bool RuleTrieLoader::Load(const std::vector<FactorType> &input,
const std::vector<FactorType> &output,
const std::string &inFile,
const RuleTableFF &ff,
RuleTrie &trie)
{
PrintUserTime(std::string("Start loading text phrase table. Moses format"));
const StaticData &staticData = StaticData::Instance();
// const std::string &factorDelimiter = staticData.GetFactorDelimiter();
std::size_t count = 0;
std::ostream *progress = NULL;
IFVERBOSE(1) progress = &std::cerr;
util::FilePiece in(inFile.c_str(), progress);
// reused variables
std::vector<float> scoreVector;
StringPiece line;
double_conversion::StringToDoubleConverter converter(double_conversion::StringToDoubleConverter::NO_FLAGS, NAN, NAN, "inf", "nan");
while(true) {
try {
line = in.ReadLine();
} catch (const util::EndOfFileException &e) {
break;
}
util::TokenIter<util::MultiCharacter> pipes(line, "|||");
StringPiece sourcePhraseString(*pipes);
StringPiece targetPhraseString(*++pipes);
StringPiece scoreString(*++pipes);
StringPiece alignString;
if (++pipes) {
StringPiece temp(*pipes);
alignString = temp;
}
if (++pipes) {
StringPiece str(*pipes); //counts
}
bool isLHSEmpty = (sourcePhraseString.find_first_not_of(" \t", 0) == std::string::npos);
if (isLHSEmpty && !staticData.IsWordDeletionEnabled()) {
TRACE_ERR( ff.GetFilePath() << ":" << count << ": pt entry contains empty target, skipping\n");
continue;
}
scoreVector.clear();
for (util::TokenIter<util::AnyCharacter, true> s(scoreString, " \t"); s; ++s) {
int processed;
float score = converter.StringToFloat(s->data(), s->length(), &processed);
UTIL_THROW_IF2(std::isnan(score), "Bad score " << *s << " on line " << count);
scoreVector.push_back(FloorScore(TransformScore(score)));
}
const std::size_t numScoreComponents = ff.GetNumScoreComponents();
if (scoreVector.size() != numScoreComponents) {
UTIL_THROW2("Size of scoreVector != number (" << scoreVector.size() << "!="
<< numScoreComponents << ") of score components on line " << count);
}
// parse source & find pt node
// constituent labels
Word *sourceLHS = NULL;
Word *targetLHS;
// create target phrase obj
TargetPhrase *targetPhrase = new TargetPhrase(&ff);
// targetPhrase->CreateFromString(Output, output, targetPhraseString, factorDelimiter, &targetLHS);
targetPhrase->CreateFromString(Output, output, targetPhraseString, &targetLHS);
// source
Phrase sourcePhrase;
// sourcePhrase.CreateFromString(Input, input, sourcePhraseString, factorDelimiter, &sourceLHS);
sourcePhrase.CreateFromString(Input, input, sourcePhraseString, &sourceLHS);
// rest of target phrase
targetPhrase->SetAlignmentInfo(alignString);
targetPhrase->SetTargetLHS(targetLHS);
//targetPhrase->SetDebugOutput(string("New Format pt ") + line);
if (++pipes) {
StringPiece sparseString(*pipes);
targetPhrase->SetSparseScore(&ff, sparseString);
}
if (++pipes) {
StringPiece propertiesString(*pipes);
targetPhrase->SetProperties(propertiesString);
}
targetPhrase->GetScoreBreakdown().Assign(&ff, scoreVector);
targetPhrase->EvaluateInIsolation(sourcePhrase, ff.GetFeaturesToApply());
TargetPhraseCollection &phraseColl = GetOrCreateTargetPhraseCollection(
trie, *sourceLHS, sourcePhrase);
phraseColl.Add(targetPhrase);
// not implemented correctly in memory pt. just delete it for now
delete sourceLHS;
count++;
}
// sort and prune each target phrase collection
if (ff.GetTableLimit()) {
SortAndPrune(trie, ff.GetTableLimit());
}
return true;
}
void PhraseDictionaryFuzzyMatch::InitializeForInput(InputType const& inputSentence)
{
#if defined __MINGW32__
char dirName[] = "moses.XXXXXX";
#else
char dirName[] = "/tmp/moses.XXXXXX";
#endif // defined
char *temp = mkdtemp(dirName);
UTIL_THROW_IF2(temp == NULL,
"Couldn't create temporary directory " << dirName);
string dirNameStr(dirName);
