void extract(SentenceAlignment &sentence) {
int countE = sentence.target.size();
int countF = sentence.source.size();
HPhraseVector inboundPhrases;
HSentenceVertices inTopLeft;
HSentenceVertices inTopRight;
HSentenceVertices inBottomLeft;
HSentenceVertices inBottomRight;
HSentenceVertices outTopLeft;
HSentenceVertices outTopRight;
HSentenceVertices outBottomLeft;
HSentenceVertices outBottomRight;
HSentenceVertices::const_iterator it;
bool relaxLimit = hierModel;
bool buildExtraStructure = phraseModel || hierModel;
// check alignments for target phrase startE...endE
// loop over extracted phrases which are compatible with the word-alignments
for(int startE=0; startE<countE; ++startE) {
for(int endE=startE; (endE<countE && (relaxLimit || endE<startE+maxPhraseLength)); ++endE) {
int minF = 9999;
int maxF = -1;
vector<int> usedF = sentence.alignedCountS;
for(int ei=startE;ei<=endE;++ei)
for(int i=0;i<sentence.alignedToT[ei].size();++i) {
int fi = sentence.alignedToT[ei][i];
if (fi<minF) { minF = fi; }
if (fi>maxF) { maxF = fi; }
--usedF[fi];
}
if (maxF >= 0 && // aligned to any source words at all
(relaxLimit || maxF-minF < maxPhraseLength)) { // source phrase within limits
// check if source words are aligned to out of bound target words
bool out_of_bounds = false;
for(int fi=minF;fi<=maxF && !out_of_bounds;++fi)
out_of_bounds = usedF[fi] > 0;
// cout << "doing if for ( " << minF << "-" << maxF << ", " << startE << "," << endE << ")\n";
if (!out_of_bounds){
// start point of source phrase may retreat over unaligned
for(int startF=minF;
(startF>=0 &&
(relaxLimit || startF>maxF-maxPhraseLength) && // within length limit
(startF==minF || sentence.alignedCountS[startF]==0)); // unaligned
--startF)
// end point of source phrase may advance over unaligned
for(int endF=maxF;
(endF<countF &&
(relaxLimit || endF<startF+maxPhraseLength) && // within length limit
(endF==maxF || sentence.alignedCountS[endF]==0)); // unaligned
++endF){ // at this point we have extracted a phrase
if(buildExtraStructure){ // phrase || hier
if(endE-startE < maxPhraseLength && endF-startF < maxPhraseLength){ // within limit
inboundPhrases.push_back(HPhrase(HPhraseVertex(startF,startE),
HPhraseVertex(endF,endE)));
insertPhraseVertices(inTopLeft, inTopRight, inBottomLeft, inBottomRight,
startF, startE, endF, endE);
} else
insertPhraseVertices(outTopLeft, outTopRight, outBottomLeft, outBottomRight,
startF, startE, endF, endE);
} else {
string orientationInfo = "";
if(wordModel) {
REO_POS wordPrevOrient, wordNextOrient;
bool connectedLeftTopP = isAligned( sentence, startF-1, startE-1 );
bool connectedRightTopP = isAligned( sentence, endF+1, startE-1 );
bool connectedLeftTopN = isAligned( sentence, endF+1, endE+1 );
bool connectedRightTopN = isAligned( sentence, startF-1, endE+1 );
wordPrevOrient = getOrientWordModel(sentence, wordType, connectedLeftTopP, connectedRightTopP, startF, endF, startE, endE, countF, 0, 1, &ge, <);
wordNextOrient = getOrientWordModel(sentence, wordType, connectedLeftTopN, connectedRightTopN, endF, startF, endE, startE, 0, countF, -1, <, &ge);
orientationInfo += getOrientString(wordPrevOrient, wordType) + " " + getOrientString(wordNextOrient, wordType);
// if(allModelsOutputFlag)
// " | | ";
}
addPhrase(sentence, startE, endE, startF, endF, orientationInfo);
}
}
}
}
}
}
if(buildExtraStructure){ // phrase || hier
string orientationInfo = "";
REO_POS wordPrevOrient, wordNextOrient, phrasePrevOrient, phraseNextOrient, hierPrevOrient, hierNextOrient;
for(size_t i = 0; i < inboundPhrases.size(); ++i){
int startF = inboundPhrases[i].first.first;
int startE = inboundPhrases[i].first.second;
int endF = inboundPhrases[i].second.first;
int endE = inboundPhrases[i].second.second;
bool connectedLeftTopP = isAligned(sentence, startF-1, startE-1);
bool connectedRightTopP = isAligned(sentence, endF+1, startE-1);
bool connectedLeftTopN = isAligned(sentence, endF+1, endE+1);
bool connectedRightTopN = isAligned(sentence, startF-1, endE+1);
if(wordModel){
wordPrevOrient = getOrientWordModel(sentence, wordType,
connectedLeftTopP, connectedRightTopP,
startF, endF, startE, endE, countF, 0, 1,
&ge, <);
wordNextOrient = getOrientWordModel(sentence, wordType,
connectedLeftTopN, connectedRightTopN,
endF, startF, endE, startE, 0, countF, -1,
<, &ge);
}
if (phraseModel) {
phrasePrevOrient = getOrientPhraseModel(sentence, phraseType,
connectedLeftTopP, connectedRightTopP,
startF, endF, startE, endE, countF-1, 0, 1, &ge, <, inBottomRight, inBottomLeft);
phraseNextOrient = getOrientPhraseModel(sentence, phraseType,
connectedLeftTopN, connectedRightTopN,
endF, startF, endE, startE, 0, countF-1, -1, <, &ge, inBottomLeft, inBottomRight);
