mirror of
https://github.com/moses-smt/mosesdecoder.git
synced 2024-12-28 14:32:38 +03:00
874 lines
28 KiB
C++
874 lines
28 KiB
C++
/***********************************************************************
|
|
Moses - factored phrase-based language decoder
|
|
Copyright (C) 2009 University of Edinburgh
|
|
|
|
This library is free software; you can redistribute it and/or
|
|
modify it under the terms of the GNU Lesser General Public
|
|
License as published by the Free Software Foundation; either
|
|
version 2.1 of the License, or (at your option) any later version.
|
|
|
|
This library is distributed in the hope that it will be useful,
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Lesser General Public License for more details.
|
|
|
|
You should have received a copy of the GNU Lesser General Public
|
|
License along with this library; if not, write to the Free Software
|
|
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
***********************************************************************/
|
|
|
|
#include <sstream>
|
|
#include <cstdio>
|
|
#include <iostream>
|
|
#include <fstream>
|
|
#include <vector>
|
|
#include <stdlib.h>
|
|
#include <assert.h>
|
|
#include <cstring>
|
|
#include <set>
|
|
|
|
#include "SafeGetline.h"
|
|
#include "tables-core.h"
|
|
#include "domain.h"
|
|
#include "PhraseAlignment.h"
|
|
#include "score.h"
|
|
#include "InputFileStream.h"
|
|
#include "OutputFileStream.h"
|
|
|
|
using namespace std;
|
|
using namespace MosesTraining;
|
|
|
|
#define LINE_MAX_LENGTH 100000
|
|
|
|
namespace MosesTraining
|
|
{
|
|
LexicalTable lexTable;
|
|
bool inverseFlag = false;
|
|
bool hierarchicalFlag = false;
|
|
bool pcfgFlag = false;
|
|
bool unpairedExtractFormatFlag = false;
|
|
bool conditionOnTargetLhsFlag = false;
|
|
bool wordAlignmentFlag = false;
|
|
bool goodTuringFlag = false;
|
|
bool kneserNeyFlag = false;
|
|
#define COC_MAX 10
|
|
bool logProbFlag = false;
|
|
int negLogProb = 1;
|
|
inline float maybeLogProb( float a ) { return logProbFlag ? negLogProb*log(a) : a; }
|
|
bool lexFlag = true;
|
|
bool unalignedFlag = false;
|
|
bool unalignedFWFlag = false;
|
|
bool outputNTLengths = false;
|
|
int countOfCounts[COC_MAX+1];
|
|
int totalDistinct = 0;
|
|
float minCountHierarchical = 0;
|
|
bool domainFlag = false;
|
|
bool domainRatioFlag = false;
|
|
bool domainSubsetFlag = false;
|
|
bool domainSparseFlag = false;
|
|
Domain *domain;
|
|
bool includeSentenceIdFlag = false;
|
|
|
|
Vocabulary vcbT;
|
|
Vocabulary vcbS;
|
|
|
|
} // namespace
|
|
|
|
vector<string> tokenize( const char [] );
|
|
|
|
void writeCountOfCounts( const string &fileNameCountOfCounts );
|
|
void processPhrasePairs( vector< PhraseAlignment > & , ostream &phraseTableFile);
|
|
PhraseAlignment* findBestAlignment(const PhraseAlignmentCollection &phrasePair );
|
|
void outputPhrasePair(const PhraseAlignmentCollection &phrasePair, float, int, ostream &phraseTableFile );
|
|
double computeLexicalTranslation( const PHRASE &, const PHRASE &, PhraseAlignment * );
|
|
double computeUnalignedPenalty( const PHRASE &, const PHRASE &, PhraseAlignment * );
|
|
set<string> functionWordList;
|
|
void loadFunctionWords( const string &fileNameFunctionWords );
|
|
double computeUnalignedFWPenalty( const PHRASE &, const PHRASE &, PhraseAlignment * );
|
|
void calcNTLengthProb(const vector< PhraseAlignment* > &phrasePairs
|
|
, map<size_t, map<size_t, float> > &sourceProb
|
|
, map<size_t, map<size_t, float> > &targetProb);
|
|
void printSourcePhrase(const PHRASE &, const PHRASE &, const PhraseAlignment &, ostream &);
|
|
void printTargetPhrase(const PHRASE &, const PHRASE &, const PhraseAlignment &, ostream &);
|
|
|
|
int main(int argc, char* argv[])
|
|
{
|
|
cerr << "Score v2.0 written by Philipp Koehn\n"
|
|
<< "scoring methods for extracted rules\n";
