moses-smt/mosesdecoder
1
1
Fork 0
mirror of https://github.com/moses-smt/mosesdecoder.git synced 2024-10-26 19:37:58 +03:00
Code Issues Packages Projects Releases Wiki Activity

Code reorganization and code cleanup.

Browse Source
This commit is contained in:
Ulrich Germann 2015-02-05 23:06:25 +00:00
parent 006297831c
commit db013975b0
1 changed files with 342 additions and 390 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

732
moses/TranslationOptionCollection.cpp
View File

@ -40,175 +40,146 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
#include "moses/FF/InputFeature.h"
#include "util/exception.hh"
#include <boost/foreach.hpp>
using namespace std;
namespace Moses
{
/** helper for pruning */
bool CompareTranslationOption(const TranslationOption *a, const TranslationOption *b)
{
return a->GetFutureScore() > b->GetFutureScore();
}
// bool CompareTranslationOption(const TranslationOption *a, const TranslationOption *b)
// {
// return a->GetFutureScore() > b->GetFutureScore();
// }
/** constructor; since translation options are indexed by coverage span, the corresponding data structure is initialized here
* This fn should be called by inherited classes
*/
TranslationOptionCollection::TranslationOptionCollection(
InputType const& src, size_t maxNoTransOptPerCoverage, float translationOptionThreshold)
/** constructor; since translation options are indexed by coverage span, the
* corresponding data structure is initialized here This fn should be
* called by inherited classe */
TranslationOptionCollection::
TranslationOptionCollection(InputType const& src, size_t maxNoTransOptPerCoverage,
float translationOptionThreshold)
: m_source(src)
,m_futureScore(src.GetSize())
,m_maxNoTransOptPerCoverage(maxNoTransOptPerCoverage)
,m_translationOptionThreshold(translationOptionThreshold)
, m_futureScore(src.GetSize())
, m_maxNoTransOptPerCoverage(maxNoTransOptPerCoverage)
, m_translationOptionThreshold(translationOptionThreshold)
{
// create 2-d vector
size_t size = src.GetSize();
for (size_t startPos = 0 ; startPos < size ; ++startPos) {
for (size_t sPos = 0 ; sPos < size ; ++sPos) {
m_collection.push_back( vector< TranslationOptionList >() );
size_t maxSize = size - startPos;
size_t maxSize = size - sPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t endPos = 0 ; endPos < maxSize ; ++endPos) {
m_collection[startPos].push_back( TranslationOptionList() );
for (size_t ePos = 0 ; ePos < maxSize ; ++ePos) {
m_collection[sPos].push_back( TranslationOptionList() );
}
}
}
/** destructor, clears out data structures */
TranslationOptionCollection::~TranslationOptionCollection()
TranslationOptionCollection::
~TranslationOptionCollection()
{
RemoveAllInColl(m_inputPathQueue);
}
void TranslationOptionCollection::Prune()
void
TranslationOptionCollection::
Prune()
{
// quit, if max size, threshold
if (m_maxNoTransOptPerCoverage == 0 && m_translationOptionThreshold == -std::numeric_limits<float>::infinity())
return;
static float no_th = -std::numeric_limits<float>::infinity();
if (m_maxNoTransOptPerCoverage == 0 && m_translationOptionThreshold == no_th)
return;
// bookkeeping for how many options used, pruned
size_t total = 0;
size_t totalPruned = 0;
// loop through all spans
size_t size = m_source.GetSize();
for (size_t startPos = 0 ; startPos < size; ++startPos) {
size_t maxSize = size - startPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t endPos = startPos ; endPos < startPos + maxSize ; ++endPos) {
// consider list for a span
TranslationOptionList &fullList = GetTranslationOptionList(startPos, endPos);
total += fullList.size();
// size pruning
if (m_maxNoTransOptPerCoverage > 0 &&
fullList.size() > m_maxNoTransOptPerCoverage) {
// sort in vector
NTH_ELEMENT4(fullList.begin(), fullList.begin() + m_maxNoTransOptPerCoverage, fullList.end(), CompareTranslationOption);
totalPruned += fullList.size() - m_maxNoTransOptPerCoverage;
// delete the rest
for (size_t i = m_maxNoTransOptPerCoverage ; i < fullList.size() ; ++i) {
delete fullList.Get(i);
for (size_t sPos = 0 ; sPos < size; ++sPos)
{
BOOST_FOREACH(TranslationOptionList& fullList, m_collection[sPos])
{
total += fullList.size();
totalPruned += fullList.SelectNBest(m_maxNoTransOptPerCoverage);
totalPruned += fullList.PruneByThreshold(m_translationOptionThreshold);
