:" + baseUnit + " lc " + leftContext + ' ' + key;
}
/** Freeze this node. Convert the set into an array to reduce memory overhead */
@Override
void freeze() {
super.freeze();
}
}
/**
* Represents the vocabulary as a lex tree with nodes in the tree representing either words (WordNode) or units
* (HMMNode). HMMNodes may be shared.
*/
class HMMTree {
private final HMMPool hmmPool;
private InitialWordNode initialNode;
private Dictionary dictionary;
private LanguageModel lm;
private final boolean addFillerWords;
private final boolean addSilenceWord = true;
private final Set<Unit> entryPoints = new HashSet<Unit>();
private Set<Unit> exitPoints = new HashSet<Unit>();
private Set<Word> allWords;
private EntryPointTable entryPointTable;
private boolean debug;
private final float languageWeight;
private final Map<Object, HMMNode[]> endNodeMap;
private WordNode sentenceEndWordNode;
private Logger logger;
/**
* Creates the HMMTree
*
* @param pool the pool of HMMs and units
* @param dictionary the dictionary containing the pronunciations
* @param lm the source of the set of words to add to the lex tree
* @param addFillerWords if <code>false</code> add filler words
* @param languageWeight the languageWeight
*/
HMMTree(HMMPool pool, Dictionary dictionary, LanguageModel lm,
boolean addFillerWords, float languageWeight) {
this.hmmPool = pool;
this.dictionary = dictionary;
this.lm = lm;
this.endNodeMap = new HashMap<Object, HMMNode[]>();
this.addFillerWords = addFillerWords;
this.languageWeight = languageWeight;
logger = Logger.getLogger(HMMTree.class.getSimpleName());
TimerPool.getTimer(this,"Create HMM Tree").start();
compile();
TimerPool.getTimer(this,"Create HMM Tree").stop();
}
/**
* Given a base unit and a left context, return the set of entry points into the lex tree
*
* @param lc the left context
* @param base the center unit
* @return the set of entry points
*/
public Node[] getEntryPoint(Unit lc, Unit base) {
EntryPoint ep = entryPointTable.getEntryPoint(base);
return ep.getEntryPointsFromLeftContext(lc).getSuccessors();
}
/**
* Gets the set of hmm nodes associated with the given end node
*
* @param endNode the end node
* @return an array of associated hmm nodes
*/
public HMMNode[] getHMMNodes(EndNode endNode) {
HMMNode[] results = endNodeMap.get(endNode.getKey());
if (results == null) {
// System.out.println("Filling cache for " + endNode.getKey()
// + " size " + endNodeMap.size());
Map<HMM, HMMNode> resultMap = new HashMap<HMM, HMMNode>();
Unit baseUnit = endNode.getBaseUnit();
Unit lc = endNode.getLeftContext();
for (Unit rc : entryPoints) {
HMM hmm = hmmPool.getHMM(baseUnit, lc, rc, HMMPosition.END);
HMMNode hmmNode = resultMap.get(hmm);
if (hmmNode == null) {
hmmNode = new HMMNode(hmm, LogMath.LOG_ONE);
resultMap.put(hmm, hmmNode);
}
hmmNode.addRC(rc);
for (Node node : endNode.getSuccessors()) {
WordNode wordNode = (WordNode)node;
hmmNode.addSuccessor(wordNode);
}
}
// cache it
results = resultMap.values().toArray(new HMMNode[resultMap.size()]);
endNodeMap.put(endNode.getKey(), results);
}
// System.out.println("GHN: " + endNode + " " + results.length);
return results;
}
/**
* Returns the word node associated with the sentence end word
*
* @return the sentence end word node
*/
public WordNode getSentenceEndWordNode() {
assert sentenceEndWordNode != null;
return sentenceEndWordNode;
}
// private Object getKey(EndNode endNode) {
// Unit base = endNode.getBaseUnit();
// Unit lc = endNode.getLeftContext();
// return null;
// }
/** Compiles the vocabulary into an HMM Tree */