|
|
|
|
if (argc < 4) {
|
|
cerr << "syntax: score extract lex phrase-table [--Inverse] [--Hierarchical] [--LogProb] [--NegLogProb] [--NoLex] [--GoodTuring] [--KneserNey] [--WordAlignment] [--UnalignedPenalty] [--UnalignedFunctionWordPenalty function-word-file] [--MinCountHierarchical count] [--OutputNTLengths] [--PCFG] [--UnpairedExtractFormat] [--ConditionOnTargetLHS] [--[Sparse]Domain[Indicator|Ratio|Subset|Bin] domain-file [bins]]\n";
|
|
exit(1);
|
|
}
|
|
string fileNameExtract = argv[1];
|
|
string fileNameLex = argv[2];
|
|
string fileNamePhraseTable = argv[3];
|
|
string fileNameCountOfCounts;
|
|
<<<<<<< HEAD
|
|
char* fileNameFunctionWords = NULL;
|
|
char* fileNameDomain = NULL;
|
|
=======
|
|
string fileNameFunctionWords;
|
|
>>>>>>> b317522563feb4ca7ff978a0de661ec2189934ea
|
|
|
|
for(int i=4; i<argc; i++) {
|
|
if (strcmp(argv[i],"inverse") == 0 || strcmp(argv[i],"--Inverse") == 0) {
|
|
inverseFlag = true;
|
|
cerr << "using inverse mode\n";
|
|
} else if (strcmp(argv[i],"--Hierarchical") == 0) {
|
|
hierarchicalFlag = true;
|
|
cerr << "processing hierarchical rules\n";
|
|
} else if (strcmp(argv[i],"--PCFG") == 0) {
|
|
pcfgFlag = true;
|
|
cerr << "including PCFG scores\n";
|
|
} else if (strcmp(argv[i],"--UnpairedExtractFormat") == 0) {
|
|
unpairedExtractFormatFlag = true;
|
|
cerr << "processing unpaired extract format\n";
|
|
} else if (strcmp(argv[i],"--ConditionOnTargetLHS") == 0) {
|
|
conditionOnTargetLhsFlag = true;
|
|
cerr << "processing unpaired extract format\n";
|
|
} else if (strcmp(argv[i],"--WordAlignment") == 0) {
|
|
wordAlignmentFlag = true;
|
|
cerr << "outputing word alignment" << endl;
|
|
} else if (strcmp(argv[i],"--NoLex") == 0) {
|
|
lexFlag = false;
|
|
cerr << "not computing lexical translation score\n";
|
|
} else if (strcmp(argv[i],"--GoodTuring") == 0) {
|
|
goodTuringFlag = true;
|
|
fileNameCountOfCounts = string(fileNamePhraseTable) + ".coc";
|
|
cerr << "adjusting phrase translation probabilities with Good Turing discounting\n";
|
|
} else if (strcmp(argv[i],"--KneserNey") == 0) {
|
|
kneserNeyFlag = true;
|
|
fileNameCountOfCounts = string(fileNamePhraseTable) + ".coc";
|
|
cerr << "adjusting phrase translation probabilities with Kneser Ney discounting\n";
|
|
} else if (strcmp(argv[i],"--UnalignedPenalty") == 0) {
|
|
unalignedFlag = true;
|
|
cerr << "using unaligned word penalty\n";
|
|
} else if (strcmp(argv[i],"--UnalignedFunctionWordPenalty") == 0) {
|
|
unalignedFWFlag = true;
|
|
if (i+1==argc) {
|
|
cerr << "ERROR: specify count of count files for Kneser Ney discounting!\n";
|
|
exit(1);
|
|
}
|
|
fileNameFunctionWords = argv[++i];
|
|
cerr << "using unaligned function word penalty with function words from " << fileNameFunctionWords << endl;
|
|
} else if (strcmp(argv[i],"--SparseDomainIndicator") == 0 ||
|
|
strcmp(argv[i],"--SparseDomainRatio") == 0 ||
|
|
strcmp(argv[i],"--SparseDomainSubset") == 0 ||
|
|
strcmp(argv[i],"--DomainIndicator") == 0 ||
|
|
strcmp(argv[i],"--DomainRatio") == 0 ||
|
|
strcmp(argv[i],"--DomainSubset") == 0) {
|
|
includeSentenceIdFlag = true;
|
|
domainFlag = true;
|
|
domainSparseFlag = strstr( argv[i], "Sparse" );
|
|
domainRatioFlag = strstr( argv[i], "Ratio" );
|
|
domainSubsetFlag = strstr( argv[i], "Subset" );
|
|
if (i+1==argc) {
|
|
cerr << "ERROR: specify domain info file with " << argv[i] << endl;
|
|
exit(1);
|
|
}
|
|
fileNameDomain = argv[++i];
|
|
} else if (strcmp(argv[i],"--LogProb") == 0) {
|
|
logProbFlag = true;
|
|
cerr << "using log-probabilities\n";
|
|
} else if (strcmp(argv[i],"--NegLogProb") == 0) {
|
|
logProbFlag = true;
|
|
negLogProb = -1;
|
|
cerr << "using negative log-probabilities\n";
|
|
} else if (strcmp(argv[i],"--MinCountHierarchical") == 0) {
|
|
minCountHierarchical = atof(argv[++i]);
|
|