}
fullList.resize(m_maxNoTransOptPerCoverage);
}
// threshold pruning
if (fullList.size() > 1 && m_translationOptionThreshold != -std::numeric_limits<float>::infinity()) {
// first, find the best score
float bestScore = -std::numeric_limits<float>::infinity();
for (size_t i=0; i < fullList.size() ; ++i) {
if (fullList.Get(i)->GetFutureScore() > bestScore)
bestScore = fullList.Get(i)->GetFutureScore();
}
//std::cerr << "best score for span " << startPos << "-" << endPos << " is " << bestScore << "\n";
// then, remove items that are worse than best score + threshold
for (size_t i=0; i < fullList.size() ; ++i) {
if (fullList.Get(i)->GetFutureScore() < bestScore + m_translationOptionThreshold) {
//std::cerr << "\tremoving item " << i << ", score " << fullList.Get(i)->GetFutureScore() << ": " << fullList.Get(i)->GetTargetPhrase() << "\n";
delete fullList.Get(i);
fullList.Remove(i);
total--;
totalPruned++;
i--;
}
//else
//{
// std::cerr << "\tkeeping item " << i << ", score " << fullList.Get(i)->GetFutureScore() << ": " << fullList.Get(i)->GetTargetPhrase() << "\n";
//}
}
} // end of threshold pruning
}
} // end of loop through all spans
VERBOSE(2," Total translation options: " << total << std::endl
<< "Total translation options pruned: " << totalPruned << std::endl);
}
/** Force a creation of a translation option where there are none for a particular source position.
* ie. where a source word has not been translated, create a translation option by
* 1. not observing the table limits on phrase/generation tables
* 2. using the handler ProcessUnknownWord()
* Call this function once translation option collection has been filled with translation options
/** Force a creation of a translation option where there are none for a
* particular source position. ie. where a source word has not been
* translated, create a translation option by
* 1. not observing the table limits on phrase/generation tables
* 2. using the handler ProcessUnknownWord()
* Call this function once translation option collection has been filled with
* translation options
*
* This function calls for unknown words is complicated by the fact it must handle different input types.
* The call stack is
* Base::ProcessUnknownWord()
* Inherited::ProcessUnknownWord(position)
* Base::ProcessOneUnknownWord()
* This function calls for unknown words is complicated by the fact it must
* handle different input types. The call stack is
* Base::ProcessUnknownWord()
* Inherited::ProcessUnknownWord(position)
* Base::ProcessOneUnknownWord()
*
*/
void TranslationOptionCollection::ProcessUnknownWord()
void
TranslationOptionCollection::
ProcessUnknownWord()
{
const vector<DecodeGraph*>& decodeGraphList = StaticData::Instance().GetDecodeGraphs();
const vector<DecodeGraph*>& decodeGraphList
= StaticData::Instance().GetDecodeGraphs();
size_t size = m_source.GetSize();
// try to translation for coverage with no trans by expanding table limit
for (size_t graphInd = 0 ; graphInd < decodeGraphList.size() ; graphInd++) {
const DecodeGraph &decodeGraph = *decodeGraphList[graphInd];
for (size_t pos = 0 ; pos < size ; ++pos) {
TranslationOptionList &fullList = GetTranslationOptionList(pos, pos);
size_t numTransOpt = fullList.size();
if (numTransOpt == 0) {
TranslationOptionList* fullList = GetTranslationOptionList(pos, pos);
// size_t numTransOpt = fullList.size();
if (!fullList || fullList->size() == 0) {
CreateTranslationOptionsForRange(decodeGraph, pos, pos, false, graphInd);
}
}
}
bool alwaysCreateDirectTranslationOption = StaticData::Instance().IsAlwaysCreateDirectTranslationOption();
bool alwaysCreateDirectTranslationOption
= StaticData::Instance().IsAlwaysCreateDirectTranslationOption();
// create unknown words for 1 word coverage where we don't have any trans options
for (size_t pos = 0 ; pos < size ; ++pos) {
TranslationOptionList &fullList = GetTranslationOptionList(pos, pos);
if (fullList.size() == 0 || alwaysCreateDirectTranslationOption)
TranslationOptionList* fullList = GetTranslationOptionList(pos, pos);
if (!fullList || fullList->size() == 0 || alwaysCreateDirectTranslationOption)
ProcessUnknownWord(pos);
}
}
/** special handling of ONE unknown words. Either add temporarily add word to translation table,
* or drop the translation.
