bool operator==(const ScoreStats& s1, const ScoreStats& s2)
{
size_t size = s1.size();
if (size != s2.size())
return false;
for (size_t k=0; k < size; k++) {
if (s1.get(k) != s2.get(k))
return false;
}
return true;
}
// really not the right place...
float sentenceLevelBleuPlusOne( ScoreStats &stats ) {
float logbleu = 0.0;
const unsigned int bleu_order = 4;
for (unsigned int j=0; j<bleu_order; j++) {
//cerr << (stats.get(2*j)+1) << "/" << (stats.get(2*j+1)+1) << " ";
logbleu += log(stats.get(2*j)+1) - log(stats.get(2*j+1)+1);
}
logbleu /= bleu_order;
float brevity = 1.0 - (float)stats.get(bleu_order*2)/stats.get(1);
if (brevity < 0.0) {
logbleu += brevity;
}
//cerr << brevity << " -> " << exp(logbleu) << endl;
return exp(logbleu);
}
void InterpolatedScorer::setScoreData(ScoreData* data)
{
size_t last = 0;
m_score_data = data;
for (ScopedVector<Scorer>::iterator itsc = m_scorers.begin();
itsc != m_scorers.end(); ++itsc) {
int numScoresScorer = (*itsc)->NumberOfScores();
ScoreData* newData =new ScoreData(*itsc);
for (size_t i = 0; i < data->size(); i++) {
ScoreArray scoreArray = data->get(i);
ScoreArray newScoreArray;
std::string istr;
std::stringstream out;
out << i;
istr = out.str();
size_t numNBest = scoreArray.size();
//cout << " Datasize " << data->size() << " NumNBest " << numNBest << endl ;
for (size_t j = 0; j < numNBest ; j++) {
ScoreStats scoreStats = data->get(i, j);
//cout << "Scorestats " << scoreStats << " i " << i << " j " << j << endl;
ScoreStats newScoreStats;
for (size_t k = last; k < size_t(numScoresScorer + last); k++) {
ScoreStatsType score = scoreStats.get(k);
newScoreStats.add(score);
}
//cout << " last " << last << " NumScores " << numScoresScorer << "newScorestats " << newScoreStats << endl;
newScoreArray.add(newScoreStats);
}
newScoreArray.setIndex(istr);
newData->add(newScoreArray);
}
//newData->dump();
// NOTE: This class takes the ownership of the heap allocated
// ScoreData objects to avoid the memory leak issues.
m_scorers_score_data.push_back(newData);
(*itsc)->setScoreData(newData);
last += numScoresScorer;
}
}
void StatisticsBasedScorer::score(const candidates_t& candidates, const diffs_t& diffs,
statscores_t& scores) const
{
if (!m_score_data) {
throw runtime_error("Score data not loaded");
}
// calculate the score for the candidates
if (m_score_data->size() == 0) {
throw runtime_error("Score data is empty");
}
if (candidates.size() == 0) {
throw runtime_error("No candidates supplied");
}
int numCounts = m_score_data->get(0,candidates[0]).size();
vector<int> totals(numCounts);
for (size_t i = 0; i < candidates.size(); ++i) {
ScoreStats stats = m_score_data->get(i,candidates[i]);
if (stats.size() != totals.size()) {
stringstream msg;
msg << "Statistics for (" << "," << candidates[i] << ") have incorrect "
<< "number of fields. Found: " << stats.size() << " Expected: "
<< totals.size();
throw runtime_error(msg.str());
}
for (size_t k = 0; k < totals.size(); ++k) {
totals[k] += stats.get(k);
}
}
scores.push_back(calculateScore(totals));
candidates_t last_candidates(candidates);
// apply each of the diffs, and get new scores
for (size_t i = 0; i < diffs.size(); ++i) {
for (size_t j = 0; j < diffs[i].size(); ++j) {
size_t sid = diffs[i][j].first;
size_t nid = diffs[i][j].second;
size_t last_nid = last_candidates[sid];
for (size_t k = 0; k < totals.size(); ++k) {
int diff = m_score_data->get(sid,nid).get(k)
- m_score_data->get(sid,last_nid).get(k);
totals[k] += diff;
}
last_candidates[sid] = nid;
}
scores.push_back(calculateScore(totals));
}
// Regularisation. This can either be none, or the min or average as described in
// Cer, Jurafsky and Manning at WMT08.
if (m_regularization_type == NONE || m_regularization_window <= 0) {
// no regularisation
return;
}
// window size specifies the +/- in each direction
statscores_t raw_scores(scores); // copy scores
for (size_t i = 0; i < scores.size(); ++i) {
size_t start = 0;
if (i >= m_regularization_window) {
start = i - m_regularization_window;
}
const size_t end = min(scores.size(), i + m_regularization_window + 1);
if (m_regularization_type == AVERAGE) {
scores[i] = score_average(raw_scores,start,end);
} else {
scores[i] = score_min(raw_scores,start,end);
}
}
}