} else
phrasePrevOrient = phraseNextOrient = UNKNOWN;
if(hierModel){
hierPrevOrient = getOrientHierModel(sentence, hierType,
connectedLeftTopP, connectedRightTopP,
startF, endF, startE, endE, countF-1, 0, 1, &ge, <, inBottomRight, inBottomLeft, outBottomRight, outBottomLeft, phrasePrevOrient);
hierNextOrient = getOrientHierModel(sentence, hierType,
connectedLeftTopN, connectedRightTopN,
endF, startF, endE, startE, 0, countF-1, -1, <, &ge, inBottomLeft, inBottomRight, outBottomLeft, outBottomRight, phraseNextOrient);
}
orientationInfo = ((wordModel)? getOrientString(wordPrevOrient, wordType) + " " + getOrientString(wordNextOrient, wordType) : "") + " | " +
((phraseModel)? getOrientString(phrasePrevOrient, phraseType) + " " + getOrientString(phraseNextOrient, phraseType) : "") + " | " +
((hierModel)? getOrientString(hierPrevOrient, hierType) + " " + getOrientString(hierNextOrient, hierType) : "");
addPhrase(sentence, startE, endE, startF, endF, orientationInfo);
}
}
}
void ExtractTask::extract(SentenceAlignment &sentence)
{
int countE = sentence.target.size();
int countF = sentence.source.size();
HPhraseVector inboundPhrases;
HSentenceVertices inTopLeft;
HSentenceVertices inTopRight;
HSentenceVertices inBottomLeft;
HSentenceVertices inBottomRight;
HSentenceVertices outTopLeft;
HSentenceVertices outTopRight;
HSentenceVertices outBottomLeft;
HSentenceVertices outBottomRight;
HSentenceVertices::const_iterator it;
bool relaxLimit = m_options.isHierModel();
bool buildExtraStructure = m_options.isPhraseModel() || m_options.isHierModel();
// check alignments for target phrase startE...endE
// loop over extracted phrases which are compatible with the word-alignments
for(int startE=0; startE<countE; startE++) {
for(int endE=startE;
(endE<countE && (relaxLimit || endE<startE+m_options.maxPhraseLength));
endE++) {
int minF = 9999;
int maxF = -1;
vector< int > usedF = sentence.alignedCountS;
for(int ei=startE; ei<=endE; ei++) {
for(size_t i=0; i<sentence.alignedToT[ei].size(); i++) {
int fi = sentence.alignedToT[ei][i];
if (fi<minF) {
minF = fi;
}
if (fi>maxF) {
maxF = fi;
}
usedF[ fi ]--;
}
}
if (maxF >= 0 && // aligned to any source words at all
(relaxLimit || maxF-minF < m_options.maxPhraseLength)) { // source phrase within limits
// check if source words are aligned to out of bound target words
bool out_of_bounds = false;
for(int fi=minF; fi<=maxF && !out_of_bounds; fi++)
if (usedF[fi]>0) {
// cout << "ouf of bounds: " << fi << "\n";
out_of_bounds = true;
}
// cout << "doing if for ( " << minF << "-" << maxF << ", " << startE << "," << endE << ")\n";
if (!out_of_bounds) {
// start point of source phrase may retreat over unaligned
for(int startF=minF;
(startF>=0 &&
(relaxLimit || startF>maxF-m_options.maxPhraseLength) && // within length limit
(startF==minF || sentence.alignedCountS[startF]==0)); // unaligned
startF--)
// end point of source phrase may advance over unaligned
for(int endF=maxF;
(endF<countF &&
(relaxLimit || endF<startF+m_options.maxPhraseLength) && // within length limit
(endF - startF + 1 > m_options.minPhraseLength) && // within length limit
(endF==maxF || sentence.alignedCountS[endF]==0)); // unaligned
endF++) { // at this point we have extracted a phrase
if(buildExtraStructure) { // phrase || hier
if(endE-startE < m_options.maxPhraseLength && endF-startF < m_options.maxPhraseLength) { // within limit
inboundPhrases.push_back(HPhrase(HPhraseVertex(startF,startE),
HPhraseVertex(endF,endE)));
insertPhraseVertices(inTopLeft, inTopRight, inBottomLeft, inBottomRight,
startF, startE, endF, endE);
} else
insertPhraseVertices(outTopLeft, outTopRight, outBottomLeft, outBottomRight,
startF, startE, endF, endE);
} else {
string orientationInfo = "";
if(m_options.isWordModel()) {
REO_POS wordPrevOrient, wordNextOrient;
bool connectedLeftTopP = isAligned( sentence, startF-1, startE-1 );
bool connectedRightTopP = isAligned( sentence, endF+1, startE-1 );
bool connectedLeftTopN = isAligned( sentence, endF+1, endE+1 );
bool connectedRightTopN = isAligned( sentence, startF-1, endE+1 );
wordPrevOrient = getOrientWordModel(sentence, m_options.isWordType(), connectedLeftTopP, connectedRightTopP, startF, endF, startE, endE, countF, 0, 1, &ge, <);
wordNextOrient = getOrientWordModel(sentence, m_options.isWordType(), connectedLeftTopN, connectedRightTopN, endF, startF, endE, startE, 0, countF, -1, <, &ge);
orientationInfo += getOrientString(wordPrevOrient, m_options.isWordType()) + " " + getOrientString(wordNextOrient, m_options.isWordType());
if(m_options.isAllModelsOutputFlag())
" | | ";
}
addPhrase(sentence, startE, endE, startF, endF, orientationInfo);
}
}
}
}
}
}
}