* This function should be called by the ProcessOneUnknownWord() in the inherited class
* At the moment, this unknown word handler is a bit of a hack, if copies over each factor from source
* to target word, or uses the 'UNK' factor.
* Ideally, this function should be in a class which can be expanded upon, for example,
* to create a morphologically aware handler.
*
* \param sourceWord the unknown word
* \param sourcePos
* \param length length covered by this word (may be > 1 for lattice input)
* \param inputScores a set of scores associated with unknown word (input scores from latties/CNs)
/** special handling of ONE unknown words. Either add temporarily add word to
* translation table, or drop the translation. This function should be
* called by the ProcessOneUnknownWord() in the inherited class At the
* moment, this unknown word handler is a bit of a hack, if copies over
* each factor from source to target word, or uses the 'UNK' factor.
* Ideally, this function should be in a class which can be expanded
* upon, for example, to create a morphologically aware handler.
*
* \param sourceWord the unknown word
* \param sourcePos
* \param length length covered by this word (may be > 1 for lattice input)
* \param inputScores a set of scores associated with unknown word (input scores from latties/CNs)
*/
void TranslationOptionCollection::ProcessOneUnknownWord(const InputPath &inputPath,
size_t sourcePos,
size_t length,
const ScorePair *inputScores)
void
TranslationOptionCollection::
ProcessOneUnknownWord(const InputPath &inputPath, size_t sourcePos,
size_t length, const ScorePair *inputScores)
{
const StaticData &staticData = StaticData::Instance();
const UnknownWordPenaltyProducer &unknownWordPenaltyProducer = UnknownWordPenaltyProducer::Instance();
const UnknownWordPenaltyProducer&
unknownWordPenaltyProducer = UnknownWordPenaltyProducer::Instance();
float unknownScore = FloorScore(TransformScore(0));
const Word &sourceWord = inputPath.GetPhrase().GetWord(0);
@ -257,14 +228,9 @@ void TranslationOptionCollection::ProcessOneUnknownWord(const InputPath &inputPa
//create a one-to-one alignment between UNKNOWN_FACTOR and its verbatim translation
targetPhrase.SetAlignmentInfo("0-0");
} else {
// drop source word. create blank trans opt
//targetPhrase.SetAlignment();
}
}
targetPhrase.GetScoreBreakdown().Assign(&unknownWordPenaltyProducer, unknownScore);
// source phrase
@ -285,35 +251,35 @@ void TranslationOptionCollection::ProcessOneUnknownWord(const InputPath &inputPa
* This matrix used in search.
