// assumes that source-side syntax labels are stored in the target non-terminal field of the rules
void SourceGHKMTreeInputMatchFeature::EvaluateWithSourceContext(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, const StackVec *stackVec
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection *estimatedScores) const
{
const Range& range = inputPath.GetWordsRange();
size_t startPos = range.GetStartPos();
size_t endPos = range.GetEndPos();
const TreeInput& treeInput = static_cast<const TreeInput&>(input);
const NonTerminalSet& treeInputLabels = treeInput.GetLabelSet(startPos,endPos);
const Word& lhsLabel = targetPhrase.GetTargetLHS();
const StaticData& staticData = StaticData::Instance();
const Word& outputDefaultNonTerminal = staticData.GetOutputDefaultNonTerminal();
std::vector<float> newScores(m_numScoreComponents,0.0); // m_numScoreComponents == 2 // first fires for matches, second for mismatches
if ( (treeInputLabels.find(lhsLabel) != treeInputLabels.end()) && (lhsLabel != outputDefaultNonTerminal) ) {
// match
newScores[0] = 1.0;
} else {
// mismatch
newScores[1] = 1.0;
}
scoreBreakdown.PlusEquals(this, newScores);
}
void ChartRuleLookupManagerMemoryPerSentence::GetChartRuleCollection(
const InputPath &inputPath,
size_t lastPos,
ChartParserCallback &outColl)
{
const Range &range = inputPath.GetWordsRange();
size_t startPos = range.GetStartPos();
size_t absEndPos = range.GetEndPos();
m_lastPos = lastPos;
m_stackVec.clear();
m_stackScores.clear();
m_outColl = &outColl;
m_unaryPos = absEndPos-1; // rules ending in this position are unary and should not be added to collection
// create/update data structure to quickly look up all chart cells that match start position and label.
UpdateCompressedMatrix(startPos, absEndPos, lastPos);
const PhraseDictionaryNodeMemory &rootNode = m_ruleTable.GetRootNode(GetParser().GetTranslationId());
// all rules starting with terminal
if (startPos == absEndPos) {
GetTerminalExtension(&rootNode, startPos);
}
// all rules starting with nonterminal
else if (absEndPos > startPos) {
GetNonTerminalExtension(&rootNode, startPos);
}
// copy temporarily stored rules to out collection
CompletedRuleCollection & rules = m_completedRules[absEndPos];
for (vector<CompletedRule*>::const_iterator iter = rules.begin(); iter != rules.end(); ++iter) {
outColl.Add((*iter)->GetTPC(), (*iter)->GetStackVector(), range);
}
rules.Clear();
}
void WordTranslationFeature::EvaluateWithSourceContext(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, const StackVec *stackVec
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection *estimatedScores) const
{
const Sentence& sentence = static_cast<const Sentence&>(input);
const AlignmentInfo &alignment = targetPhrase.GetAlignTerm();
// process aligned words
for (AlignmentInfo::const_iterator alignmentPoint = alignment.begin(); alignmentPoint != alignment.end(); alignmentPoint++) {
const Phrase& sourcePhrase = inputPath.GetPhrase();
int sourceIndex = alignmentPoint->first;
int targetIndex = alignmentPoint->second;
Word ws = sourcePhrase.GetWord(sourceIndex);
if (m_factorTypeSource == 0 && ws.IsNonTerminal()) continue;
Word wt = targetPhrase.GetWord(targetIndex);
if (m_factorTypeSource == 0 && wt.IsNonTerminal()) continue;
StringPiece sourceWord = ws.GetFactor(m_factorTypeSource)->GetString();
StringPiece targetWord = wt.GetFactor(m_factorTypeTarget)->GetString();
if (m_ignorePunctuation) {
// check if source or target are punctuation
char firstChar = sourceWord[0];
CharHash::const_iterator charIterator = m_punctuationHash.find( firstChar );
if(charIterator != m_punctuationHash.end())
continue;
firstChar = targetWord[0];
charIterator = m_punctuationHash.find( firstChar );
if(charIterator != m_punctuationHash.end())
continue;
}
if (!m_unrestricted) {
if (FindStringPiece(m_vocabSource, sourceWord) == m_vocabSource.end())
sourceWord = "OTHER";
if (FindStringPiece(m_vocabTarget, targetWord) == m_vocabTarget.end())
targetWord = "OTHER";
}
if (m_simple) {
// construct feature name
util::StringStream featureName;
featureName << m_description << "_";
featureName << sourceWord;
featureName << "~";
featureName << targetWord;
scoreBreakdown.SparsePlusEquals(featureName.str(), 1);
}
if (m_domainTrigger && !m_sourceContext) {
