void OutputBestSurface(std::ostream &out, const Hypothesis *hypo, const std::vector<FactorType> &outputFactorOrder,
char reportSegmentation, bool reportAllFactors)
{
if (hypo != NULL) {
// recursively retrace this best path through the lattice, starting from the end of the hypothesis sentence
OutputBestSurface(out, hypo->GetPrevHypo(), outputFactorOrder, reportSegmentation, reportAllFactors);
OutputSurface(out, *hypo, outputFactorOrder, reportSegmentation, reportAllFactors);
}
}
void OutputBestHypo(const Moses::TrellisPath &path, long /*translationId*/, char reportSegmentation, bool reportAllFactors, std::ostream &out)
{
const std::vector<const Hypothesis *> &edges = path.GetEdges();
for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) {
const Hypothesis &edge = *edges[currEdge];
OutputSurface(out, edge, StaticData::Instance().GetOutputFactorOrder(), reportSegmentation, reportAllFactors);
}
out << endl;
}
void OutputSurface(std::ostream &out, const ChartHypothesis *hypo, const std::vector<FactorType> &outputFactorOrder
,bool reportSegmentation, bool reportAllFactors)
{
if ( hypo != NULL) {
//OutputSurface(out, hypo->GetCurrTargetPhrase(), outputFactorOrder, reportAllFactors);
const vector<const ChartHypothesis*> &prevHypos = hypo->GetPrevHypos();
vector<const ChartHypothesis*>::const_iterator iter;
for (iter = prevHypos.begin(); iter != prevHypos.end(); ++iter) {
const ChartHypothesis *prevHypo = *iter;
OutputSurface(out, prevHypo, outputFactorOrder, reportSegmentation, reportAllFactors);
}
}
}
void OutputNBest(std::ostream& out
, const Moses::TrellisPathList &nBestList
, const std::vector<Moses::FactorType>& outputFactorOrder
, long translationId
, char reportSegmentation)
{
const StaticData &staticData = StaticData::Instance();
bool reportAllFactors = staticData.GetReportAllFactorsNBest();
bool includeSegmentation = staticData.NBestIncludesSegmentation();
bool includeWordAlignment = staticData.PrintAlignmentInfoInNbest();
TrellisPathList::const_iterator iter;
for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) {
const TrellisPath &path = **iter;
const std::vector<const Hypothesis *> &edges = path.GetEdges();
// print the surface factor of the translation
out << translationId << " ||| ";
for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) {
const Hypothesis &edge = *edges[currEdge];
OutputSurface(out, edge, outputFactorOrder, reportSegmentation, reportAllFactors);
}
out << " |||";
// print scores with feature names
OutputAllFeatureScores(path.GetScoreBreakdown(), out );
// total
out << " ||| " << path.GetTotalScore();
//phrase-to-phrase segmentation
if (includeSegmentation) {
out << " |||";
for (int currEdge = (int)edges.size() - 2 ; currEdge >= 0 ; currEdge--) {
const Hypothesis &edge = *edges[currEdge];
const WordsRange &sourceRange = edge.GetCurrSourceWordsRange();
WordsRange targetRange = path.GetTargetWordsRange(edge);
out << " " << sourceRange.GetStartPos();
if (sourceRange.GetStartPos() < sourceRange.GetEndPos()) {
out << "-" << sourceRange.GetEndPos();
}
out<< "=" << targetRange.GetStartPos();
if (targetRange.GetStartPos() < targetRange.GetEndPos()) {
out<< "-" << targetRange.GetEndPos();
}
}
}
if (includeWordAlignment) {
out << " ||| ";
for (int currEdge = (int)edges.size() - 2 ; currEdge >= 0 ; currEdge--) {
const Hypothesis &edge = *edges[currEdge];
const WordsRange &sourceRange = edge.GetCurrSourceWordsRange();
WordsRange targetRange = path.GetTargetWordsRange(edge);
const int sourceOffset = sourceRange.GetStartPos();
const int targetOffset = targetRange.GetStartPos();
const AlignmentInfo &ai = edge.GetCurrTargetPhrase().GetAlignTerm();
OutputAlignment(out, ai, sourceOffset, targetOffset);
}
}
if (StaticData::Instance().IsPathRecoveryEnabled()) {
out << " ||| ";
OutputInput(out, edges[0]);
}
out << endl;
}
out << std::flush;
}
void IOWrapper::OutputNBestList(const ChartTrellisPathList &nBestList, const ChartHypothesis *bestHypo, const TranslationSystem* system, long translationId)
{
std::ostringstream out;
// Check if we're writing to std::cout.