* Call this function once translation option collection has been filled with translation options
*/
void TranslationOptionCollection::CalcFutureScore()
void
TranslationOptionCollection::
CalcFutureScore()
{
// setup the matrix (ignore lower triangle, set upper triangle to -inf
size_t size = m_source.GetSize(); // the width of the matrix
for(size_t row=0; row<size; row++) {
for(size_t row=0; row < size; row++) {
for(size_t col=row; col<size; col++) {
m_futureScore.SetScore(row, col, -numeric_limits<float>::infinity());
}
}
// walk all the translation options and record the cheapest option for each span
for (size_t startPos = 0 ; startPos < size ; ++startPos) {
size_t maxSize = m_source.GetSize() - startPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t endPos = startPos ; endPos < startPos + maxSize ; ++endPos) {
TranslationOptionList &transOptList = GetTranslationOptionList(startPos, endPos);
TranslationOptionList::const_iterator iterTransOpt;
for(iterTransOpt = transOptList.begin() ; iterTransOpt != transOptList.end() ; ++iterTransOpt) {
const TranslationOption &transOpt = **iterTransOpt;
float score = transOpt.GetFutureScore();
if (score > m_futureScore.GetScore(startPos, endPos))
m_futureScore.SetScore(startPos, endPos, score);
}
for (size_t sPos = 0 ; sPos < size ; ++sPos)
{
size_t ePos = sPos;
BOOST_FOREACH(TranslationOptionList& tol, m_collection[sPos])
{
TranslationOptionList::const_iterator toi;
for(toi = tol.begin() ; toi != tol.end() ; ++toi) {
const TranslationOption& to = **toi;
float score = to.GetFutureScore();
if (score > m_futureScore.GetScore(sPos, ePos))
m_futureScore.SetScore(sPos, ePos, score);
}
++ePos;
}
}
}
// now fill all the cells in the strictly upper triangle
// there is no way to modify the diagonal now, in the case
@ -324,42 +290,52 @@ void TranslationOptionCollection::CalcFutureScore()
for(size_t colstart = 1; colstart < size ; colstart++) {
for(size_t diagshift = 0; diagshift < size-colstart ; diagshift++) {
size_t startPos = diagshift;
size_t endPos = colstart+diagshift;
for(size_t joinAt = startPos; joinAt < endPos ; joinAt++) {
float joinedScore = m_futureScore.GetScore(startPos, joinAt)
+ m_futureScore.GetScore(joinAt+1, endPos);
/* // uncomment to see the cell filling scheme
TRACE_ERR( "[" <<startPos<<","<<endPos<<"] <-? ["<<startPos<<","<<joinAt<<"]+["<<joinAt+1<<","<<endPos
<< "] (colstart: "<<colstart<<", diagshift: "<<diagshift<<")"<<endl);
*/
if (joinedScore > m_futureScore.GetScore(startPos, endPos))
m_futureScore.SetScore(startPos, endPos, joinedScore);
size_t sPos = diagshift;
size_t ePos = colstart+diagshift;
for(size_t joinAt = sPos; joinAt < ePos ; joinAt++) {
float joinedScore = m_futureScore.GetScore(sPos, joinAt)
+ m_futureScore.GetScore(joinAt+1, ePos);
// uncomment to see the cell filling scheme
// TRACE_ERR("[" << sPos << "," << ePos << "] <-? ["
// << sPos << "," << joinAt << "]+["
// << joinAt+1 << "," << ePos << "] (colstart: "
// << colstart << ", diagshift: " << diagshift << ")"
// << endl);
if (joinedScore > m_futureScore.GetScore(sPos, ePos))
m_futureScore.SetScore(sPos, ePos, joinedScore);
}
}
}
IFVERBOSE(3)
{
int total = 0;
for(size_t row = 0; row < size; row++)
{
size_t col = row;
BOOST_FOREACH(TranslationOptionList& tol, m_collection[row])
{
// size_t maxSize = size - row;
// size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
// maxSize = std::min(maxSize, maxSizePhrase);
IFVERBOSE(3) {
int total = 0;
for(size_t row=0; row<size; row++) {
size_t maxSize = size - row;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for(size_t col=row; col<row+maxSize; col++) {
int count = GetTranslationOptionList(row, col).size();
TRACE_ERR( "translation options spanning from "
<< row <<" to "<< col <<" is "
<< count <<endl);
total += count;
}
// for(size_t col=row; col<row+maxSize; col++) {
int count = tol.size();
TRACE_ERR( "translation options spanning from "
<< row <<" to "<< col <<" is "
<< count <<endl);
total += count;
++col;
}
}
TRACE_ERR( "translation options generated in total: "<< total << endl);
for(size_t row=0; row<size; row++)
for(size_t col=row; col<size; col++)
TRACE_ERR( "future cost from "<< row <<" to "<< col <<" is "
<< m_futureScore.GetScore(row, col) <<endl);
}
TRACE_ERR( "translation options generated in total: "<< total << endl);
for(size_t row=0; row<size; row++)
for(size_t col=row; col<size; col++)
TRACE_ERR( "future cost from "<< row <<" to "<< col <<" is "<< m_futureScore.GetScore(row, col) <<endl);
}
}
@ -368,7 +344,9 @@ void TranslationOptionCollection::CalcFutureScore()
* for a particular input sentence. This implies applying all
* translation and generation steps. Also computes future cost matrix.
