FFState* TargetBigramFeature::Evaluate(const Hypothesis& cur_hypo,
const FFState* prev_state,
ScoreComponentCollection* accumulator) const
{
const TargetBigramState* tbState = dynamic_cast<const TargetBigramState*>(prev_state);
assert(tbState);
// current hypothesis target phrase
const Phrase& targetPhrase = cur_hypo.GetCurrTargetPhrase();
if (targetPhrase.GetSize() == 0) {
return new TargetBigramState(*tbState);
}
// extract all bigrams w1 w2 from current hypothesis
for (size_t i = 0; i < targetPhrase.GetSize(); ++i) {
const Factor* f1 = NULL;
if (i == 0) {
f1 = tbState->GetWord().GetFactor(m_factorType);
} else {
f1 = targetPhrase.GetWord(i-1).GetFactor(m_factorType);
}
const Factor* f2 = targetPhrase.GetWord(i).GetFactor(m_factorType);
const StringPiece w1 = f1->GetString();
const StringPiece w2 = f2->GetString();
// skip bigrams if they don't belong to a given restricted vocabulary
if (m_vocab.size() &&
(FindStringPiece(m_vocab, w1) == m_vocab.end() || FindStringPiece(m_vocab, w2) == m_vocab.end())) {
continue;
}
string name(w1.data(), w1.size());
name += ":";
name.append(w2.data(), w2.size());
accumulator->PlusEquals(this,name,1);
}
if (cur_hypo.GetWordsBitmap().IsComplete()) {
const StringPiece w1 = targetPhrase.GetWord(targetPhrase.GetSize()-1).GetFactor(m_factorType)->GetString();
const string& w2 = EOS_;
if (m_vocab.empty() || (FindStringPiece(m_vocab, w1) != m_vocab.end())) {
string name(w1.data(), w1.size());
name += ":";
name += w2;
accumulator->PlusEquals(this,name,1);
}
return NULL;
}
return new TargetBigramState(targetPhrase.GetWord(targetPhrase.GetSize()-1));
}
void GlobalLexicalModelUnlimited::Evaluate(const Hypothesis& cur_hypo, ScoreComponentCollection* accumulator) const
{
const Sentence& input = *(m_local->input);
const TargetPhrase& targetPhrase = cur_hypo.GetCurrTargetPhrase();
for(size_t targetIndex = 0; targetIndex < targetPhrase.GetSize(); targetIndex++ ) {
StringPiece targetString = targetPhrase.GetWord(targetIndex).GetString(0); // TODO: change for other factors
if (m_ignorePunctuation) {
// check if first char is punctuation
char firstChar = targetString[0];
CharHash::const_iterator charIterator = m_punctuationHash.find( firstChar );
if(charIterator != m_punctuationHash.end())
continue;
}
if (m_biasFeature) {
stringstream feature;
feature << "glm_";
feature << targetString;
feature << "~";
feature << "**BIAS**";
accumulator->SparsePlusEquals(feature.str(), 1);
}
boost::unordered_set<uint64_t> alreadyScored;
for(size_t sourceIndex = 0; sourceIndex < input.GetSize(); sourceIndex++ ) {
const StringPiece sourceString = input.GetWord(sourceIndex).GetString(0);
// TODO: change for other factors
if (m_ignorePunctuation) {
// check if first char is punctuation
char firstChar = sourceString[0];
CharHash::const_iterator charIterator = m_punctuationHash.find( firstChar );
if(charIterator != m_punctuationHash.end())
continue;
}
const uint64_t sourceHash = util::MurmurHashNative(sourceString.data(), sourceString.size());
if ( alreadyScored.find(sourceHash) == alreadyScored.end()) {
bool sourceExists, targetExists;
if (!m_unrestricted) {
sourceExists = FindStringPiece(m_vocabSource, sourceString ) != m_vocabSource.end();
targetExists = FindStringPiece(m_vocabTarget, targetString) != m_vocabTarget.end();
}
// no feature if vocab is in use and both words are not in restricted vocabularies
if (m_unrestricted || (sourceExists && targetExists)) {
if (m_sourceContext) {
if (sourceIndex == 0) {
// add <s> trigger feature for source
stringstream feature;
feature << "glm_";
feature << targetString;
feature << "~";
feature << "<s>,";
feature << sourceString;
accumulator->SparsePlusEquals(feature.str(), 1);
alreadyScored.insert(sourceHash);
}
// add source words to the right of current source word as context
for(int contextIndex = sourceIndex+1; contextIndex < input.GetSize(); contextIndex++ ) {
StringPiece contextString = input.GetWord(contextIndex).GetString(0); // TODO: change for other factors