private void compile() {
collectEntryAndExitUnits();
entryPointTable = new EntryPointTable(entryPoints);
addWords();
entryPointTable.createEntryPointMaps();
freeze();
}
/** Dumps the tree */
void dumpTree() {
System.out.println("Dumping Tree ...");
Map<Node, Node> dupNode = new HashMap<Node, Node>();
dumpTree(0, getInitialNode(), dupNode);
System.out.println("... done Dumping Tree");
}
/**
* Dumps the tree
*
* @param level the level of the dump
* @param node the root of the tree to dump
* @param dupNode map of visited nodes
*/
private void dumpTree(int level, Node node, Map<Node, Node> dupNode) {
if (dupNode.get(node) == null) {
dupNode.put(node, node);
System.out.println(Utilities.pad(level) + node);
if (!(node instanceof WordNode)) {
for (Node nextNode : node.getSuccessors()) {
dumpTree(level + 1, nextNode, dupNode);
}
}
}
}
/** Collects all of the entry and exit points for the vocabulary. */
private void collectEntryAndExitUnits() {
Collection<Word> words = getAllWords();
for (Word word : words) {
for (int j = 0; j < word.getPronunciations().length; j++) {
Pronunciation p = word.getPronunciations()[j];
Unit first = p.getUnits()[0];
Unit last = p.getUnits()[p.getUnits().length - 1];
entryPoints.add(first);
exitPoints.add(last);
}
}
if (debug) {
System.out.println("Entry Points: " + entryPoints.size());
System.out.println("Exit Points: " + exitPoints.size());
}
}
/**
* Called after the lex tree is built. Frees all temporary structures. After this is called, no more words can be
* added to the lex tree.
*/
private void freeze() {
entryPointTable.freeze();
dictionary = null;
lm = null;
exitPoints = null;
allWords = null;
}
/** Adds the given collection of words to the lex tree */
private void addWords() {
Set<Word> words = getAllWords();
for (Word word : words) {
addWord(word);
}
}
/**
* Adds a single word to the lex tree
*
* @param word the word to add
*/
private void addWord(Word word) {
float prob = getWordUnigramProbability(word);
Pronunciation[] pronunciations = word.getPronunciations();
for (Pronunciation pronunciation : pronunciations) {
addPronunciation(pronunciation, prob);
}
}
/**
* Adds the given pronunciation to the lex tree
*
* @param pronunciation the pronunciation
* @param probability the unigram probability
*/
private void addPronunciation(Pronunciation pronunciation,
float probability) {
Unit baseUnit;
Unit lc;
Unit rc;
Node curNode;
WordNode wordNode;
Unit[] units = pronunciation.getUnits();
baseUnit = units[0];
EntryPoint ep = entryPointTable.getEntryPoint(baseUnit);
ep.addProbability(probability);
if (units.length > 1) {
curNode = ep.getNode();
lc = baseUnit;
for (int i = 1; i < units.length - 1; i++) {
baseUnit = units[i];
rc = units[i + 1];
HMM hmm = hmmPool.getHMM(baseUnit, lc, rc, HMMPosition.INTERNAL);
if (hmm == null) {
logger.severe("Missing HMM for unit " + baseUnit.getName() + " with lc=" + lc.getName() + " rc=" + rc.getName());
} else {
curNode = curNode.addSuccessor(hmm, probability);
}
lc = baseUnit; // next lc is this baseUnit
}
// now add the last unit as an end unit
baseUnit = units[units.length - 1];
EndNode endNode = new EndNode(baseUnit, lc, probability);
curNode = curNode.addSuccessor(endNode, probability);
wordNode = curNode.addSuccessor(pronunciation, probability);
if (wordNode.getWord().isSentenceEndWord()) {
sentenceEndWordNode = wordNode;
}
} else {
ep.addSingleUnitWord(pronunciation);
}
}
/**
* Gets the unigram probability for the given word
*
* @param word the word
* @return the unigram probability for the word.