if (m_nBestOutputCollector->OutputIsCout()) {
// Set precision only if we're writing the n-best list to cout. This is to
// preserve existing behaviour, but should probably be done either way.
IOWrapper::FixPrecision(out);
// The output from -output-hypo-score is always written to std::cout.
if (StaticData::Instance().GetOutputHypoScore()) {
if (bestHypo != NULL) {
out << bestHypo->GetTotalScore() << " ";
} else {
out << "0 ";
}
}
}
bool labeledOutput = StaticData::Instance().IsLabeledNBestList();
//bool includeAlignment = StaticData::Instance().NBestIncludesAlignment();
ChartTrellisPathList::const_iterator iter;
for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) {
const ChartTrellisPath &path = **iter;
//cerr << path << endl << endl;
Moses::Phrase outputPhrase = path.GetOutputPhrase();
// delete 1st & last
CHECK(outputPhrase.GetSize() >= 2);
outputPhrase.RemoveWord(0);
outputPhrase.RemoveWord(outputPhrase.GetSize() - 1);
// print the surface factor of the translation
out << translationId << " ||| ";
OutputSurface(out, outputPhrase, m_outputFactorOrder, false);
out << " ||| ";
// print the scores in a hardwired order
// before each model type, the corresponding command-line-like name must be emitted
// MERT script relies on this
// lm
const LMList& lml = system->GetLanguageModels();
if (lml.size() > 0) {
if (labeledOutput)
out << "lm:";
LMList::const_iterator lmi = lml.begin();
for (; lmi != lml.end(); ++lmi) {
out << " " << path.GetScoreBreakdown().GetScoreForProducer(*lmi);
}
}
std::string lastName = "";
// output stateful sparse features
const vector<const StatefulFeatureFunction*>& sff = system->GetStatefulFeatureFunctions();
for( size_t i=0; i<sff.size(); i++ )
if (sff[i]->GetNumScoreComponents() == ScoreProducer::unlimited)
OutputSparseFeatureScores( out, path, sff[i], lastName );
// translation components
const vector<PhraseDictionaryFeature*>& pds = system->GetPhraseDictionaries();
if (pds.size() > 0) {
for( size_t i=0; i<pds.size(); i++ ) {
size_t pd_numinputscore = pds[i]->GetNumInputScores();
vector<float> scores = path.GetScoreBreakdown().GetScoresForProducer( pds[i] );
for (size_t j = 0; j<scores.size(); ++j){
if (labeledOutput && (i == 0) ){
if ((j == 0) || (j == pd_numinputscore)){
lastName = pds[i]->GetScoreProducerWeightShortName(j);
out << " " << lastName << ":";
}
}
out << " " << scores[j];
}
}
}
// word penalty
if (labeledOutput)
out << " w:";
out << " " << path.GetScoreBreakdown().GetScoreForProducer(system->GetWordPenaltyProducer());
// generation
const vector<GenerationDictionary*>& gds = system->GetGenerationDictionaries();
if (gds.size() > 0) {
for( size_t i=0; i<gds.size(); i++ ) {
size_t pd_numinputscore = gds[i]->GetNumInputScores();
vector<float> scores = path.GetScoreBreakdown().GetScoresForProducer( gds[i] );
for (size_t j = 0; j<scores.size(); ++j){
if (labeledOutput && (i == 0) ){
if ((j == 0) || (j == pd_numinputscore)){
lastName = gds[i]->GetScoreProducerWeightShortName(j);
out << " " << lastName << ":";
}
}
out << " " << scores[j];
}
}
}
// output stateless sparse features
lastName = "";
const vector<const StatelessFeatureFunction*>& slf = system->GetStatelessFeatureFunctions();