bool contextExists;
if (!m_unrestricted)
contextExists = FindStringPiece(m_vocabSource, contextString ) != m_vocabSource.end();
if (m_unrestricted || contextExists) {
stringstream feature;
feature << "glm_";
feature << targetString;
feature << "~";
feature << sourceString;
feature << ",";
feature << contextString;
accumulator->SparsePlusEquals(feature.str(), 1);
alreadyScored.insert(sourceHash);
}
}
} else if (m_biphrase) {
// --> look backwards for constructing context
int globalTargetIndex = cur_hypo.GetSize() - targetPhrase.GetSize() + targetIndex;
// 1) source-target pair, trigger source word (can be discont.) and adjacent target word (bigram)
StringPiece targetContext;
if (globalTargetIndex > 0)
targetContext = cur_hypo.GetWord(globalTargetIndex-1).GetString(0); // TODO: change for other factors
else
targetContext = "<s>";
if (sourceIndex == 0) {
StringPiece sourceTrigger = "<s>";
AddFeature(accumulator, sourceTrigger, sourceString,
targetContext, targetString);
} else
for(int contextIndex = sourceIndex-1; contextIndex >= 0; contextIndex-- ) {
StringPiece sourceTrigger = input.GetWord(contextIndex).GetString(0); // TODO: change for other factors
bool sourceTriggerExists = false;
if (!m_unrestricted)
sourceTriggerExists = FindStringPiece(m_vocabSource, sourceTrigger ) != m_vocabSource.end();
if (m_unrestricted || sourceTriggerExists)
AddFeature(accumulator, sourceTrigger, sourceString,
targetContext, targetString);
}
// 2) source-target pair, adjacent source word (bigram) and trigger target word (can be discont.)
StringPiece sourceContext;
if (sourceIndex-1 >= 0)
sourceContext = input.GetWord(sourceIndex-1).GetString(0); // TODO: change for other factors
else
sourceContext = "<s>";
if (globalTargetIndex == 0) {
string targetTrigger = "<s>";
AddFeature(accumulator, sourceContext, sourceString,
targetTrigger, targetString);
} else
for(int globalContextIndex = globalTargetIndex-1; globalContextIndex >= 0; globalContextIndex-- ) {
StringPiece targetTrigger = cur_hypo.GetWord(globalContextIndex).GetString(0); // TODO: change for other factors
bool targetTriggerExists = false;
if (!m_unrestricted)
targetTriggerExists = FindStringPiece(m_vocabTarget, targetTrigger ) != m_vocabTarget.end();
if (m_unrestricted || targetTriggerExists)
AddFeature(accumulator, sourceContext, sourceString,
targetTrigger, targetString);
}
} else if (m_bitrigger) {
// allow additional discont. triggers on both sides
int globalTargetIndex = cur_hypo.GetSize() - targetPhrase.GetSize() + targetIndex;
if (sourceIndex == 0) {
StringPiece sourceTrigger = "<s>";
bool sourceTriggerExists = true;
if (globalTargetIndex == 0) {
string targetTrigger = "<s>";
bool targetTriggerExists = true;
if (m_unrestricted || (sourceTriggerExists && targetTriggerExists))
AddFeature(accumulator, sourceTrigger, sourceString,
targetTrigger, targetString);
} else {
// iterate backwards over target
for(int globalContextIndex = globalTargetIndex-1; globalContextIndex >= 0; globalContextIndex-- ) {
StringPiece targetTrigger = cur_hypo.GetWord(globalContextIndex).GetString(0); // TODO: change for other factors
bool targetTriggerExists = false;
if (!m_unrestricted)
targetTriggerExists = FindStringPiece(m_vocabTarget, targetTrigger ) != m_vocabTarget.end();
if (m_unrestricted || (sourceTriggerExists && targetTriggerExists))
AddFeature(accumulator, sourceTrigger, sourceString,
targetTrigger, targetString);
}
}
}
// iterate over both source and target
else {
// iterate backwards over source
for(int contextIndex = sourceIndex-1; contextIndex >= 0; contextIndex-- ) {
StringPiece sourceTrigger = input.GetWord(contextIndex).GetString(0); // TODO: change for other factors
bool sourceTriggerExists = false;
if (!m_unrestricted)
sourceTriggerExists = FindStringPiece(m_vocabSource, sourceTrigger ) != m_vocabSource.end();
if (globalTargetIndex == 0) {
string targetTrigger = "<s>";
bool targetTriggerExists = true;
if (m_unrestricted || (sourceTriggerExists && targetTriggerExists))
