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;
}
void ScoreStats::Copy(const ScoreStats &stats)
{
m_available_size = stats.available();
m_entries = stats.size();
m_array = new ScoreStatsType[m_available_size];
memcpy(m_array, stats.getArray(), GetArraySizeWithBytes());
}
int main(int argc, char **argv)
{
if (argc == 1) {
cerr << "Usage: ./sentence-bleu ref1 [ref2 ...] < candidate > bleu-scores" << endl;
return 1;
}
vector<string> refFiles(argv + 1, argv + argc);
// TODO all of these are empty for now
string config;
string factors;
string filter;
BleuScorer scorer(config);
scorer.setFactors(factors);
scorer.setFilter(filter);
// initialize reference streams
vector<boost::shared_ptr<ifstream> > refStreams;
for (vector<string>::const_iterator refFile=refFiles.begin(); refFile!=refFiles.end(); ++refFile) {
TRACE_ERR("Loading reference from " << *refFile << endl);
boost::shared_ptr<ifstream> ifs(new ifstream(refFile->c_str()));
UTIL_THROW_IF2(!ifs, "Cannot open " << *refFile);
refStreams.push_back(ifs);
}
// load sentences, preparing statistics, score
string hypothesisLine;
size_t sid = 0;
while (getline(std::cin, hypothesisLine)) {
Reference ref;
if (!scorer.GetNextReferenceFromStreams(refStreams, ref)) {
UTIL_THROW2("Missing references");
}
ScoreStats scoreStats;
scorer.CalcBleuStats(ref, hypothesisLine, scoreStats);
vector<float> stats(scoreStats.getArray(), scoreStats.getArray() + scoreStats.size());
std::cout << smoothedSentenceBleu(stats) << std::endl;
++sid;
}
return 0;
}
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);
}
}
}