*/
private float getWordUnigramProbability(Word word) {
float prob = LogMath.LOG_ONE;
if (!word.isFiller()) {
Word[] wordArray = new Word[1];
wordArray[0] = word;
prob = lm.getProbability((new WordSequence(wordArray)));
// System.out.println("gwup: " + word + " " + prob);
prob *= languageWeight;
}
return prob;
}
/**
* Returns the entire set of words, including filler words
*
* @return the set of all words (as Word objects)
*/
private Set<Word> getAllWords() {
if (allWords == null) {
allWords = new HashSet<Word>();
for (String spelling : lm.getVocabulary()) {
Word word = dictionary.getWord(spelling);
if (word != null) {
allWords.add(word);
}
}
if (addFillerWords) {
allWords.addAll(Arrays.asList(dictionary.getFillerWords()));
} else if (addSilenceWord) {
allWords.add(dictionary.getSilenceWord());
}
}
return allWords;
}
/**
* Returns the initial node for this lex tree
*
* @return the initial lex node
*/
InitialWordNode getInitialNode() {
return initialNode;
}
/** The EntryPoint table is used to manage the set of entry points into the lex tree. */
class EntryPointTable {
private final Map<Unit, EntryPoint> entryPoints;
/**
* Create the entry point table give the set of all possible entry point units
*
* @param entryPointCollection the set of possible entry points
*/
EntryPointTable(Collection<Unit> entryPointCollection) {
entryPoints = new HashMap<Unit, EntryPoint>();
for (Unit unit : entryPointCollection) {
entryPoints.put(unit, new EntryPoint(unit));
}
}
/**
* Given a CI unit, return the EntryPoint object that manages the entry point for the unit
*
* @param baseUnit the unit of interest (A ci unit)
* @return the object that manages the entry point for the unit
*/
EntryPoint getEntryPoint(Unit baseUnit) {
return entryPoints.get(baseUnit);
}
/** Creates the entry point maps for all entry points. */
void createEntryPointMaps() {
for (EntryPoint ep : entryPoints.values()) {
ep.createEntryPointMap();
}
}
/** Freezes the entry point table */
void freeze() {
for (EntryPoint ep : entryPoints.values()) {
ep.freeze();
}
}
/** Dumps the entry point table */
void dump() {
for (EntryPoint ep : entryPoints.values()) {
ep.dump();
}
}
}
/** Manages a single entry point. */
class EntryPoint {
final Unit baseUnit;
final Node baseNode; // second units and beyond start here
final Map<Unit, Node> unitToEntryPointMap;
List<Pronunciation> singleUnitWords;
int nodeCount;
Set<Unit> rcSet;
float totalProbability;
/**
* Creates an entry point for the given unit
*
* @param baseUnit the EntryPoint is created for this unit
*/
EntryPoint(Unit baseUnit) {
this.baseUnit = baseUnit;
this.baseNode = new Node(LogMath.LOG_ZERO);
this.unitToEntryPointMap = new HashMap<Unit, Node>();
this.singleUnitWords = new ArrayList<Pronunciation>();
this.totalProbability = LogMath.LOG_ZERO;
}
/**
* Given a left context get a node that represents a single set of entry points into this unit
*
* @param leftContext the left context of interest
* @return the node representing the entry point
*/
Node getEntryPointsFromLeftContext(Unit leftContext) {
return unitToEntryPointMap.get(leftContext);
}
/**
* Accumulates the probability for this entry point
*
* @param probability a new probability
*/
void addProbability(float probability) {
if (probability > totalProbability) {
totalProbability = probability;
}
}
/**
* Returns the probability for all words reachable from this node
*
* @return the log probability
*/
float getProbability() {
return totalProbability;
}
/** Once we have built the full entry point we can eliminate some fields */
void freeze() {
for (Node node : unitToEntryPointMap.values()) {
node.freeze();
}
singleUnitWords = null;
rcSet = null;
}
/**
* Gets the base node for this entry point
*
* @return the base node
*/
Node getNode() {
return baseNode;
}
/**
* Adds a one-unit word to this entry point. Such single unit words need to be dealt with specially.