AddFeature(accumulator, sourceTrigger, sourceString,
targetTrigger, targetString);
} else {
// iterate backwards over target
for(int globalContextIndex = globalTargetIndex-1; globalContextIndex >= 0; globalContextIndex-- ) {
StringPiece targetTrigger = cur_hypo.GetWord(globalContextIndex).GetString(0); // TODO: change for other factors
bool targetTriggerExists = false;
if (!m_unrestricted)
targetTriggerExists = FindStringPiece(m_vocabTarget, targetTrigger ) != m_vocabTarget.end();
if (m_unrestricted || (sourceTriggerExists && targetTriggerExists))
AddFeature(accumulator, sourceTrigger, sourceString,
targetTrigger, targetString);
}
}
}
}
} else {
stringstream feature;
feature << "glm_";
feature << targetString;
feature << "~";
feature << sourceString;
accumulator->SparsePlusEquals(feature.str(), 1);
alreadyScored.insert(sourceHash);
}
}
}
}
}
}
void WordTranslationFeature::EvaluateWhenApplied
(const Hypothesis& hypo,
ScoreComponentCollection* accumulator) const
{
const Sentence& input = static_cast<const Sentence&>(hypo.GetInput());
const TranslationOption& transOpt = hypo.GetTranslationOption();
const TargetPhrase& targetPhrase = hypo.GetCurrTargetPhrase();
const AlignmentInfo &alignment = targetPhrase.GetAlignTerm();
// process aligned words
for (AlignmentInfo::const_iterator alignmentPoint = alignment.begin(); alignmentPoint != alignment.end(); alignmentPoint++) {
const Phrase& sourcePhrase = transOpt.GetInputPath().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
stringstream featureName;
featureName << m_description << "_";
featureName << sourceWord;
featureName << "~";
featureName << targetWord;
accumulator->SparsePlusEquals(featureName.str(), 1);
}
if (m_domainTrigger && !m_sourceContext) {
const bool use_topicid = input.GetUseTopicId();
const bool use_topicid_prob = input.GetUseTopicIdAndProb();
if (use_topicid || use_topicid_prob) {
if(use_topicid) {
// use topicid as trigger
const long topicid = input.GetTopicId();
stringstream feature;
feature << m_description << "_";
if (topicid == -1)
feature << "unk";
else
feature << topicid;
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
accumulator->SparsePlusEquals(feature.str(), 1);
} else {
// use topic probabilities
const vector<string> &topicid_prob = *(input.GetTopicIdAndProb());
if (atol(topicid_prob[0].c_str()) == -1) {
stringstream feature;
feature << m_description << "_unk_";
feature << sourceWord;
feature << "~";
feature << targetWord;
accumulator->SparsePlusEquals(feature.str(), 1);
} else {
for (size_t i=0; i+1 < topicid_prob.size(); i+=2) {
stringstream feature;
feature << m_description << "_";
feature << topicid_prob[i];
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
accumulator->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;
stringstream feature;
feature << m_description << "_";
feature << sourceTrigger;
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
accumulator->SparsePlusEquals(feature.str(), 1);
}
}
}
if (m_sourceContext) {
size_t globalSourceIndex = hypo.GetTranslationOption().GetStartPos() + sourceIndex;
if (!m_domainTrigger && globalSourceIndex == 0) {
// add <s> trigger feature for source
stringstream feature;
feature << m_description << "_";
feature << "<s>,";
feature << sourceWord;
feature << "~";
feature << targetWord;
accumulator->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) {
stringstream feature;
feature << m_description << "_";
feature << sourceTrigger;
feature << "_";
feature << sourceWord;
feature << "~";
feature << targetWord;
accumulator->SparsePlusEquals(feature.str(), 1);
}
} else if (m_unrestricted || sourceTriggerExists) {
stringstream feature;
feature << m_description << "_";
if (contextIndex < globalSourceIndex) {
feature << sourceTrigger;
feature << ",";
feature << sourceWord;
} else {
feature << sourceWord;
feature << ",";
feature << sourceTrigger;
}
feature << "~";
feature << targetWord;
accumulator->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 PhrasePairFeature::EvaluateWhenApplied(