*/
void TranslationOptionCollection::CreateTranslationOptions()
void
TranslationOptionCollection::
CreateTranslationOptions()
{
// loop over all substrings of the source sentence, look them up
// in the phraseDictionary (which is the- possibly filtered-- phrase
@ -376,158 +354,142 @@ void TranslationOptionCollection::CreateTranslationOptions()
// for all phrases
// there may be multiple decoding graphs (factorizations of decoding)
const vector <DecodeGraph*> &decodeGraphList = StaticData::Instance().GetDecodeGraphs();
const vector <DecodeGraph*> &decodeGraphList
= StaticData::Instance().GetDecodeGraphs();
// length of the sentence
const size_t size = m_source.GetSize();
// loop over all decoding graphs, each generates translation options
for (size_t graphInd = 0 ; graphInd < decodeGraphList.size() ; graphInd++) {
if (decodeGraphList.size() > 1) {
VERBOSE(3,"Creating translation options from decoding graph " << graphInd << endl);
for (size_t gidx = 0 ; gidx < decodeGraphList.size() ; gidx++)
{
if (decodeGraphList.size() > 1)
VERBOSE(3,"Creating translation options from decoding graph " << gidx << endl);
const DecodeGraph& dg = *decodeGraphList[gidx];
size_t backoff = dg.GetBackoff();
// iterate over spans
for (size_t sPos = 0 ; sPos < size; sPos++)
{
size_t maxSize = size - sPos; // don't go over end of sentence
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t ePos = sPos ; ePos < sPos + maxSize ; ePos++)
{
if (gidx && backoff &&
(ePos-sPos+1 <= backoff || // size exceeds backoff limit (HUH? UG) or ...
m_collection[sPos][ePos-sPos].size() > 0))
{
VERBOSE(3,"No backoff to graph " << gidx << " for span [" << sPos << ";" << ePos << "]" << endl);
continue;
}
CreateTranslationOptionsForRange(dg, sPos, ePos, true, gidx);
}
}
}
const DecodeGraph &decodeGraph = *decodeGraphList[graphInd];
size_t backoff = decodeGraph.GetBackoff();
// generate phrases that start at startPos ...
for (size_t startPos = 0 ; startPos < size; startPos++) {
size_t maxSize = size - startPos; // don't go over end of sentence
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
// ... and that end at endPos
for (size_t endPos = startPos ; endPos < startPos + maxSize ; endPos++) {
if (graphInd > 0 && // only skip subsequent graphs
backoff != 0 && // use of backoff specified
(endPos-startPos+1 <= backoff || // size exceeds backoff limit or ...