*
* @param p the pronunciation of the single unit word
*/
void addSingleUnitWord(Pronunciation p) {
singleUnitWords.add(p);
}
/**
* Gets the set of possible right contexts that we can transition to from this entry point
*
* @return the set of possible transition points.
*/
private Collection<Unit> getEntryPointRC() {
if (rcSet == null) {
rcSet = new HashSet<Unit>();
for (Node node : baseNode.getSuccessorMap().values()) {
UnitNode unitNode = (UnitNode) node;
rcSet.add(unitNode.getBaseUnit());
}
}
return rcSet;
}
/**
* A version of createEntryPointMap that compresses common hmms across all entry points.
*/
void createEntryPointMap() {
HashMap<HMM, Node> map = new HashMap<HMM, Node>();
HashMap<HMM, HMMNode> singleUnitMap = new HashMap<HMM, HMMNode>();
for (Unit lc : exitPoints) {
Node epNode = new Node(LogMath.LOG_ZERO);
for (Unit rc : getEntryPointRC()) {
HMM hmm = hmmPool.getHMM(baseUnit, lc, rc, HMMPosition.BEGIN);
Node addedNode;
if ((addedNode = map.get(hmm)) == null) {
addedNode = epNode.addSuccessor(hmm, getProbability());
map.put(hmm, addedNode);
} else {
epNode.putSuccessor(hmm, addedNode);
}
nodeCount++;
connectEntryPointNode(addedNode, rc);
}
connectSingleUnitWords(lc, epNode, singleUnitMap);
unitToEntryPointMap.put(lc, epNode);
}
}
/**
* Connects the single unit words associated with this entry point. The singleUnitWords list contains all
* single unit pronunciations that have as their sole unit, the unit associated with this entry point. Entry
* points for these words are added to the epNode for all possible left (exit) and right (entry) contexts.
*
* @param lc the left context
* @param epNode the entry point node
*/
private void connectSingleUnitWords(Unit lc, Node epNode, HashMap<HMM, HMMNode> map) {
if (!singleUnitWords.isEmpty()) {
for (Unit rc : entryPoints) {
HMM hmm = hmmPool.getHMM(baseUnit, lc, rc, HMMPosition.SINGLE);
HMMNode tailNode;
if (( tailNode = map.get(hmm)) == null) {
tailNode = (HMMNode)
epNode.addSuccessor(hmm, getProbability());
map.put(hmm, tailNode);
} else {
epNode.putSuccessor(hmm, tailNode);
}
WordNode wordNode;
tailNode.addRC(rc);
nodeCount++;
for (Pronunciation p : singleUnitWords) {
if (p.getWord() == dictionary.getSentenceStartWord()) {
initialNode = new InitialWordNode(p, tailNode);
} else {
float prob = getWordUnigramProbability(p.getWord());
wordNode = tailNode.addSuccessor(p, prob);
if (p.getWord() ==
dictionary.getSentenceEndWord()) {
sentenceEndWordNode = wordNode;
}
}
nodeCount++;
}
}
}
}
/**
* Connect the entry points that match the given rc to the given epNode
*
* @param epNode add matching successors here
* @param rc the next unit
*/
private void connectEntryPointNode(Node epNode, Unit rc) {
for (Node node : baseNode.getSuccessors()) {
UnitNode successor = (UnitNode) node;
if (successor.getBaseUnit() == rc) {
epNode.addSuccessor(successor);
}
}
}
/** Dumps the entry point */
void dump() {
System.out.println("EntryPoint " + baseUnit + " RC Followers: "
+ getEntryPointRC().size());
int count = 0;
Collection<Unit> rcs = getEntryPointRC();
System.out.print(" ");
for (Unit rc : rcs) {
System.out.print(Utilities.pad(rc.getName(), 4));
if (count++ >= 12) {
count = 0;
System.out.println();
System.out.print(" ");
}
}
System.out.println();
}
}
}