const Hypothesis& hypo,
ScoreComponentCollection* accumulator) const
{
const TargetPhrase& target = hypo.GetCurrTargetPhrase();
const Phrase& source = hypo.GetTranslationOption().GetInputPath().GetPhrase();
if (m_simple) {
ostringstream namestr;
namestr << "pp_";
namestr << source.GetWord(0).GetFactor(m_sourceFactorId)->GetString();
for (size_t i = 1; i < source.GetSize(); ++i) {
const Factor* sourceFactor = source.GetWord(i).GetFactor(m_sourceFactorId);
namestr << ",";
namestr << sourceFactor->GetString();
}
namestr << "~";
namestr << target.GetWord(0).GetFactor(m_targetFactorId)->GetString();
for (size_t i = 1; i < target.GetSize(); ++i) {
const Factor* targetFactor = target.GetWord(i).GetFactor(m_targetFactorId);
namestr << ",";
namestr << targetFactor->GetString();
}
accumulator->SparsePlusEquals(namestr.str(),1);
}
if (m_domainTrigger) {
const Sentence& input = static_cast<const Sentence&>(hypo.GetInput());
const bool use_topicid = input.GetUseTopicId();
const bool use_topicid_prob = input.GetUseTopicIdAndProb();
// compute pair
ostringstream pair;
pair << source.GetWord(0).GetFactor(m_sourceFactorId)->GetString();
for (size_t i = 1; i < source.GetSize(); ++i) {
const Factor* sourceFactor = source.GetWord(i).GetFactor(m_sourceFactorId);
pair << ",";
pair << sourceFactor->GetString();
}
pair << "~";
pair << target.GetWord(0).GetFactor(m_targetFactorId)->GetString();
for (size_t i = 1; i < target.GetSize(); ++i) {
const Factor* targetFactor = target.GetWord(i).GetFactor(m_targetFactorId);
pair << ",";
pair << targetFactor->GetString();
}
if (use_topicid || use_topicid_prob) {
if(use_topicid) {
// use topicid as trigger
const long topicid = input.GetTopicId();
stringstream feature;
feature << "pp_";
if (topicid == -1)
feature << "unk";
else
feature << topicid;
feature << "_";
feature << pair.str();
accumulator->SparsePlusEquals(feature.str(), 1);
} else {
// use topic probabilities
const vector<string> &topicid_prob = *(input.GetTopicIdAndProb());
if (atol(topicid_prob[0].c_str()) == -1) {
stringstream feature;
feature << "pp_unk_";
feature << pair.str();
accumulator->SparsePlusEquals(feature.str(), 1);
} else {
for (size_t i=0; i+1 < topicid_prob.size(); i+=2) {
stringstream feature;
feature << "pp_";
feature << topicid_prob[i];
feature << "_";
feature << pair.str();
accumulator->SparsePlusEquals(feature.str(), atof((topicid_prob[i+1]).c_str()));
}
}
}
} else {
// range over domain trigger words
const long docid = input.GetDocumentId();
for (set<string>::const_iterator p = m_vocabDomain[docid].begin(); p != m_vocabDomain[docid].end(); ++p) {
string sourceTrigger = *p;
ostringstream namestr;
namestr << "pp_";
namestr << sourceTrigger;
namestr << "_";
namestr << pair.str();
accumulator->SparsePlusEquals(namestr.str(),1);
}
}
}
if (m_sourceContext) {
const Sentence& input = static_cast<const Sentence&>(hypo.GetInput());
// range over source words to get context
for(size_t contextIndex = 0; contextIndex < input.GetSize(); contextIndex++ ) {
StringPiece sourceTrigger = input.GetWord(contextIndex).GetFactor(m_sourceFactorId)->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;
}
bool sourceTriggerExists = false;
if (!m_unrestricted)
sourceTriggerExists = FindStringPiece(m_vocabSource, sourceTrigger ) != m_vocabSource.end();
if (m_unrestricted || sourceTriggerExists) {
ostringstream namestr;
namestr << "pp_";
namestr << sourceTrigger;
namestr << "~";
namestr << source.GetWord(0).GetFactor(m_sourceFactorId)->GetString();
for (size_t i = 1; i < source.GetSize(); ++i) {
const Factor* sourceFactor = source.GetWord(i).GetFactor(m_sourceFactorId);
namestr << ",";
namestr << sourceFactor->GetString();
}
namestr << "~";
namestr << target.GetWord(0).GetFactor(m_targetFactorId)->GetString();
for (size_t i = 1; i < target.GetSize(); ++i) {
const Factor* targetFactor = target.GetWord(i).GetFactor(m_targetFactorId);
namestr << ",";
namestr << targetFactor->GetString();
}
accumulator->SparsePlusEquals(namestr.str(),1);
}
}
}
}