cerr << "dropping all phrase pairs occurring less than " << minCountHierarchical << " times\n";
|
|
minCountHierarchical -= 0.00001; // account for rounding
|
|
} else if (strcmp(argv[i],"--OutputNTLengths") == 0) {
|
|
outputNTLengths = true;
|
|
} else {
|
|
cerr << "ERROR: unknown option " << argv[i] << endl;
|
|
exit(1);
|
|
}
|
|
}
|
|
|
|
// lexical translation table
|
|
if (lexFlag)
|
|
lexTable.load( fileNameLex );
|
|
|
|
// function word list
|
|
if (unalignedFWFlag)
|
|
loadFunctionWords( fileNameFunctionWords );
|
|
|
|
// load domain information
|
|
if (domainFlag) {
|
|
if (inverseFlag) {
|
|
domainFlag = false;
|
|
includeSentenceIdFlag = false;
|
|
}
|
|
else {
|
|
domain = new Domain;
|
|
domain->load( fileNameDomain );
|
|
}
|
|
}
|
|
|
|
// compute count of counts for Good Turing discounting
|
|
if (goodTuringFlag || kneserNeyFlag) {
|
|
for(int i=1; i<=COC_MAX; i++) countOfCounts[i] = 0;
|
|
}
|
|
|
|
// sorted phrase extraction file
|
|
Moses::InputFileStream extractFile(fileNameExtract);
|
|
|
|
if (extractFile.fail()) {
|
|
cerr << "ERROR: could not open extract file " << fileNameExtract << endl;
|
|
exit(1);
|
|
}
|
|
istream &extractFileP = extractFile;
|
|
|
|
// output file: phrase translation table
|
|
ostream *phraseTableFile;
|
|
|
|
if (fileNamePhraseTable == "-") {
|
|
phraseTableFile = &cout;
|
|
}
|
|
else {
|
|
Moses::OutputFileStream *outputFile = new Moses::OutputFileStream();
|
|
bool success = outputFile->Open(fileNamePhraseTable);
|
|
if (!success) {
|
|
cerr << "ERROR: could not open file phrase table file "
|
|
<< fileNamePhraseTable << endl;
|
|
exit(1);
|
|
}
|
|
phraseTableFile = outputFile;
|
|
}
|
|
|
|
// loop through all extracted phrase translations
|
|
float lastCount = 0.0f;
|
|
float lastPcfgSum = 0.0f;
|
|
vector< PhraseAlignment > phrasePairsWithSameF;
|
|
int i=0;
|
|
char line[LINE_MAX_LENGTH],lastLine[LINE_MAX_LENGTH];
|
|
lastLine[0] = '\0';
|
|
PhraseAlignment *lastPhrasePair = NULL;
|
|
while(true) {
|
|
if (extractFileP.eof()) break;
|
|
if (++i % 100000 == 0) cerr << "." << flush;
|
|
SAFE_GETLINE((extractFileP), line, LINE_MAX_LENGTH, '\n', __FILE__);
|
|
if (extractFileP.eof()) break;
|
|
|
|
// identical to last line? just add count
|
|
if (strcmp(line,lastLine) == 0) {
|
|
lastPhrasePair->count += lastCount;
|
|
lastPhrasePair->pcfgSum += lastPcfgSum;
|
|
continue;
|
|
}
|
|
strcpy( lastLine, line );
|
|
|
|
// create new phrase pair
|
|
PhraseAlignment phrasePair;
|
|
phrasePair.create( line, i, includeSentenceIdFlag );
|
|
lastCount = phrasePair.count;
|
|
lastPcfgSum = phrasePair.pcfgSum;
|
|
|
|
// only differs in count? just add count
|
|
if (lastPhrasePair != NULL && lastPhrasePair->equals( phrasePair )) {
|
|
lastPhrasePair->count += phrasePair.count;
|
|
lastPhrasePair->pcfgSum += phrasePair.pcfgSum;
|
|
continue;
|
|
}
|
|
|
|
// if new source phrase, process last batch
|
|
if (lastPhrasePair != NULL &&
|
|
lastPhrasePair->GetSource() != phrasePair.GetSource()) {
|
|
processPhrasePairs( phrasePairsWithSameF, *phraseTableFile );
|
|
phrasePairsWithSameF.clear();
|
|
lastPhrasePair = NULL;
|
|
}
|
|
|
|
// add phrase pairs to list, it's now the last one
|
|
phrasePairsWithSameF.push_back( phrasePair );
|
|
lastPhrasePair = &phrasePairsWithSameF.back();
|
|
}
|
|
processPhrasePairs( phrasePairsWithSameF, *phraseTableFile );
|
|
|
|
phraseTableFile->flush();
|
|
if (phraseTableFile != &cout) {
|
|
delete phraseTableFile;
|
|
}
|
|
|
|
// output count of count statistics
|
|
if (goodTuringFlag || kneserNeyFlag) {
|
|
writeCountOfCounts( fileNameCountOfCounts );
|
|
}
|
|
}
|
|
|
|
void writeCountOfCounts( const string &fileNameCountOfCounts )
|
|
{
|
|
// open file
|
|
Moses::OutputFileStream countOfCountsFile;
|
|
bool success = countOfCountsFile.Open(fileNameCountOfCounts.c_str());
|
|
if (!success) {