const bool use_topicid = sentence.GetUseTopicId();
const bool use_topicid_prob = sentence.GetUseTopicIdAndProb();
if (use_topicid || use_topicid_prob) {
if(use_topicid) {
// use topicid as trigger
const long topicid = sentence.GetTopicId();
util::StringStream feature;
feature << m_description << "_";
if (topicid == -1)
feature << "unk";
else
feature << topicid;
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
scoreBreakdown.SparsePlusEquals(feature.str(), 1);
} else {
// use topic probabilities
const vector<string> &topicid_prob = *(input.GetTopicIdAndProb());
if (atol(topicid_prob[0].c_str()) == -1) {
util::StringStream feature;
feature << m_description << "_unk_";
feature << sourceWord;
feature << "~";
feature << targetWord;
scoreBreakdown.SparsePlusEquals(feature.str(), 1);
} else {
for (size_t i=0; i+1 < topicid_prob.size(); i+=2) {
util::StringStream feature;
feature << m_description << "_";
feature << topicid_prob[i];
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
scoreBreakdown.SparsePlusEquals(feature.str(), atof((topicid_prob[i+1]).c_str()));
}
}
}
} else {
// range over domain trigger words (keywords)
const long docid = input.GetDocumentId();
for (boost::unordered_set<std::string>::const_iterator p = m_vocabDomain[docid].begin(); p != m_vocabDomain[docid].end(); ++p) {
string sourceTrigger = *p;
util::StringStream feature;
feature << m_description << "_";
feature << sourceTrigger;
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
scoreBreakdown.SparsePlusEquals(feature.str(), 1);
}
}
}
if (m_sourceContext) {
size_t globalSourceIndex = inputPath.GetWordsRange().GetStartPos() + sourceIndex;
if (!m_domainTrigger && globalSourceIndex == 0) {
// add <s> trigger feature for source
util::StringStream feature;
feature << m_description << "_";
feature << "<s>,";
feature << sourceWord;
feature << "~";
feature << targetWord;
scoreBreakdown.SparsePlusEquals(feature.str(), 1);
}
// range over source words to get context
for(size_t contextIndex = 0; contextIndex < input.GetSize(); contextIndex++ ) {
if (contextIndex == globalSourceIndex) continue;
StringPiece sourceTrigger = input.GetWord(contextIndex).GetFactor(m_factorTypeSource)->GetString();
if (m_ignorePunctuation) {
// check if trigger is punctuation
char firstChar = sourceTrigger[0];
CharHash::const_iterator charIterator = m_punctuationHash.find( firstChar );
if(charIterator != m_punctuationHash.end())
continue;
}
const long docid = input.GetDocumentId();
bool sourceTriggerExists = false;
if (m_domainTrigger)
sourceTriggerExists = FindStringPiece(m_vocabDomain[docid], sourceTrigger ) != m_vocabDomain[docid].end();
else if (!m_unrestricted)
sourceTriggerExists = FindStringPiece(m_vocabSource, sourceTrigger ) != m_vocabSource.end();
if (m_domainTrigger) {
if (sourceTriggerExists) {
util::StringStream feature;
feature << m_description << "_";
feature << sourceTrigger;
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
scoreBreakdown.SparsePlusEquals(feature.str(), 1);
}
} else if (m_unrestricted || sourceTriggerExists) {
util::StringStream feature;
feature << m_description << "_";
if (contextIndex < globalSourceIndex) {
feature << sourceTrigger;
feature << ",";
feature << sourceWord;
} else {
feature << sourceWord;
feature << ",";
feature << sourceTrigger;
}
feature << "~";
feature << targetWord;
scoreBreakdown.SparsePlusEquals(feature.str(), 1);
}
}
}
if (m_targetContext) {
throw runtime_error("Can't use target words outside current translation option in a stateless feature");
/*
size_t globalTargetIndex = cur_hypo.GetCurrTargetWordsRange().GetStartPos() + targetIndex;
if (globalTargetIndex == 0) {
// add <s> trigger feature for source
stringstream feature;
feature << "wt_";
feature << sourceWord;
feature << "~";
feature << "<s>,";
feature << targetWord;
accumulator->SparsePlusEquals(feature.str(), 1);
}
// range over target words (up to current position) to get context
for(size_t contextIndex = 0; contextIndex < globalTargetIndex; contextIndex++ ) {
string targetTrigger = cur_hypo.GetWord(contextIndex).GetFactor(m_factorTypeTarget)->GetString();
if (m_ignorePunctuation) {
// check if trigger is punctuation
char firstChar = targetTrigger.at(0);
CharHash::const_iterator charIterator = m_punctuationHash.find( firstChar );