string inFileName(dirNameStr + "/in");
ofstream inFile(inFileName.c_str());
for (size_t i = 1; i < inputSentence.GetSize() - 1; ++i) {
inFile << inputSentence.GetWord(i);
}
inFile << endl;
inFile.close();
long translationId = inputSentence.GetTranslationId();
string ptFileName = m_FuzzyMatchWrapper->Extract(translationId, dirNameStr);
// populate with rules for this sentence
PhraseDictionaryNodeMemory &rootNode = m_collection[translationId];
FormatType format = MosesFormat;
// data from file
InputFileStream inStream(ptFileName);
// copied from class LoaderStandard
PrintUserTime("Start loading fuzzy-match phrase model");
const StaticData &staticData = StaticData::Instance();
const std::string& factorDelimiter = staticData.GetFactorDelimiter();
string lineOrig;
size_t count = 0;
while(getline(inStream, lineOrig)) {
const string *line;
if (format == HieroFormat) { // reformat line
UTIL_THROW(util::Exception, "Cannot be Hiero format");
//line = ReformatHieroRule(lineOrig);
} else {
// do nothing to format of line
line = &lineOrig;
}
vector<string> tokens;
vector<float> scoreVector;
TokenizeMultiCharSeparator(tokens, *line , "|||" );
if (tokens.size() != 4 && tokens.size() != 5) {
UTIL_THROW2("Syntax error at " << ptFileName << ":" << count);
}
const string &sourcePhraseString = tokens[0]
, &targetPhraseString = tokens[1]
, &scoreString = tokens[2]
, &alignString = tokens[3];
bool isLHSEmpty = (sourcePhraseString.find_first_not_of(" \t", 0) == string::npos);
if (isLHSEmpty && !staticData.IsWordDeletionEnabled()) {
TRACE_ERR( ptFileName << ":" << count << ": pt entry contains empty target, skipping\n");
continue;
}
Tokenize<float>(scoreVector, scoreString);
const size_t numScoreComponents = GetNumScoreComponents();
if (scoreVector.size() != numScoreComponents) {
UTIL_THROW2("Size of scoreVector != number (" << scoreVector.size() << "!="
<< numScoreComponents << ") of score components on line " << count);
}
UTIL_THROW_IF2(scoreVector.size() != numScoreComponents,
"Number of scores incorrectly specified");
// parse source & find pt node
// constituent labels
Word *sourceLHS;
Word *targetLHS;
// source
Phrase sourcePhrase( 0);
// sourcePhrase.CreateFromString(Input, m_input, sourcePhraseString, factorDelimiter, &sourceLHS);
sourcePhrase.CreateFromString(Input, m_input, sourcePhraseString, &sourceLHS);
// create target phrase obj
TargetPhrase *targetPhrase = new TargetPhrase(this);
// targetPhrase->CreateFromString(Output, m_output, targetPhraseString, factorDelimiter, &targetLHS);
targetPhrase->CreateFromString(Output, m_output, targetPhraseString, &targetLHS);
// rest of target phrase
targetPhrase->SetAlignmentInfo(alignString);
targetPhrase->SetTargetLHS(targetLHS);
//targetPhrase->SetDebugOutput(string("New Format pt ") + line);
// component score, for n-best output
std::transform(scoreVector.begin(),scoreVector.end(),scoreVector.begin(),TransformScore);
std::transform(scoreVector.begin(),scoreVector.end(),scoreVector.begin(),FloorScore);
targetPhrase->GetScoreBreakdown().Assign(this, scoreVector);
targetPhrase->EvaluateInIsolation(sourcePhrase, GetFeaturesToApply());
TargetPhraseCollection &phraseColl = GetOrCreateTargetPhraseCollection(rootNode, sourcePhrase, *targetPhrase, sourceLHS);
phraseColl.Add(targetPhrase);
count++;
if (format == HieroFormat) { // reformat line
delete line;
} else {
// do nothing
}
}
// sort and prune each target phrase collection
SortAndPrune(rootNode);
//removedirectoryrecursively(dirName);
}
TargetPhraseVectorPtr PhraseDecoder::DecodeCollection(
TargetPhraseVectorPtr tpv, BitWrapper<> &encodedBitStream,
const Phrase &sourcePhrase, bool topLevel, bool eval)
{
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());
targetPhrase = &tpv->back();