for( size_t i=0; i<slf.size(); i++ ) {
if (slf[i]->GetNumScoreComponents() == ScoreProducer::unlimited) {
OutputSparseFeatureScores( out, path, slf[i], lastName );
}
}
// total
out << " ||| " << path.GetTotalScore();
/*
if (includeAlignment) {
*m_nBestStream << " |||";
for (int currEdge = (int)edges.size() - 2 ; currEdge >= 0 ; currEdge--)
{
const ChartHypothesis &edge = *edges[currEdge];
WordsRange sourceRange = edge.GetCurrSourceWordsRange();
WordsRange targetRange = edge.GetCurrTargetWordsRange();
*m_nBestStream << " " << sourceRange.GetStartPos();
if (sourceRange.GetStartPos() < sourceRange.GetEndPos()) {
*m_nBestStream << "-" << sourceRange.GetEndPos();
}
*m_nBestStream << "=" << targetRange.GetStartPos();
if (targetRange.GetStartPos() < targetRange.GetEndPos()) {
*m_nBestStream << "-" << targetRange.GetEndPos();
}
}
}
*/
out << endl;
}
out <<std::flush;
CHECK(m_nBestOutputCollector);
m_nBestOutputCollector->Write(translationId, out.str());
}
void OutputNBest(std::ostream& out, const Moses::TrellisPathList &nBestList, const std::vector<Moses::FactorType>& outputFactorOrder, const TranslationSystem* system, long translationId, bool reportSegmentation)
{
const StaticData &staticData = StaticData::Instance();
bool labeledOutput = staticData.IsLabeledNBestList();
bool reportAllFactors = staticData.GetReportAllFactorsNBest();
bool includeSegmentation = staticData.NBestIncludesSegmentation();
bool includeWordAlignment = staticData.PrintAlignmentInfoInNbest();
TrellisPathList::const_iterator iter;
for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) {
const TrellisPath &path = **iter;
const std::vector<const Hypothesis *> &edges = path.GetEdges();
// print the surface factor of the translation
out << translationId << " ||| ";
for (int currEdge = (int)edges.size() - 1 ; currEdge >= 0 ; currEdge--) {
const Hypothesis &edge = *edges[currEdge];
OutputSurface(out, edge, outputFactorOrder, reportSegmentation, reportAllFactors);
}
out << " |||";
// print scores with feature names
OutputAllFeatureScores( out, system, path );
string lastName;
// translation components
const vector<PhraseDictionaryFeature*>& pds = system->GetPhraseDictionaries();
if (pds.size() > 0) {
for( size_t i=0; i<pds.size(); i++ ) {
size_t pd_numinputscore = pds[i]->GetNumInputScores();
vector<float> scores = path.GetScoreBreakdown().GetScoresForProducer( pds[i] );
for (size_t j = 0; j<scores.size(); ++j){
if (labeledOutput && (i == 0) ){
if ((j == 0) || (j == pd_numinputscore)){
lastName = pds[i]->GetScoreProducerWeightShortName(j);
out << " " << lastName << ":";
}
}
out << " " << scores[j];
}
}
}
// generation
const vector<GenerationDictionary*>& gds = system->GetGenerationDictionaries();
if (gds.size() > 0) {
for( size_t i=0; i<gds.size(); i++ ) {
size_t pd_numinputscore = gds[i]->GetNumInputScores();
vector<float> scores = path.GetScoreBreakdown().GetScoresForProducer( gds[i] );
for (size_t j = 0; j<scores.size(); ++j){
if (labeledOutput && (i == 0) ){
if ((j == 0) || (j == pd_numinputscore)){
lastName = gds[i]->GetScoreProducerWeightShortName(j);
out << " " << lastName << ":";
}
}
out << " " << scores[j];
}
}
}