|
|
cerr << "ERROR: could not open count-of-counts file "
|
|
<< fileNameCountOfCounts << endl;
|
|
return;
|
|
}
|
|
|
|
// Kneser-Ney needs the total number of phrase pairs
|
|
countOfCountsFile << totalDistinct << endl;
|
|
|
|
// write out counts
|
|
for(int i=1; i<=COC_MAX; i++) {
|
|
countOfCountsFile << countOfCounts[ i ] << endl;
|
|
}
|
|
countOfCountsFile.Close();
|
|
}
|
|
|
|
void processPhrasePairs( vector< PhraseAlignment > &phrasePair, ostream &phraseTableFile )
|
|
{
|
|
if (phrasePair.size() == 0) return;
|
|
|
|
// group phrase pairs based on alignments that matter
|
|
// (i.e. that re-arrange non-terminals)
|
|
PhrasePairGroup phrasePairGroup;
|
|
|
|
float totalSource = 0;
|
|
|
|
//cerr << "phrasePair.size() = " << phrasePair.size() << endl;
|
|
|
|
// loop through phrase pairs
|
|
for(size_t i=0; i<phrasePair.size(); i++) {
|
|
// add to total count
|
|
PhraseAlignment &currPhrasePair = phrasePair[i];
|
|
|
|
totalSource += phrasePair[i].count;
|
|
|
|
// check for matches
|
|
//cerr << "phrasePairGroup.size() = " << phrasePairGroup.size() << endl;
|
|
|
|
PhraseAlignmentCollection phraseAlignColl;
|
|
phraseAlignColl.push_back(&currPhrasePair);
|
|
pair<PhrasePairGroup::iterator, bool> retInsert;
|
|
retInsert = phrasePairGroup.insert(phraseAlignColl);
|
|
if (!retInsert.second)
|
|
{ // already exist. Add to that collection instead
|
|
PhraseAlignmentCollection &existingColl = const_cast<PhraseAlignmentCollection&>(*retInsert.first);
|
|
existingColl.push_back(&currPhrasePair);
|
|
}
|
|
|
|
}
|
|
|
|
// output the distinct phrase pairs, one at a time
|
|
const PhrasePairGroup::SortedColl &sortedColl = phrasePairGroup.GetSortedColl();
|
|
PhrasePairGroup::SortedColl::const_iterator iter;
|
|
|
|
for(iter = sortedColl.begin(); iter != sortedColl.end(); ++iter)
|
|
{
|
|
const PhraseAlignmentCollection &group = **iter;
|
|
outputPhrasePair( group, totalSource, phrasePairGroup.GetSize(), phraseTableFile );
|
|
}
|
|
|
|
}
|
|
|
|
PhraseAlignment* findBestAlignment(const PhraseAlignmentCollection &phrasePair )
|
|
{
|
|
float bestAlignmentCount = -1;
|
|
PhraseAlignment* bestAlignment = NULL;
|
|
|
|
for(size_t i=0; i<phrasePair.size(); i++) {
|
|
size_t alignInd;
|
|
if (inverseFlag)
|
|
{ // count backwards, so that alignments for ties will be the same for both normal & inverse scores
|
|
alignInd = phrasePair.size() - i - 1;
|
|
}
|
|
else {
|
|
alignInd = i;
|
|
}
|
|
|
|
if (phrasePair[alignInd]->count > bestAlignmentCount) {
|
|
bestAlignmentCount = phrasePair[alignInd]->count;
|
|
bestAlignment = phrasePair[alignInd];
|
|
}
|
|
}
|
|
|
|
return bestAlignment;
|
|
}
|
|
|
|
|
|
void calcNTLengthProb(const map<size_t, map<size_t, size_t> > &lengths
|
|
, size_t total
|
|
, map<size_t, map<size_t, float> > &probs)
|
|
{
|
|
map<size_t, map<size_t, size_t> >::const_iterator iterOuter;
|
|
for (iterOuter = lengths.begin(); iterOuter != lengths.end(); ++iterOuter)
|
|
{
|
|
size_t sourcePos = iterOuter->first;
|
|
const map<size_t, size_t> &inner = iterOuter->second;
|
|
|
|
map<size_t, size_t>::const_iterator iterInner;
|
|
for (iterInner = inner.begin(); iterInner != inner.end(); ++iterInner)
|
|
{
|
|
size_t length = iterInner->first;
|
|
size_t count = iterInner->second;
|
|
float prob = (float) count / (float) total;
|
|
probs[sourcePos][length] = prob;
|
|
}
|
|
}
|
|
}
|
|
|
|
void calcNTLengthProb(const vector< PhraseAlignment* > &phrasePairs
|
|
, map<size_t, map<size_t, float> > &sourceProb
|
|
, map<size_t, map<size_t, float> > &targetProb)
|
|
{
|
|
map<size_t, map<size_t, size_t> > sourceLengths, targetLengths;
|
|
// 1st = position in source phrase, 2nd = length, 3rd = count
|
|
map<size_t, size_t> totals;
|
|
// 1st = position in source phrase, 2nd = total counts
|
|
// each source pos should have same count?