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);
} else {
// false positive consistency check
if(rank >= tpv->size()-1)
return TargetPhraseVectorPtr();
}
// 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.GetAlignTerm().begin();
it != subTp.GetAlignTerm().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->GetScoreBreakdown().Assign(&m_phraseDictionary, scores);
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->SetAlignTerm(alignment);
}
if(eval) {
targetPhrase->EvaluateInIsolation(sourcePhrase, m_phraseDictionary.GetFeaturesToApply());
}
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;
}
void ChartParserUnknown::Process(const Word &sourceWord, const WordsRange &range, ChartParserCallback &to)
{
// unknown word, add as trans opt
const StaticData &staticData = StaticData::Instance();
const UnknownWordPenaltyProducer &unknownWordPenaltyProducer = UnknownWordPenaltyProducer::Instance();
size_t isDigit = 0;
if (staticData.GetDropUnknown()) {
const Factor *f = sourceWord[0]; // TODO hack. shouldn't know which factor is surface
const StringPiece s = f->GetString();
isDigit = s.find_first_of("0123456789");
if (isDigit == string::npos)
isDigit = 0;
else
isDigit = 1;
// modify the starting bitmap
}
Phrase* unksrc = new Phrase(1);
unksrc->AddWord() = sourceWord;
Word &newWord = unksrc->GetWord(0);
newWord.SetIsOOV(true);
m_unksrcs.push_back(unksrc);
//TranslationOption *transOpt;
if (! staticData.GetDropUnknown() || isDigit) {
// loop
const UnknownLHSList &lhsList = staticData.GetUnknownLHS();
UnknownLHSList::const_iterator iterLHS;
for (iterLHS = lhsList.begin(); iterLHS != lhsList.end(); ++iterLHS) {
const string &targetLHSStr = iterLHS->first;
float prob = iterLHS->second;
// lhs
//const Word &sourceLHS = staticData.GetInputDefaultNonTerminal();
Word *targetLHS = new Word(true);
targetLHS->CreateFromString(Output, staticData.GetOutputFactorOrder(), targetLHSStr, true);
UTIL_THROW_IF2(targetLHS->GetFactor(0) == NULL, "Null factor for target LHS");
// add to dictionary
TargetPhrase *targetPhrase = new TargetPhrase();
Word &targetWord = targetPhrase->AddWord();
targetWord.CreateUnknownWord(sourceWord);
// scores
float unknownScore = FloorScore(TransformScore(prob));
targetPhrase->GetScoreBreakdown().Assign(&unknownWordPenaltyProducer, unknownScore);
targetPhrase->Evaluate(*unksrc);
targetPhrase->SetTargetLHS(targetLHS);
targetPhrase->SetAlignmentInfo("0-0");
if (staticData.IsDetailedTreeFragmentsTranslationReportingEnabled()) {
targetPhrase->SetProperty("Tree","[ " + (*targetLHS)[0]->GetString().as_string() + " "+sourceWord[0]->GetString().as_string()+" ]");
}
// chart rule
to.AddPhraseOOV(*targetPhrase, m_cacheTargetPhraseCollection, range);
} // for (iterLHS
} else {
// drop source word. create blank trans opt
float unknownScore = FloorScore(-numeric_limits<float>::infinity());
TargetPhrase *targetPhrase = new TargetPhrase();
// loop
const UnknownLHSList &lhsList = staticData.GetUnknownLHS();
UnknownLHSList::const_iterator iterLHS;
for (iterLHS = lhsList.begin(); iterLHS != lhsList.end(); ++iterLHS) {
const string &targetLHSStr = iterLHS->first;
//float prob = iterLHS->second;
Word *targetLHS = new Word(true);
targetLHS->CreateFromString(Output, staticData.GetOutputFactorOrder(), targetLHSStr, true);
UTIL_THROW_IF2(targetLHS->GetFactor(0) == NULL, "Null factor for target LHS");
targetPhrase->GetScoreBreakdown().Assign(&unknownWordPenaltyProducer, unknownScore);
targetPhrase->Evaluate(*unksrc);
targetPhrase->SetTargetLHS(targetLHS);
// chart rule
to.AddPhraseOOV(*targetPhrase, m_cacheTargetPhraseCollection, range);
}
}
}
ChartTranslationOption::ChartTranslationOption(const TargetPhrase &targetPhrase)
:m_targetPhrase(targetPhrase)
,m_scoreBreakdown(targetPhrase.GetScoreBreakdown())
{
}
