void ExtractTask::saveHieroAlignment( int startT, int endT, int startS, int endS
, const WordIndex &indexS, const WordIndex &indexT, HoleCollection &holeColl, ExtractedRule &rule)
{
// print alignment of words
for(int ti=startT; ti<=endT; ti++) {
WordIndex::const_iterator p = indexT.find(ti);
if (p != indexT.end()) { // does word still exist?
for(unsigned int i=0; i<m_sentence.alignedToT[ti].size(); i++) {
int si = m_sentence.alignedToT[ti][i];
std::string sourceSymbolIndex = IntToString(indexS.find(si)->second);
std::string targetSymbolIndex = IntToString(p->second);
rule.alignment += sourceSymbolIndex + "-" + targetSymbolIndex + " ";
if (! m_options.onlyDirectFlag)
rule.alignmentInv += targetSymbolIndex + "-" + sourceSymbolIndex + " ";
}
}
}
// print alignment of non terminals
HoleList::const_iterator iterHole;
for (iterHole = holeColl.GetHoles().begin(); iterHole != holeColl.GetHoles().end(); ++iterHole) {
const Hole &hole = *iterHole;
std::string sourceSymbolIndex = IntToString(hole.GetPos(0));
std::string targetSymbolIndex = IntToString(hole.GetPos(1));
rule.alignment += sourceSymbolIndex + "-" + targetSymbolIndex + " ";
if (!m_options.onlyDirectFlag)
rule.alignmentInv += targetSymbolIndex + "-" + sourceSymbolIndex + " ";
}
rule.alignment.erase(rule.alignment.size()-1);
if (!m_options.onlyDirectFlag) {
rule.alignmentInv.erase(rule.alignmentInv.size()-1);
}
}
void ApParser::train(SentenceReader* sentenceReader, char const* modelFile)
{
WordIndex labelIndex;
vector<string> labels;
vector<string> predLabels;
// collect events
list<Tanl::Classifier::Event*> events;
WordCounts predCount; // count predicate occurrences
int evCount = 0;
Tanl::Classifier::PID pID = 1; // leave 0 for bias
// create inverted index of predicate names
// used to create vector of pIDs
EventStream eventStream(sentenceReader, &info);
while (eventStream.hasNext()) {
Tanl::Classifier::Event* ev = eventStream.next();
events.push_back(ev);
evCount++; // count them explicitly, since size() is costly
if (config.verbose) {
if (evCount % 10000 == 0)
cerr << '+' << flush;
else if (evCount % 1000 == 0)
cerr << '.' << flush;
}
vector<string>& ec = ev->features; // ec = {p1, ... , pn}
for (unsigned j = 0; j < ec.size(); j++) {
string& pred = ec[j];
// decide whether to retain it (# occurrences > cutoff)
if (predIndex.find(pred.c_str()) == predIndex.end()) {
// not yet among those retained
WordCounts::iterator wcit = predCount.find(pred);
// increment # of occurrences
int count;
if (wcit == predCount.end())
count = predCount[pred] = 1;
else
count = ++wcit->second;
if (count >= config.featureCutoff) {
predLabels.push_back(pred); // accept it into predLabels
predIndex[pred.c_str()] = pID++;
predCount.erase(pred);
}
}
}
}
if (config.verbose)
cerr << endl;
// build cases
Cases cases;
cases.reserve(evCount);
int n = 0;
Tanl::Classifier::ClassID oID = 0;
while (!events.empty()) {
Tanl::Classifier::Event* ev = events.front();
events.pop_front();
cases.push_back(Case());
X& x = cases[n].first; // features
// add features
vector<string>& ec = ev->features; // ec = {p1, ... , pn}
char const* c = ev->className.c_str();
for (unsigned j = 0; j < ec.size(); j++) {
string& pred = ec[j];
WordIndex::const_iterator pit = predIndex.find(pred.c_str());
if (pit != predIndex.end()) {
x.push_back(pit->second);
}
}
if (x.size()) {
if (labelIndex.find(c) == labelIndex.end()) {
labelIndex[c] = oID++;
labels.push_back(c);
}
cases[n].second = labelIndex[c];
n++;
if (config.verbose) {
if (n % 10000 == 0)
cerr << '+' << flush;
else if (n % 1000 == 0)
cerr << '.' << flush;
}
x.push_back(0); // bias
}
delete ev;
}
cases.resize(n);
if (config.verbose)
cerr << endl;
int predSize = predLabels.size();
predSize++; // bias
APSV ap(labels.size(), predSize);
ofstream ofs(modelFile, ios::binary | ios::trunc);
// dump configuration settings
config.writeHeader(ofs);
// dump labels
ofs << labels.size() << endl;
FOR_EACH (vector<string>, labels, pit)
ofs << *pit << endl;
// dump predLabels
ofs << predLabels.size() << endl;
FOR_EACH (vector<string>, predLabels, pit)
ofs << *pit << endl;
// free memory
predIndex.clear();
WordIndex().swap(predIndex); // STL map do not deallocate. resize(0) has no effect
labelIndex.clear();
WordIndex().swap(labelIndex);
// clear memory for unfrequent entities
info.clearRareEntities();
// perform training
ap.train(cases, iter);
// dump parameters
ap.save(ofs);
// dump global info
info.save(ofs);
}