|
|
|
|
vector< PhraseAlignment* >::const_iterator iterOuter;
|
|
for (iterOuter = phrasePairs.begin(); iterOuter != phrasePairs.end(); ++iterOuter)
|
|
{
|
|
const PhraseAlignment &phrasePair = **iterOuter;
|
|
const std::map<size_t, std::pair<size_t, size_t> > &ntLengths = phrasePair.GetNTLengths();
|
|
|
|
std::map<size_t, std::pair<size_t, size_t> >::const_iterator iterInner;
|
|
for (iterInner = ntLengths.begin(); iterInner != ntLengths.end(); ++iterInner)
|
|
{
|
|
size_t sourcePos = iterInner->first;
|
|
size_t sourceLength = iterInner->second.first;
|
|
size_t targetLength = iterInner->second.second;
|
|
|
|
sourceLengths[sourcePos][sourceLength]++;
|
|
targetLengths[sourcePos][targetLength]++;
|
|
|
|
totals[sourcePos]++;
|
|
}
|
|
}
|
|
|
|
if (totals.size() == 0)
|
|
{ // no non-term. Don't bother
|
|
return;
|
|
}
|
|
|
|
size_t total = totals.begin()->second;
|
|
if (totals.size() > 1)
|
|
{
|
|
assert(total == (++totals.begin())->second );
|
|
}
|
|
|
|
calcNTLengthProb(sourceLengths, total, sourceProb);
|
|
calcNTLengthProb(targetLengths, total, targetProb);
|
|
|
|
}
|
|
|
|
void outputNTLengthProbs(ostream &phraseTableFile, const map<size_t, map<size_t, float> > &probs, const string &prefix)
|
|
{
|
|
map<size_t, map<size_t, float> >::const_iterator iterOuter;
|
|
for (iterOuter = probs.begin(); iterOuter != probs.end(); ++iterOuter)
|
|
{
|
|
size_t sourcePos = iterOuter->first;
|
|
const map<size_t, float> &inner = iterOuter->second;
|
|
|
|
map<size_t, float>::const_iterator iterInner;
|
|
for (iterInner = inner.begin(); iterInner != inner.end(); ++iterInner)
|
|
{
|
|
size_t length = iterInner->first;
|
|
float prob = iterInner->second;
|
|
|
|
phraseTableFile << sourcePos << "|" << prefix << "|" << length << "=" << prob << " ";
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void outputPhrasePair(const PhraseAlignmentCollection &phrasePair, float totalCount, int distinctCount, ostream &phraseTableFile )
|
|
{
|
|
if (phrasePair.size() == 0) return;
|
|
|
|
PhraseAlignment *bestAlignment = findBestAlignment( phrasePair );
|
|
|
|
// compute count
|
|
float count = 0;
|
|
for(size_t i=0; i<phrasePair.size(); i++) {
|
|
count += phrasePair[i]->count;
|
|
}
|
|
|
|
// compute domain counts
|
|
map< string, float > domainCount;
|
|
if (domainFlag) {
|
|
for(size_t i=0; i<phrasePair.size(); i++) {
|
|
string d = domain->getDomainOfSentence( phrasePair[i]->sentenceId );
|
|
if (domainCount.find( d ) == domainCount.end())
|
|
domainCount[ d ] = phrasePair[i]->count;
|
|
else
|
|
domainCount[ d ] += phrasePair[i]->count;
|
|
}
|
|
}
|
|
|
|
// collect count of count statistics
|
|
if (goodTuringFlag || kneserNeyFlag) {
|
|
totalDistinct++;
|
|
int countInt = count + 0.99999;
|
|
if(countInt <= COC_MAX)
|
|
countOfCounts[ countInt ]++;
|
|
}
|
|
|
|
// compute PCFG score
|
|
float pcfgScore = 0;
|
|
if (pcfgFlag && !inverseFlag) {
|
|
float pcfgSum = 0;
|
|
for(size_t i=0; i<phrasePair.size(); ++i) {
|
|
pcfgSum += phrasePair[i]->pcfgSum;
|
|
}
|
|
pcfgScore = pcfgSum / count;
|
|
}
|
|
|
|
// output phrases
|
|
const PHRASE &phraseS = phrasePair[0]->GetSource();
|
|
const PHRASE &phraseT = phrasePair[0]->GetTarget();
|
|
|
|
// do not output if hierarchical and count below threshold
|
|
if (hierarchicalFlag && count < minCountHierarchical) {
|
|
for(size_t j=0; j<phraseS.size()-1; j++) {
|
|
if (isNonTerminal(vcbS.getWord( phraseS[j] )))
|
|
return;
|
|
}
|
|
}
|
|
|
|
// source phrase (unless inverse)
|
|
if (! inverseFlag) {
|
|
printSourcePhrase(phraseS, phraseT, *bestAlignment, phraseTableFile);