if(charIterator != m_punctuationHash.end())
continue;
}
bool targetTriggerExists = false;
if (!m_unrestricted)
targetTriggerExists = m_vocabTarget.find( targetTrigger ) != m_vocabTarget.end();
if (m_unrestricted || targetTriggerExists) {
stringstream feature;
feature << "wt_";
feature << sourceWord;
feature << "~";
feature << targetTrigger;
feature << ",";
feature << targetWord;
accumulator->SparsePlusEquals(feature.str(), 1);
}
}*/
}
}
}
void ChartRuleLookupManagerOnDisk::GetChartRuleCollection(
const InputPath &inputPath,
size_t lastPos,
ChartParserCallback &outColl)
{
const StaticData &staticData = StaticData::Instance();
const Word &defaultSourceNonTerm = staticData.GetInputDefaultNonTerminal();
const WordsRange &range = inputPath.GetWordsRange();
size_t relEndPos = range.GetEndPos() - range.GetStartPos();
size_t absEndPos = range.GetEndPos();
// MAIN LOOP. create list of nodes of target phrases
DottedRuleStackOnDisk &expandableDottedRuleList = *m_expandableDottedRuleListVec[range.GetStartPos()];
// sort save nodes so only do nodes with most counts
expandableDottedRuleList.SortSavedNodes();
const DottedRuleStackOnDisk::SavedNodeColl &savedNodeColl = expandableDottedRuleList.GetSavedNodeColl();
//cerr << "savedNodeColl=" << savedNodeColl.size() << " ";
const ChartCellLabel &sourceWordLabel = GetSourceAt(absEndPos);
for (size_t ind = 0; ind < (savedNodeColl.size()) ; ++ind) {
const SavedNodeOnDisk &savedNode = *savedNodeColl[ind];
const DottedRuleOnDisk &prevDottedRule = savedNode.GetDottedRule();
const OnDiskPt::PhraseNode &prevNode = prevDottedRule.GetLastNode();
size_t startPos = prevDottedRule.IsRoot() ? range.GetStartPos() : prevDottedRule.GetWordsRange().GetEndPos() + 1;
// search for terminal symbol
if (startPos == absEndPos) {
OnDiskPt::Word *sourceWordBerkeleyDb = m_dbWrapper.ConvertFromMoses(m_inputFactorsVec, sourceWordLabel.GetLabel());
if (sourceWordBerkeleyDb != NULL) {
const OnDiskPt::PhraseNode *node = prevNode.GetChild(*sourceWordBerkeleyDb, m_dbWrapper);
if (node != NULL) {
// TODO figure out why source word is needed from node, not from sentence
// prob to do with factors or non-term
//const Word &sourceWord = node->GetSourceWord();
DottedRuleOnDisk *dottedRule = new DottedRuleOnDisk(*node, sourceWordLabel, prevDottedRule);
expandableDottedRuleList.Add(relEndPos+1, dottedRule);
// cache for cleanup
m_sourcePhraseNode.push_back(node);
}
delete sourceWordBerkeleyDb;
}
}
// search for non-terminals
size_t endPos, stackInd;
if (startPos > absEndPos)
continue;
else if (startPos == range.GetStartPos() && range.GetEndPos() > range.GetStartPos()) {
// start.
endPos = absEndPos - 1;
stackInd = relEndPos;
} else {
endPos = absEndPos;
stackInd = relEndPos + 1;
}
// get target nonterminals in this span from chart
const ChartCellLabelSet &chartNonTermSet =
GetTargetLabelSet(startPos, endPos);
//const Word &defaultSourceNonTerm = staticData.GetInputDefaultNonTerminal()
// ,&defaultTargetNonTerm = staticData.GetOutputDefaultNonTerminal();
// go through each SOURCE lhs
const NonTerminalSet &sourceLHSSet = GetParser().GetInputPath(startPos, endPos).GetNonTerminalSet();
NonTerminalSet::const_iterator iterSourceLHS;
for (iterSourceLHS = sourceLHSSet.begin(); iterSourceLHS != sourceLHSSet.end(); ++iterSourceLHS) {
const Word &sourceLHS = *iterSourceLHS;
OnDiskPt::Word *sourceLHSBerkeleyDb = m_dbWrapper.ConvertFromMoses(m_inputFactorsVec, sourceLHS);
if (sourceLHSBerkeleyDb == NULL) {
delete sourceLHSBerkeleyDb;
continue; // vocab not in pt. node definately won't be in there
}
const OnDiskPt::PhraseNode *sourceNode = prevNode.GetChild(*sourceLHSBerkeleyDb, m_dbWrapper);
delete sourceLHSBerkeleyDb;
if (sourceNode == NULL)
continue; // didn't find source node
// go through each TARGET lhs
ChartCellLabelSet::const_iterator iterChartNonTerm;
for (iterChartNonTerm = chartNonTermSet.begin(); iterChartNonTerm != chartNonTermSet.end(); ++iterChartNonTerm) {
if (*iterChartNonTerm == NULL) {
continue;
}
const ChartCellLabel &cellLabel = **iterChartNonTerm;
bool doSearch = true;
if (m_dictionary.m_maxSpanDefault != NOT_FOUND) {
// for Hieu's source syntax
bool isSourceSyntaxNonTerm = sourceLHS != defaultSourceNonTerm;
size_t nonTermNumWordsCovered = endPos - startPos + 1;
doSearch = isSourceSyntaxNonTerm ?
