void Data::sampleRankedPairs( const std::string &rankedpairfile ) {
cout << "Sampling ranked pairs." << endl;
ofstream *outFile = new ofstream();
outFile->open( rankedpairfile.c_str() );
ostream *out = outFile;
const unsigned int n_samplings = 5000;
const unsigned int n_samples = 50;
const float min_diff = 0.05;
// loop over all sentences
for(unsigned int S=0; S<featdata->size(); S++) {
unsigned int n_translations = featdata->get(S).size();
// sample a fixed number of times
vector< SampledPair* > samples;
vector< float > scores;
for(unsigned int i=0; i<n_samplings; i++) {
unsigned int translation1 = rand() % n_translations;
float bleu1 = sentenceLevelBleuPlusOne(scoredata->get(S,translation1));
unsigned int translation2 = rand() % n_translations;
float bleu2 = sentenceLevelBleuPlusOne(scoredata->get(S,translation2));
if (abs(bleu1-bleu2) < min_diff)
continue;
samples.push_back( new SampledPair( translation1, translation2, bleu1-bleu2) );
scores.push_back( 1.0 - abs(bleu1-bleu2) );
}
//cerr << "sampled " << samples.size() << " pairs\n";
float min_diff = -1.0;
if (samples.size() > n_samples) {
nth_element(scores.begin(), scores.begin()+(n_samples-1), scores.end());
min_diff = 0.99999-scores[n_samples-1];
//cerr << "min_diff = " << min_diff << endl;
}
unsigned int collected = 0;
for(unsigned int i=0; i<samples.size() && collected < n_samples; i++) {
if (samples[i]->getDiff() >= min_diff) {
collected++;
*out << "1";
outputSample( *out, featdata->get(S,samples[i]->getTranslation1()),
featdata->get(S,samples[i]->getTranslation2()) );
*out << endl;
*out << "0";
outputSample( *out, featdata->get(S,samples[i]->getTranslation2()),
featdata->get(S,samples[i]->getTranslation1()) );
*out << endl;
}
delete samples[i];
}
//cerr << "collected " << collected << endl;
}
out->flush();
outFile->close();
}
vector<float> BleuScorer::ScoreNbestList(const string& scoreFile, const string& featureFile) {
vector<string> scoreFiles;
vector<string> featureFiles;
scoreFiles.push_back(scoreFile);
featureFiles.push_back(featureFile);
vector<FeatureDataIterator> featureDataIters;
vector<ScoreDataIterator> scoreDataIters;
for (size_t i = 0; i < featureFiles.size(); ++i) {
featureDataIters.push_back(FeatureDataIterator(featureFiles[i]));
scoreDataIters.push_back(ScoreDataIterator(scoreFiles[i]));
}
vector<pair<size_t,size_t> > hypotheses;
if (featureDataIters[0] == FeatureDataIterator::end()) {
cerr << "Error: at the end of feature data iterator" << endl;
exit(1);
}
for (size_t i = 0; i < featureFiles.size(); ++i) {
if (featureDataIters[i] == FeatureDataIterator::end()) {
cerr << "Error: Feature file " << i << " ended prematurely" << endl;
exit(1);
}
if (scoreDataIters[i] == ScoreDataIterator::end()) {
cerr << "Error: Score file " << i << " ended prematurely" << endl;
exit(1);
}
if (featureDataIters[i]->size() != scoreDataIters[i]->size()) {
cerr << "Error: features and scores have different size" << endl;
exit(1);
}
for (size_t j = 0; j < featureDataIters[i]->size(); ++j) {
hypotheses.push_back(pair<size_t,size_t>(i,j));
}
}
// score the nbest list
vector<float> bleuScores;
for (size_t i=0; i < hypotheses.size(); ++i) {
pair<size_t,size_t> translation = hypotheses[i];
float bleu = sentenceLevelBleuPlusOne(scoreDataIters[translation.first]->operator[](translation.second));
bleuScores.push_back(bleu);
}
return bleuScores;
}