|
|
phraseTableFile << " ||| ";
|
|
}
|
|
|
|
// target phrase
|
|
printTargetPhrase(phraseS, phraseT, *bestAlignment, phraseTableFile);
|
|
phraseTableFile << " ||| ";
|
|
|
|
// source phrase (if inverse)
|
|
if (inverseFlag) {
|
|
printSourcePhrase(phraseS, phraseT, *bestAlignment, phraseTableFile);
|
|
phraseTableFile << " ||| ";
|
|
}
|
|
|
|
// lexical translation probability
|
|
if (lexFlag) {
|
|
double lexScore = computeLexicalTranslation( phraseS, phraseT, bestAlignment);
|
|
phraseTableFile << maybeLogProb( lexScore );
|
|
}
|
|
|
|
// unaligned word penalty
|
|
if (unalignedFlag) {
|
|
double penalty = computeUnalignedPenalty( phraseS, phraseT, bestAlignment);
|
|
phraseTableFile << " " << maybeLogProb( penalty );
|
|
}
|
|
|
|
// unaligned function word penalty
|
|
if (unalignedFWFlag) {
|
|
double penalty = computeUnalignedFWPenalty( phraseS, phraseT, bestAlignment);
|
|
phraseTableFile << " " << maybeLogProb( penalty );
|
|
}
|
|
|
|
// target-side PCFG score
|
|
if (pcfgFlag && !inverseFlag) {
|
|
phraseTableFile << " " << maybeLogProb( pcfgScore );
|
|
}
|
|
|
|
// domain count features
|
|
if (domainFlag) {
|
|
if (domainSparseFlag) {
|
|
// sparse, subset
|
|
if (domainSubsetFlag) {
|
|
typedef vector< string >::const_iterator I;
|
|
phraseTableFile << " doms";
|
|
for (I i = domain->list.begin(); i != domain->list.end(); i++ ) {
|
|
if (domainCount.find( *i ) != domainCount.end() ) {
|
|
phraseTableFile << "_" << *i;
|
|
}
|
|
}
|
|
phraseTableFile << " 1";
|
|
}
|
|
// sparse, indicator or ratio
|
|
else {
|
|
typedef map< string, float >::const_iterator I;
|
|
for (I i=domainCount.begin(); i != domainCount.end(); i++) {
|
|
if (domainRatioFlag) {
|
|
phraseTableFile << " domr_" << i->first << " " << (i->second / count);
|
|
}
|
|
else {
|
|
phraseTableFile << " dom_" << i->first << " 1";
|
|
}
|
|
}
|
|
}
|
|
}
|
|
// core, subset
|
|
else if (domainSubsetFlag) {
|
|
if (domain->list.size() > 6) {
|
|
cerr << "ERROR: too many domains for core domain subset features\n";
|
|
exit(1);
|
|
}
|
|
size_t bitmap = 0;
|
|
for(size_t bit = 0; bit < domain->list.size(); bit++) {
|
|
if (domainCount.find( domain->list[ bit ] ) != domainCount.end()) {
|
|
bitmap += 1 << bit;
|
|
}
|
|
}
|
|
for(size_t i = 1; i < (1 << domain->list.size()); i++) {
|
|
phraseTableFile << " " << maybeLogProb( (bitmap == i) ? 2.718 : 1 );
|
|
}
|
|
}
|
|
// core, indicator or ratio
|
|
else {
|
|
typedef vector< string >::const_iterator I;
|
|
for (I i = domain->list.begin(); i != domain->list.end(); i++ ) {
|
|
if (domainCount.find( *i ) == domainCount.end() ) {
|
|
phraseTableFile << " " << maybeLogProb( 1 );
|
|
}
|
|
else if (domainRatioFlag) {
|
|
phraseTableFile << " " << maybeLogProb( exp( domainCount[ *i ] / count ) );
|
|
}
|
|
else {
|
|
phraseTableFile << " " << maybeLogProb( 2.718 );
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
phraseTableFile << " ||| ";
|
|
|
|
// alignment info for non-terminals
|
|
if (! inverseFlag) {
|
|
if (hierarchicalFlag) {
|
|
// always output alignment if hiero style, but only for non-terms
|
|
assert(phraseT.size() == bestAlignment->alignedToT.size() + 1);
|
|
for(size_t j = 0; j < phraseT.size() - 1; j++) {
|
|
if (isNonTerminal(vcbT.getWord( phraseT[j] ))) {
|
|
if (bestAlignment->alignedToT[ j ].size() != 1) {
|
|
cerr << "Error: unequal numbers of non-terminals. Make sure the text does not contain words in square brackets (like [xxx])." << endl;
|
|
phraseTableFile.flush();
|
|
assert(bestAlignment->alignedToT[ j ].size() == 1);
|
|
}
|
|
