mosesdecoder/mert/Scorer.cpp
bhaddow 83f234cf17 Implementation of Cer et al mert regularisation. Use with argument such
as --scconfig regtype:min,regwin:3 in extractor and mert. Only tested
on toy example so far.


git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@1860 1f5c12ca-751b-0410-a591-d2e778427230
2008-06-24 19:27:18 +00:00

98 lines
3.0 KiB
C++

#include "Scorer.h"
//regularisation strategies
static float score_min(const statscores_t& scores, size_t start, size_t end) {
float min = numeric_limits<float>::max();
for (size_t i = start; i < end; ++i) {
if (scores[i] < min) {
min = scores[i];
}
}
return min;
}
static float score_average(const statscores_t& scores, size_t start, size_t end) {
if ((end - start) < 1) {
//shouldn't happen
return 0;
}
float total = 0;
for (size_t j = start; j < end; ++j) {
total += scores[j];
}
return total / (end - start);
}
void StatisticsBasedScorer::score(const candidates_t& candidates, const diffs_t& diffs,
statscores_t& scores) {
if (!_scoreData) {
throw runtime_error("Score data not loaded");
}
//calculate the score for the candidates
if (_scoreData->size() == 0) {
throw runtime_error("Score data is empty");
}
if (candidates.size() == 0) {
throw runtime_error("No candidates supplied");
}
int numCounts = _scoreData->get(0,candidates[0]).size();
vector<int> totals(numCounts);
for (size_t i = 0; i < candidates.size(); ++i) {
ScoreStats stats = _scoreData->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 = _scoreData->get(sid,nid).get(k)
- _scoreData->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 (_regularisationStrategy == REG_NONE || _regularisationWindow <= 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 >= _regularisationWindow) {
start = i - _regularisationWindow;
}
size_t end = min(scores.size(), i + _regularisationWindow+1);
if (_regularisationStrategy == REG_AVERAGE) {
scores[i] = score_average(raw_scores,start,end);
} else {
scores[i] = score_min(raw_scores,start,end);
}
}
}