m_collection[startPos][endPos-startPos].size() > 0)) { // no phrases found so far
VERBOSE(3,"No backoff to graph " << graphInd << " for span [" << startPos << ";" << endPos << "]" << endl);
// do not create more options
continue;
}
// create translation options for that range
CreateTranslationOptionsForRange( decodeGraph, startPos, endPos, true, graphInd);
}
}
}
VERBOSE(3,"Translation Option Collection\n " << *this << endl);
ProcessUnknownWord();
EvaluateWithSourceContext();
// Prune
Prune();
Sort();
// future score matrix
CalcFutureScore();
// Cached lex reodering costs
CacheLexReordering();
CalcFutureScore(); // future score matrix
CacheLexReordering(); // Cached lex reodering costs
}
void TranslationOptionCollection::CreateTranslationOptionsForRange(
const DecodeGraph &decodeGraph
, size_t startPos
, size_t endPos
, bool adhereTableLimit
, size_t graphInd
, InputPath &inputPath)
bool
TranslationOptionCollection::
CreateTranslationOptionsForRange
(const DecodeGraph& dg, size_t sPos, size_t ePos,
bool adhereTableLimit, size_t gidx, InputPath &inputPath)
{
if ((StaticData::Instance().GetXmlInputType() != XmlExclusive) || !HasXmlOptionsOverlappingRange(startPos,endPos)) {
if ((StaticData::Instance().GetXmlInputType() != XmlExclusive)
|| !HasXmlOptionsOverlappingRange(sPos,ePos))
{
// partial trans opt stored in here
PartialTranslOptColl* oldPtoc = new PartialTranslOptColl;
size_t totalEarlyPruned = 0;
// partial trans opt stored in here
PartialTranslOptColl* oldPtoc = new PartialTranslOptColl;
size_t totalEarlyPruned = 0;
// initial translation step
list <const DecodeStep* >::const_iterator iterStep = decodeGraph.begin();
const DecodeStep &decodeStep = **iterStep;
// initial translation step
list <const DecodeStep* >::const_iterator iterStep = dg.begin();
const DecodeStep &dstep = **iterStep;
const PhraseDictionary &phraseDictionary = *decodeStep.GetPhraseDictionaryFeature();
const TargetPhraseCollection *targetPhrases = inputPath.GetTargetPhrases(phraseDictionary);
const PhraseDictionary &pdict = *dstep.GetPhraseDictionaryFeature();
const TargetPhraseCollection *targetPhrases = inputPath.GetTargetPhrases(pdict);
static_cast<const DecodeStepTranslation&>(dstep).ProcessInitialTranslation
(m_source, *oldPtoc, sPos, ePos, adhereTableLimit, inputPath, targetPhrases);
static_cast<const DecodeStepTranslation&>(decodeStep).ProcessInitialTranslation
(m_source, *oldPtoc
, startPos, endPos, adhereTableLimit
, inputPath, targetPhrases);
SetInputScore(inputPath, *oldPtoc);
SetInputScore(inputPath, *oldPtoc);
// do rest of decode steps
int indexStep = 0;
// do rest of decode steps
int indexStep = 0;
for (++iterStep ; iterStep != dg.end() ; ++iterStep) {
const DecodeStep *dstep = *iterStep;
PartialTranslOptColl* newPtoc = new PartialTranslOptColl;
// go thru each intermediate trans opt just created
const vector<TranslationOption*>& partTransOptList = oldPtoc->GetList();
vector<TranslationOption*>::const_iterator iterPartialTranslOpt;
for (iterPartialTranslOpt = partTransOptList.begin() ; iterPartialTranslOpt != partTransOptList.end() ; ++iterPartialTranslOpt) {
TranslationOption &inputPartialTranslOpt = **iterPartialTranslOpt;
for (++iterStep ; iterStep != decodeGraph.end() ; ++iterStep) {
const DecodeStep *decodeStep = *iterStep;
PartialTranslOptColl* newPtoc = new PartialTranslOptColl;
// go thru each intermediate trans opt just created
const vector<TranslationOption*>& partTransOptList = oldPtoc->GetList();
vector<TranslationOption*>::const_iterator iterPartialTranslOpt;
for (iterPartialTranslOpt = partTransOptList.begin() ; iterPartialTranslOpt != partTransOptList.end() ; ++iterPartialTranslOpt) {
TranslationOption &inputPartialTranslOpt = **iterPartialTranslOpt;
if (const DecodeStepTranslation *translateStep = dynamic_cast<const DecodeStepTranslation*>(decodeStep) ) {