int sourcePos = *(bestAlignment->alignedToT[ j ].begin());
|
|
phraseTableFile << sourcePos << "-" << j << " ";
|
|
}
|
|
}
|
|
} else if (wordAlignmentFlag) {
|
|
// alignment info in pb model
|
|
for(size_t j=0; j<bestAlignment->alignedToT.size(); j++) {
|
|
const set< size_t > &aligned = bestAlignment->alignedToT[j];
|
|
for (set< size_t >::const_iterator p(aligned.begin()); p != aligned.end(); ++p) {
|
|
phraseTableFile << *p << "-" << j << " ";
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
// counts
|
|
|
|
phraseTableFile << " ||| " << totalCount << " " << count;
|
|
if (kneserNeyFlag)
|
|
phraseTableFile << " " << distinctCount;
|
|
|
|
// nt lengths
|
|
if (outputNTLengths)
|
|
{
|
|
phraseTableFile << " ||| ";
|
|
|
|
if (!inverseFlag)
|
|
{
|
|
map<size_t, map<size_t, float> > sourceProb, targetProb;
|
|
// 1st sourcePos, 2nd = length, 3rd = prob
|
|
|
|
calcNTLengthProb(phrasePair, sourceProb, targetProb);
|
|
|
|
outputNTLengthProbs(phraseTableFile, sourceProb, "S");
|
|
outputNTLengthProbs(phraseTableFile, targetProb, "T");
|
|
}
|
|
}
|
|
|
|
phraseTableFile << endl;
|
|
}
|
|
|
|
double computeUnalignedPenalty( const PHRASE &phraseS, const PHRASE &phraseT, PhraseAlignment *alignment )
|
|
{
|
|
// unaligned word counter
|
|
double unaligned = 1.0;
|
|
// only checking target words - source words are caught when computing inverse
|
|
for(size_t ti=0; ti<alignment->alignedToT.size(); ti++) {
|
|
const set< size_t > & srcIndices = alignment->alignedToT[ ti ];
|
|
if (srcIndices.empty()) {
|
|
unaligned *= 2.718;
|
|
}
|
|
}
|
|
return unaligned;
|
|
}
|
|
|
|
double computeUnalignedFWPenalty( const PHRASE &phraseS, const PHRASE &phraseT, PhraseAlignment *alignment )
|
|
{
|
|
// unaligned word counter
|
|
double unaligned = 1.0;
|
|
// only checking target words - source words are caught when computing inverse
|
|
for(size_t ti=0; ti<alignment->alignedToT.size(); ti++) {
|
|
const set< size_t > & srcIndices = alignment->alignedToT[ ti ];
|
|
if (srcIndices.empty() && functionWordList.find( vcbT.getWord( phraseT[ ti ] ) ) != functionWordList.end()) {
|
|
unaligned *= 2.718;
|
|
}
|
|
}
|
|
return unaligned;
|
|
}
|
|
|
|
void loadFunctionWords( const string &fileName )
|
|
{
|
|
cerr << "Loading function word list from " << fileName;
|
|
ifstream inFile;
|
|
inFile.open(fileName.c_str());
|
|
if (inFile.fail()) {
|
|
cerr << " - ERROR: could not open file\n";
|
|
exit(1);
|
|
}
|
|
istream *inFileP = &inFile;
|
|
|
|
char line[LINE_MAX_LENGTH];
|
|
while(true) {
|
|
SAFE_GETLINE((*inFileP), line, LINE_MAX_LENGTH, '\n', __FILE__);
|
|
if (inFileP->eof()) break;
|
|
vector<string> token = tokenize( line );
|
|
if (token.size() > 0)
|
|
functionWordList.insert( token[0] );
|
|
}
|
|
inFile.close();
|
|
|
|
cerr << " - read " << functionWordList.size() << " function words\n";
|
|
inFile.close();
|
|
}
|
|
|
|
double computeLexicalTranslation( const PHRASE &phraseS, const PHRASE &phraseT, PhraseAlignment *alignment )
|
|
{
|
|
// lexical translation probability
|
|
double lexScore = 1.0;
|
|
int null = vcbS.getWordID("NULL");
|
|
// all target words have to be explained
|
|
for(size_t ti=0; ti<alignment->alignedToT.size(); ti++) {
|
|
const set< size_t > & srcIndices = alignment->alignedToT[ ti ];
|
|
if (srcIndices.empty()) {
|
|
// explain unaligned word by NULL
|
|
lexScore *= lexTable.permissiveLookup( null, phraseT[ ti ] );
|
|
} else {
|
|
// go through all the aligned words to compute average
|
|
double thisWordScore = 0;
|
|
for (set< size_t >::const_iterator p(srcIndices.begin()); p != srcIndices.end(); ++p) {
