void PhraseTableCreator::EncodeTargetPhraseREnc(std::vector<std::string>& s,
std::vector<std::string>& t,
std::set<AlignPoint>& a,
std::ostream& os)
{
std::stringstream encodedTargetPhrase;
std::vector<std::vector<size_t> > a2(t.size());
for(std::set<AlignPoint>::iterator it = a.begin(); it != a.end(); it++)
a2[it->second].push_back(it->first);
for(size_t i = 0; i < t.size(); i++) {
unsigned idxTarget = GetOrAddTargetSymbolId(t[i]);
unsigned encodedSymbol = -1;
unsigned bestSrcPos = s.size();
unsigned bestDiff = s.size();
unsigned bestRank = m_lexicalTable.size();
unsigned badRank = m_lexicalTable.size();
for(std::vector<size_t>::iterator it = a2[i].begin(); it != a2[i].end(); it++) {
unsigned idxSource = GetSourceSymbolId(s[*it]);
size_t r = GetRank(idxSource, idxTarget);
if(r != badRank) {
if(r < bestRank) {
bestRank = r;
bestSrcPos = *it;
bestDiff = abs(*it-i);
} else if(r == bestRank && unsigned(abs(*it-i)) < bestDiff) {
bestSrcPos = *it;
bestDiff = abs(*it-i);
}
}
}
if(bestRank != badRank && bestSrcPos < s.size()) {
if(bestSrcPos == i)
encodedSymbol = EncodeREncSymbol3(bestRank);
else
encodedSymbol = EncodeREncSymbol2(bestSrcPos, bestRank);
a.erase(AlignPoint(bestSrcPos, i));
} else {
encodedSymbol = EncodeREncSymbol1(idxTarget);
}
os.write((char*)&encodedSymbol, sizeof(encodedSymbol));
m_symbolCounter.Increase(encodedSymbol);
}
unsigned stopSymbolId = GetTargetSymbolId(m_phraseStopSymbol);
unsigned encodedSymbol = EncodeREncSymbol1(stopSymbolId);
os.write((char*)&encodedSymbol, sizeof(encodedSymbol));
m_symbolCounter.Increase(encodedSymbol);
}
std::string PhraseTableCreator::EncodeLine(std::vector<std::string>& tokens, size_t ownRank)
{
std::string sourcePhraseStr = tokens[0];
std::string targetPhraseStr = tokens[1];
std::string scoresStr = tokens[2];
std::string alignmentStr = "";
if(tokens.size() > 3)
alignmentStr = tokens[3];
std::vector<std::string> s = Tokenize(sourcePhraseStr);
size_t phraseLength = s.size();
if(m_maxPhraseLength < phraseLength)
m_maxPhraseLength = phraseLength;
std::vector<std::string> t = Tokenize(targetPhraseStr);
std::vector<float> scores = Tokenize<float>(scoresStr);
if(scores.size() != m_numScoreComponent) {
std::cerr << "Error: Wrong number of scores detected ("
<< scores.size() << " != " << m_numScoreComponent << ") :" << std::endl;
std::cerr << "Line: " << tokens[0] << " ||| " << tokens[1] << " ||| " << tokens[3] << " ..." << std::endl;
abort();
}
std::set<AlignPoint> a;
if(m_coding != None || m_useAlignmentInfo) {
std::vector<size_t> positions = Tokenize<size_t>(alignmentStr, " \t-");
for(size_t i = 0; i < positions.size(); i += 2) {
a.insert(AlignPoint(positions[i], positions[i+1]));
}
}
std::stringstream encodedTargetPhrase;
if(m_coding == PREnc) {
EncodeTargetPhrasePREnc(s, t, a, ownRank, encodedTargetPhrase);
} else if(m_coding == REnc) {
EncodeTargetPhraseREnc(s, t, a, encodedTargetPhrase);
} else {
EncodeTargetPhraseNone(t, encodedTargetPhrase);
}
EncodeScores(scores, encodedTargetPhrase);
if(m_useAlignmentInfo)
EncodeAlignment(a, encodedTargetPhrase);
return encodedTargetPhrase.str();
}
TargetPhraseVectorPtr PhraseDecoder::DecodeCollection(
TargetPhraseVectorPtr tpv, BitWrapper<> &encodedBitStream,
const Phrase &sourcePhrase, bool topLevel)
{
bool extending = tpv->size();
size_t bitsLeft = encodedBitStream.TellFromEnd();
typedef std::pair<size_t, size_t> AlignPointSizeT;
std::vector<int> sourceWords;
if(m_coding == REnc)
{
for(size_t i = 0; i < sourcePhrase.GetSize(); i++)
{
std::string sourceWord
= sourcePhrase.GetWord(i).GetString(*m_input, false);
unsigned idx = GetSourceSymbolId(sourceWord);
sourceWords.push_back(idx);
}
}
unsigned phraseStopSymbol = 0;
AlignPoint alignStopSymbol(-1, -1);
std::vector<float> scores;
std::set<AlignPointSizeT> alignment;
enum DecodeState { New, Symbol, Score, Alignment, Add } state = New;
size_t srcSize = sourcePhrase.GetSize();
TargetPhrase* targetPhrase = NULL;
while(encodedBitStream.TellFromEnd())
{
if(state == New)
{
// Creating new TargetPhrase on the heap
tpv->push_back(TargetPhrase(Output));
targetPhrase = &tpv->back();
targetPhrase->SetSourcePhrase(sourcePhrase);
alignment.clear();