nonTermNumWordsCovered <= m_dictionary.m_maxSpanLabelled :
nonTermNumWordsCovered <= m_dictionary.m_maxSpanDefault;
}
if (doSearch) {
OnDiskPt::Word *chartNonTermBerkeleyDb = m_dbWrapper.ConvertFromMoses(m_outputFactorsVec, cellLabel.GetLabel());
if (chartNonTermBerkeleyDb == NULL)
continue;
const OnDiskPt::PhraseNode *node = sourceNode->GetChild(*chartNonTermBerkeleyDb, m_dbWrapper);
delete chartNonTermBerkeleyDb;
if (node == NULL)
continue;
// found matching entry
//const Word &sourceWord = node->GetSourceWord();
DottedRuleOnDisk *dottedRule = new DottedRuleOnDisk(*node, cellLabel, prevDottedRule);
expandableDottedRuleList.Add(stackInd, dottedRule);
m_sourcePhraseNode.push_back(node);
}
} // for (iterChartNonTerm
delete sourceNode;
} // for (iterLabelListf
// return list of target phrases
DottedRuleCollOnDisk &nodes = expandableDottedRuleList.Get(relEndPos + 1);
// source LHS
DottedRuleCollOnDisk::const_iterator iterDottedRuleColl;
for (iterDottedRuleColl = nodes.begin(); iterDottedRuleColl != nodes.end(); ++iterDottedRuleColl) {
// node of last source word
const DottedRuleOnDisk &prevDottedRule = **iterDottedRuleColl;
if (prevDottedRule.Done())
continue;
prevDottedRule.Done(true);
const OnDiskPt::PhraseNode &prevNode = prevDottedRule.GetLastNode();
//get node for each source LHS
const NonTerminalSet &lhsSet = GetParser().GetInputPath(range.GetStartPos(), range.GetEndPos()).GetNonTerminalSet();
NonTerminalSet::const_iterator iterLabelSet;
for (iterLabelSet = lhsSet.begin(); iterLabelSet != lhsSet.end(); ++iterLabelSet) {
const Word &sourceLHS = *iterLabelSet;
OnDiskPt::Word *sourceLHSBerkeleyDb = m_dbWrapper.ConvertFromMoses(m_inputFactorsVec, sourceLHS);
if (sourceLHSBerkeleyDb == NULL)
continue;
const TargetPhraseCollection *targetPhraseCollection = NULL;
const OnDiskPt::PhraseNode *node = prevNode.GetChild(*sourceLHSBerkeleyDb, m_dbWrapper);
if (node) {
uint64_t tpCollFilePos = node->GetValue();
std::map<uint64_t, const TargetPhraseCollection*>::const_iterator iterCache = m_cache.find(tpCollFilePos);
if (iterCache == m_cache.end()) {
const OnDiskPt::TargetPhraseCollection *tpcollBerkeleyDb = node->GetTargetPhraseCollection(m_dictionary.GetTableLimit(), m_dbWrapper);
std::vector<float> weightT = staticData.GetWeights(&m_dictionary);
targetPhraseCollection
= tpcollBerkeleyDb->ConvertToMoses(m_inputFactorsVec
,m_outputFactorsVec
,m_dictionary
,weightT
,m_dbWrapper.GetVocab()
,true);
delete tpcollBerkeleyDb;
m_cache[tpCollFilePos] = targetPhraseCollection;
} else {
// just get out of cache
targetPhraseCollection = iterCache->second;
}
UTIL_THROW_IF2(targetPhraseCollection == NULL, "Error");
if (!targetPhraseCollection->IsEmpty()) {
AddCompletedRule(prevDottedRule, *targetPhraseCollection,
range, outColl);
}
} // if (node)
delete node;
delete sourceLHSBerkeleyDb;
}
}
} // for (size_t ind = 0; ind < savedNodeColl.size(); ++ind)
//cerr << numDerivations << " ";
}