vector<const Factor*> doMBR(const TrellisPathList& nBestList){
// cerr << "Sentence " << sent << " has " << sents.size() << " candidate translations" << endl;
float marginal = 0;
vector<float> joint_prob_vec;
vector< vector<const Factor*> > translations;
float joint_prob;
vector< map < vector <const Factor *>, int > > ngram_stats;
TrellisPathList::const_iterator iter;
//TrellisPath* hyp = NULL;
for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter)
{
const TrellisPath &path = **iter;
joint_prob = UntransformScore(StaticData::Instance().GetMBRScale() * path.GetScoreBreakdown().InnerProduct(StaticData::Instance().GetAllWeights()));
marginal += joint_prob;
joint_prob_vec.push_back(joint_prob);
//Cache ngram counts
map < vector < const Factor *>, int > counts;
vector<const Factor*> translation;
GetOutputFactors(path, translation);
//TO DO
extract_ngrams(translation,counts);
ngram_stats.push_back(counts);
translations.push_back(translation);
}
vector<float> mbr_loss;
float bleu, weightedLoss;
float weightedLossCumul = 0;
float minMBRLoss = 1000000;
int minMBRLossIdx = -1;
/* Main MBR computation done here */
for (int i = 0; i < nBestList.GetSize(); i++){
weightedLossCumul = 0;
for (int j = 0; j < nBestList.GetSize(); j++){
if ( i != j) {
bleu = calculate_score(translations, j, i,ngram_stats );
weightedLoss = ( 1 - bleu) * ( joint_prob_vec[j]/marginal);
weightedLossCumul += weightedLoss;
if (weightedLossCumul > minMBRLoss)
break;
}
}
if (weightedLossCumul < minMBRLoss){
minMBRLoss = weightedLossCumul;
minMBRLossIdx = i;
}
}
/* Find sentence that minimises Bayes Risk under 1- BLEU loss */
return translations[minMBRLossIdx];
}
vector< vector<const Word*> > MosesDecoder::runDecoder(const std::string& source,
size_t sentenceid,
size_t nBestSize,
float bleuObjectiveWeight,
float bleuScoreWeight,
vector< ScoreComponentCollection>& featureValues,
vector< float>& bleuScores,
vector< float>& modelScores,
size_t numReturnedTranslations,
bool realBleu,
bool distinct,
size_t rank,
size_t epoch,
SearchAlgorithm& search,
string filename)
{
// run the decoder
m_manager = new Moses::Manager(*m_sentence);
m_manager->Decode();
TrellisPathList nBestList;
m_manager->CalcNBest(nBestSize, nBestList, distinct);
// optionally print nbest to file (to extract scores and features.. currently just for sentence bleu scoring)
/*if (filename != "") {
ofstream out(filename.c_str());
if (!out) {
ostringstream msg;
msg << "Unable to open " << filename;
throw runtime_error(msg.str());
}
// TODO: handle sentence id (for now always 0)
//OutputNBest(out, nBestList, StaticData::Instance().GetOutputFactorOrder(), 0, false);
out.close();
}*/
// read off the feature values and bleu scores for each sentence in the nbest list
Moses::TrellisPathList::const_iterator iter;
for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter) {
const Moses::TrellisPath &path = **iter;
featureValues.push_back(path.GetScoreBreakdown());
float bleuScore, dynBleuScore, realBleuScore;
if (realBleu) realBleuScore = m_bleuScoreFeature->CalculateBleu(path.GetTargetPhrase());
else dynBleuScore = getBleuScore(featureValues.back());
bleuScore = realBleu ? realBleuScore : dynBleuScore;
bleuScores.push_back(bleuScore);
//std::cout << "Score breakdown: " << path.GetScoreBreakdown() << endl;
float scoreWithoutBleu = path.GetTotalScore() - (bleuObjectiveWeight * bleuScoreWeight * bleuScore);
modelScores.push_back(scoreWithoutBleu);
if (iter != nBestList.begin())
cerr << endl;
cerr << "Rank " << rank << ", epoch " << epoch << ", \"" << path.GetTargetPhrase() << "\", score: "
<< scoreWithoutBleu << ", Bleu: " << bleuScore << ", total: " << path.GetTotalScore();
if (m_bleuScoreFeature->Enabled() && realBleu)
cerr << " (d-bleu: " << dynBleuScore << ", r-bleu: " << realBleuScore << ") ";
// set bleu score to zero in the feature vector since we do not want to optimise its weight
setBleuScore(featureValues.back(), 0);
}
// prepare translations to return
vector< vector<const Word*> > translations;
for (size_t i=0; i < numReturnedTranslations && i < nBestList.GetSize(); ++i) {
const TrellisPath &path = nBestList.at(i);
Phrase phrase = path.GetTargetPhrase();
vector<const Word*> translation;
for (size_t pos = 0; pos < phrase.GetSize(); ++pos) {
const Word &word = phrase.GetWord(pos);
Word *newWord = new Word(word);
translation.push_back(newWord);
}
translations.push_back(translation);
}
return translations;
}
const TrellisPath doMBR(const TrellisPathList& nBestList){
float marginal = 0;
vector<float> joint_prob_vec;
vector< vector<const Factor*> > translations;
float joint_prob;
vector< map < vector <const Factor *>, int > > ngram_stats;
TrellisPathList::const_iterator iter;
// get max score to prevent underflow
float maxScore = -1e20;
for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter)
{
const TrellisPath &path = **iter;
float score = StaticData::Instance().GetMBRScale()
* path.GetScoreBreakdown().InnerProduct(StaticData::Instance().GetAllWeights());
if (maxScore < score) maxScore = score;
}
for (iter = nBestList.begin() ; iter != nBestList.end() ; ++iter)
{
const TrellisPath &path = **iter;
joint_prob = UntransformScore(StaticData::Instance().GetMBRScale() * path.GetScoreBreakdown().InnerProduct(StaticData::Instance().GetAllWeights()) - maxScore);
marginal += joint_prob;
joint_prob_vec.push_back(joint_prob);
// get words in translation
vector<const Factor*> translation;
GetOutputFactors(path, translation);
// collect n-gram counts
map < vector < const Factor *>, int > counts;
extract_ngrams(translation,counts);
ngram_stats.push_back(counts);
translations.push_back(translation);
}
vector<float> mbr_loss;
float bleu, weightedLoss;
float weightedLossCumul = 0;
float minMBRLoss = 1000000;
int minMBRLossIdx = -1;
/* Main MBR computation done here */
iter = nBestList.begin();
for (unsigned int i = 0; i < nBestList.GetSize(); i++){
weightedLossCumul = 0;
for (unsigned int j = 0; j < nBestList.GetSize(); j++){
if ( i != j) {
bleu = calculate_score(translations, j, i,ngram_stats );
weightedLoss = ( 1 - bleu) * ( joint_prob_vec[j]/marginal);
weightedLossCumul += weightedLoss;
if (weightedLossCumul > minMBRLoss)
break;
}
}
if (weightedLossCumul < minMBRLoss){
minMBRLoss = weightedLossCumul;
minMBRLossIdx = i;
}
iter++;
}
/* Find sentence that minimises Bayes Risk under 1- BLEU loss */
return nBestList.at(minMBRLossIdx);
//return translations[minMBRLossIdx];
}