const PhraseDictionary &phraseDictionary = *translateStep->GetPhraseDictionaryFeature();
const TargetPhraseCollection *targetPhrases = inputPath.GetTargetPhrases(phraseDictionary);
translateStep->Process(inputPartialTranslOpt
, *decodeStep
, *newPtoc
, this
, adhereTableLimit
, targetPhrases);
} else {
const DecodeStepGeneration *genStep = dynamic_cast<const DecodeStepGeneration*>(decodeStep);
assert(genStep);
genStep->Process(inputPartialTranslOpt
, *decodeStep
, *newPtoc
, this
, adhereTableLimit);
}
}
// last but 1 partial trans not required anymore
totalEarlyPruned += newPtoc->GetPrunedCount();
delete oldPtoc;
oldPtoc = newPtoc;
indexStep++;
} // for (++iterStep
// add to fully formed translation option list
PartialTranslOptColl &lastPartialTranslOptColl = *oldPtoc;
const vector<TranslationOption*>& partTransOptList = lastPartialTranslOptColl.GetList();
vector<TranslationOption*>::const_iterator iterColl;
for (iterColl = partTransOptList.begin() ; iterColl != partTransOptList.end() ; ++iterColl) {
TranslationOption *transOpt = *iterColl;
if (StaticData::Instance().GetXmlInputType() != XmlConstraint || !ViolatesXmlOptionsConstraint(startPos,endPos,transOpt)) {
if (const DecodeStepTranslation *tstep = dynamic_cast<const DecodeStepTranslation*>(dstep) )
{
const PhraseDictionary &pdict = *tstep->GetPhraseDictionaryFeature();
const TargetPhraseCollection *targetPhrases = inputPath.GetTargetPhrases(pdict);
tstep->Process(inputPartialTranslOpt, *dstep, *newPtoc,
this, adhereTableLimit, targetPhrases);
} else {
const DecodeStepGeneration *genStep = dynamic_cast<const DecodeStepGeneration*>(dstep);
assert(genStep);
genStep->Process(inputPartialTranslOpt, *dstep, *newPtoc,
this, adhereTableLimit);
}
}
// last but 1 partial trans not required anymore
totalEarlyPruned += newPtoc->GetPrunedCount();
delete oldPtoc;
oldPtoc = newPtoc;
indexStep++;
} // for (++iterStep
// add to fully formed translation option list
PartialTranslOptColl &lastPartialTranslOptColl = *oldPtoc;
const vector<TranslationOption*>& partTransOptList = lastPartialTranslOptColl.GetList();
vector<TranslationOption*>::const_iterator iterColl;
for (iterColl = partTransOptList.begin() ; iterColl != partTransOptList.end() ; ++iterColl) {
TranslationOption *transOpt = *iterColl;
if (StaticData::Instance().GetXmlInputType() != XmlConstraint || !ViolatesXmlOptionsConstraint(sPos,ePos,transOpt)) {
Add(transOpt);
}
}
}
lastPartialTranslOptColl.DetachAll();
totalEarlyPruned += oldPtoc->GetPrunedCount();
delete oldPtoc;
// TRACE_ERR( "Early translation options pruned: " << totalEarlyPruned << endl);
} // if ((StaticData::Instance().GetXmlInputType() != XmlExclusive) || !HasXmlOptionsOverlappingRange(startPos,endPos))
if (graphInd == 0 && StaticData::Instance().GetXmlInputType() != XmlPassThrough && HasXmlOptionsOverlappingRange(startPos,endPos)) {
CreateXmlOptionsForRange(startPos, endPos);
} // if ((StaticData::Instance().GetXmlInputType() != XmlExclusive) || !HasXmlOptionsOverlappingRange(sPos,ePos))
if (gidx == 0 && StaticData::Instance().GetXmlInputType() != XmlPassThrough && HasXmlOptionsOverlappingRange(sPos,ePos)) {
CreateXmlOptionsForRange(sPos, ePos);
}
return true;
}
void TranslationOptionCollection::SetInputScore(const InputPath &inputPath, PartialTranslOptColl &oldPtoc)
void
TranslationOptionCollection::
SetInputScore(const InputPath &inputPath, PartialTranslOptColl &oldPtoc)
{
const ScorePair *inputScore = inputPath.GetInputScore();
if (inputScore == NULL) {
return;
}
const ScorePair* inputScore = inputPath.GetInputScore();
if (inputScore == NULL) return;
const InputFeature &inputFeature = InputFeature::Instance();
@ -541,79 +503,73 @@ void TranslationOptionCollection::SetInputScore(const InputPath &inputPath, Part
}
}
void TranslationOptionCollection::EvaluateWithSourceContext()