scores.clear();
state = Symbol;
}
if(state == Symbol)
{
unsigned symbol = m_symbolTree->Read(encodedBitStream);
if(symbol == phraseStopSymbol)
{
state = Score;
}
else
{
if(m_coding == REnc)
{
std::string wordString;
size_t type = GetREncType(symbol);
if(type == 1)
{
unsigned decodedSymbol = DecodeREncSymbol1(symbol);
wordString = GetTargetSymbol(decodedSymbol);
}
else if (type == 2)
{
size_t rank = DecodeREncSymbol2Rank(symbol);
size_t srcPos = DecodeREncSymbol2Position(symbol);
if(srcPos >= sourceWords.size())
return TargetPhraseVectorPtr();
wordString = GetTargetSymbol(GetTranslation(sourceWords[srcPos], rank));
if(m_phraseDictionary.m_useAlignmentInfo)
{
size_t trgPos = targetPhrase->GetSize();
alignment.insert(AlignPoint(srcPos, trgPos));
}
}
else if(type == 3)
{
size_t rank = DecodeREncSymbol3(symbol);
size_t srcPos = targetPhrase->GetSize();
if(srcPos >= sourceWords.size())
return TargetPhraseVectorPtr();
wordString = GetTargetSymbol(GetTranslation(sourceWords[srcPos], rank));
if(m_phraseDictionary.m_useAlignmentInfo)
{
size_t trgPos = srcPos;
alignment.insert(AlignPoint(srcPos, trgPos));
}
}
Word word;
word.CreateFromString(Output, *m_output, wordString, false);
targetPhrase->AddWord(word);
}
else if(m_coding == PREnc)
{
// if the symbol is just a word
if(GetPREncType(symbol) == 1)
{
unsigned decodedSymbol = DecodePREncSymbol1(symbol);
Word word;
word.CreateFromString(Output, *m_output,
GetTargetSymbol(decodedSymbol), false);
targetPhrase->AddWord(word);
}
// if the symbol is a subphrase pointer
else
{
int left = DecodePREncSymbol2Left(symbol);
int right = DecodePREncSymbol2Right(symbol);
unsigned rank = DecodePREncSymbol2Rank(symbol);
int srcStart = left + targetPhrase->GetSize();
int srcEnd = srcSize - right - 1;
// false positive consistency check
if(0 > srcStart || srcStart > srcEnd || unsigned(srcEnd) >= srcSize)
return TargetPhraseVectorPtr();
// false positive consistency check
if(m_maxRank && rank > m_maxRank)
return TargetPhraseVectorPtr();
// set subphrase by default to itself
TargetPhraseVectorPtr subTpv = tpv;
// if range smaller than source phrase retrieve subphrase
if(unsigned(srcEnd - srcStart + 1) != srcSize)
{
Phrase subPhrase = sourcePhrase.GetSubString(WordsRange(srcStart, srcEnd));
subTpv = CreateTargetPhraseCollection(subPhrase, false);
}
// false positive consistency check
if(subTpv != NULL && rank < subTpv->size())
{
// insert the subphrase into the main target phrase
TargetPhrase& subTp = subTpv->at(rank);
if(m_phraseDictionary.m_useAlignmentInfo)
{
// reconstruct the alignment data based on the alignment of the subphrase
for(AlignmentInfo::const_iterator it = subTp.GetAlignmentInfo().begin();
it != subTp.GetAlignmentInfo().end(); it++)
{
alignment.insert(AlignPointSizeT(srcStart + it->first,
targetPhrase->GetSize() + it->second));
}
}
targetPhrase->Append(subTp);
}
else
return TargetPhraseVectorPtr();
}
}
else
{
Word word;
word.CreateFromString(Output, *m_output,
GetTargetSymbol(symbol), false);
targetPhrase->AddWord(word);
}
}
}
else if(state == Score)
{
size_t idx = m_multipleScoreTrees ? scores.size() : 0;
float score = m_scoreTrees[idx]->Read(encodedBitStream);
scores.push_back(score);
if(scores.size() == m_numScoreComponent)
{
targetPhrase->SetScore(m_feature, scores, ScoreComponentCollection() /*sparse*/,*m_weight, m_weightWP, *m_languageModels);
if(m_containsAlignmentInfo)
state = Alignment;
else
state = Add;
}
}
else if(state == Alignment)
{
AlignPoint alignPoint = m_alignTree->Read(encodedBitStream);
if(alignPoint == alignStopSymbol)
{
state = Add;
}
else
{
if(m_phraseDictionary.m_useAlignmentInfo)
alignment.insert(AlignPointSizeT(alignPoint));
}
}
if(state == Add)
{
if(m_phraseDictionary.m_useAlignmentInfo)
targetPhrase->SetAlignmentInfo(alignment);
if(m_coding == PREnc)
{
if(!m_maxRank || tpv->size() <= m_maxRank)
bitsLeft = encodedBitStream.TellFromEnd();
if(!topLevel && m_maxRank && tpv->size() >= m_maxRank)
break;
}
if(encodedBitStream.TellFromEnd() <= 8)
break;
state = New;
}
}
if(m_coding == PREnc && !extending)
{
bitsLeft = bitsLeft > 8 ? bitsLeft : 0;
m_decodingCache.Cache(sourcePhrase, tpv, bitsLeft, m_maxRank);
}
return tpv;
}