|
|
thisWordScore += lexTable.permissiveLookup( phraseS[ *p ], phraseT[ ti ] );
|
|
}
|
|
lexScore *= thisWordScore / (double)srcIndices.size();
|
|
}
|
|
}
|
|
return lexScore;
|
|
}
|
|
|
|
void LexicalTable::load( const string &fileName )
|
|
{
|
|
cerr << "Loading lexical translation table from " << fileName;
|
|
ifstream inFile;
|
|
inFile.open(fileName.c_str());
|
|
if (inFile.fail()) {
|
|
cerr << " - ERROR: could not open file\n";
|
|
exit(1);
|
|
}
|
|
istream *inFileP = &inFile;
|
|
|
|
char line[LINE_MAX_LENGTH];
|
|
|
|
int i=0;
|
|
while(true) {
|
|
i++;
|
|
if (i%100000 == 0) cerr << "." << flush;
|
|
SAFE_GETLINE((*inFileP), line, LINE_MAX_LENGTH, '\n', __FILE__);
|
|
if (inFileP->eof()) break;
|
|
|
|
vector<string> token = tokenize( line );
|
|
if (token.size() != 3) {
|
|
cerr << "line " << i << " in " << fileName
|
|
<< " has wrong number of tokens, skipping:\n"
|
|
<< token.size() << " " << token[0] << " " << line << endl;
|
|
continue;
|
|
}
|
|
|
|
double prob = atof( token[2].c_str() );
|
|
WORD_ID wordT = vcbT.storeIfNew( token[0] );
|
|
WORD_ID wordS = vcbS.storeIfNew( token[1] );
|
|
ltable[ wordS ][ wordT ] = prob;
|
|
}
|
|
cerr << endl;
|
|
}
|
|
|
|
void printSourcePhrase(const PHRASE &phraseS, const PHRASE &phraseT,
|
|
const PhraseAlignment &bestAlignment, ostream &out)
|
|
{
|
|
// output source symbols, except root, in rule table format
|
|
for (std::size_t i = 0; i < phraseS.size()-1; ++i) {
|
|
const std::string &word = vcbS.getWord(phraseS[i]);
|
|
if (!unpairedExtractFormatFlag || !isNonTerminal(word)) {
|
|
out << word << " ";
|
|
continue;
|
|
}
|
|
// get corresponding target non-terminal and output pair
|
|
std::set<std::size_t> alignmentPoints = bestAlignment.alignedToS[i];
|
|
assert(alignmentPoints.size() == 1);
|
|
int j = *(alignmentPoints.begin());
|
|
if (inverseFlag) {
|
|
out << vcbT.getWord(phraseT[j]) << word << " ";
|
|
} else {
|
|
out << word << vcbT.getWord(phraseT[j]) << " ";
|
|
}
|
|
}
|
|
// output source root symbol
|
|
if (conditionOnTargetLhsFlag && !inverseFlag) {
|
|
out << "[X]";
|
|
} else {
|
|
out << vcbS.getWord(phraseS.back());
|
|
}
|
|
}
|
|
|
|
void printTargetPhrase(const PHRASE &phraseS, const PHRASE &phraseT,
|
|
const PhraseAlignment &bestAlignment, ostream &out)
|
|
{
|
|
// output target symbols, except root, in rule table format
|
|
for (std::size_t i = 0; i < phraseT.size()-1; ++i) {
|
|
const std::string &word = vcbT.getWord(phraseT[i]);
|
|
if (!unpairedExtractFormatFlag || !isNonTerminal(word)) {
|
|
out << word << " ";
|
|
continue;
|
|
}
|
|
// get corresponding source non-terminal and output pair
|
|
std::set<std::size_t> alignmentPoints = bestAlignment.alignedToT[i];
|
|
assert(alignmentPoints.size() == 1);
|
|
int j = *(alignmentPoints.begin());
|
|
if (inverseFlag) {
|
|
out << word << vcbS.getWord(phraseS[j]) << " ";
|
|
} else {
|
|
out << vcbS.getWord(phraseS[j]) << word << " ";
|
|
}
|
|
}
|
|
// output target root symbol
|
|
if (conditionOnTargetLhsFlag) {
|
|
if (inverseFlag) {
|
|
out << "[X]";
|
|
} else {
|
|
out << vcbS.getWord(phraseS.back());
|
|
}
|
|
} else {
|
|
out << vcbT.getWord(phraseT.back());
|
|
}
|
|
}
|
|
|
|
std::pair<PhrasePairGroup::Coll::iterator,bool> PhrasePairGroup::insert ( const PhraseAlignmentCollection& obj )
|
|
{
|
|
std::pair<iterator,bool> ret = m_coll.insert(obj);
|
|
|
|
if (ret.second)
|
|
{ // obj inserted. Also add to sorted vector
|
|
const PhraseAlignmentCollection &insertedObj = *ret.first;
|
|
m_sortedColl.push_back(&insertedObj);
|
|
}
|
|
|
|
return ret;
|
|
}
|
|
|
|
|