// total
out << " ||| " << path.GetTotalScore();
//phrase-to-phrase segmentation
if (includeSegmentation) {
out << " |||";
for (int currEdge = (int)edges.size() - 2 ; currEdge >= 0 ; currEdge--) {
const Hypothesis &edge = *edges[currEdge];
const WordsRange &sourceRange = edge.GetCurrSourceWordsRange();
WordsRange targetRange = path.GetTargetWordsRange(edge);
out << " " << sourceRange.GetStartPos();
if (sourceRange.GetStartPos() < sourceRange.GetEndPos()) {
out << "-" << sourceRange.GetEndPos();
}
out<< "=" << targetRange.GetStartPos();
if (targetRange.GetStartPos() < targetRange.GetEndPos()) {
out<< "-" << targetRange.GetEndPos();
}
}
}
if (includeWordAlignment) {
out << " ||| ";
for (int currEdge = (int)edges.size() - 2 ; currEdge >= 0 ; currEdge--) {
const Hypothesis &edge = *edges[currEdge];
const WordsRange &sourceRange = edge.GetCurrSourceWordsRange();
WordsRange targetRange = path.GetTargetWordsRange(edge);
const int sourceOffset = sourceRange.GetStartPos();
const int targetOffset = targetRange.GetStartPos();
const AlignmentInfo &ai = edge.GetCurrTargetPhrase().GetAlignTerm();
OutputAlignment(out, ai, sourceOffset, targetOffset);
}
}
if (StaticData::Instance().IsPathRecoveryEnabled()) {
out << "|||";
OutputInput(out, edges[0]);
}
out << endl;
}
out << std::flush;
}
void Manager::OutputNBestList(OutputCollector *collector,
const KBestExtractor::KBestVec &nBestList,
long translationId) const
{
const StaticData &staticData = StaticData::Instance();
const std::vector<FactorType> &outputFactorOrder =
staticData.GetOutputFactorOrder();
std::ostringstream out;
if (collector->OutputIsCout()) {
// Set precision only if we're writing the n-best list to cout. This is to
// preserve existing behaviour, but should probably be done either way.
FixPrecision(out);
}
bool includeWordAlignment = staticData.PrintAlignmentInfoInNbest();
bool PrintNBestTrees = staticData.PrintNBestTrees();
for (KBestExtractor::KBestVec::const_iterator p = nBestList.begin();
p != nBestList.end(); ++p) {
const KBestExtractor::Derivation &derivation = **p;
// get the derivation's target-side yield
Phrase outputPhrase = KBestExtractor::GetOutputPhrase(derivation);
// delete <s> and </s>
UTIL_THROW_IF2(outputPhrase.GetSize() < 2,
"Output phrase should have contained at least 2 words (beginning and end-of-sentence)");
outputPhrase.RemoveWord(0);
outputPhrase.RemoveWord(outputPhrase.GetSize() - 1);
// print the translation ID, surface factors, and scores
out << translationId << " ||| ";
OutputSurface(out, outputPhrase, outputFactorOrder, false);
out << " ||| ";
OutputAllFeatureScores(derivation.scoreBreakdown, out);
out << " ||| " << derivation.score;
// optionally, print word alignments
if (includeWordAlignment) {
out << " ||| ";
Alignments align;
OutputAlignmentNBest(align, derivation, 0);
for (Alignments::const_iterator q = align.begin(); q != align.end();
++q) {
out << q->first << "-" << q->second << " ";
}
}
// optionally, print tree
if (PrintNBestTrees) {
TreePointer tree = KBestExtractor::GetOutputTree(derivation);
out << " ||| " << tree->GetString();
}
out << std::endl;
}
assert(collector);
collector->Write(translationId, out.str());
}