void
TranslationOptionCollection::
EvaluateWithSourceContext()
{
const size_t size = m_source.GetSize();
for (size_t startPos = 0 ; startPos < size ; ++startPos) {
size_t maxSize = m_source.GetSize() - startPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t endPos = startPos ; endPos < startPos + maxSize ; ++endPos) {
TranslationOptionList &transOptList = GetTranslationOptionList(startPos, endPos);
TranslationOptionList::const_iterator iterTransOpt;
for(iterTransOpt = transOptList.begin() ; iterTransOpt != transOptList.end() ; ++iterTransOpt) {
TranslationOption &transOpt = **iterTransOpt;
transOpt.EvaluateWithSourceContext(m_source);
}
for (size_t sPos = 0 ; sPos < size ; ++sPos)
{
BOOST_FOREACH(TranslationOptionList& tol, m_collection[sPos])
{
typedef TranslationOptionList::const_iterator to_iter;
for(to_iter i = tol.begin() ; i != tol.end() ; ++i)
(*i)->EvaluateWithSourceContext(m_source);
}
}
}
}
void TranslationOptionCollection::Sort()
void
TranslationOptionCollection::
Sort()
{
static TranslationOption::Better cmp;
size_t size = m_source.GetSize();
for (size_t startPos = 0 ; startPos < size; ++startPos) {
size_t maxSize = size - startPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t endPos = startPos ; endPos < startPos + maxSize; ++endPos) {
TranslationOptionList &transOptList = GetTranslationOptionList(startPos, endPos);
std::sort(transOptList.begin(), transOptList.end(), CompareTranslationOption);
for (size_t sPos = 0 ; sPos < size; ++sPos)
{
BOOST_FOREACH(TranslationOptionList& tol, m_collection.at(sPos))
std::sort(tol.begin(), tol.end(), cmp);
}
}
}
/** Check if this range overlaps with any XML options. This doesn't need to be an exact match, only an overlap.
* by default, we don't support XML options. subclasses need to override this function.
* called by CreateTranslationOptionsForRange()
* \param startPos first position in input sentence
* \param sPos first position in input sentence
* \param lastPos last position in input sentence
*/
bool TranslationOptionCollection::HasXmlOptionsOverlappingRange(size_t, size_t) const
{
return false;
//not implemented for base class
}
bool
TranslationOptionCollection::
HasXmlOptionsOverlappingRange(size_t, size_t) const
{ return false; }
/** Check if an option conflicts with any constraint XML options. Okay, if XML option is substring in source and target.
* by default, we don't support XML options. subclasses need to override this function.
* called by CreateTranslationOptionsForRange()
* \param startPos first position in input sentence
* \param sPos first position in input sentence
* \param lastPos last position in input sentence
*/
bool TranslationOptionCollection::ViolatesXmlOptionsConstraint(size_t, size_t, TranslationOption *) const
{
return false;
//not implemented for base class
}
bool
TranslationOptionCollection::
ViolatesXmlOptionsConstraint(size_t, size_t, TranslationOption*) const
{ return false; }
/** Populates the current Collection with XML options exactly covering the range specified. Default implementation does nothing.
* called by CreateTranslationOptionsForRange()
* \param startPos first position in input sentence
* \param sPos first position in input sentence
* \param lastPos last position in input sentence
*/
void TranslationOptionCollection::CreateXmlOptionsForRange(size_t, size_t)
{
//not implemented for base class
};
void
TranslationOptionCollection::
CreateXmlOptionsForRange(size_t, size_t)
{ }
/** Add translation option to the list
* \param translationOption translation option to be added */
void TranslationOptionCollection::Add(TranslationOption *translationOption)
void
TranslationOptionCollection::
Add(TranslationOption *translationOption)
{
const WordsRange &coverage = translationOption->GetSourceWordsRange();
@ -632,30 +588,27 @@ TO_STRING_BODY(TranslationOptionCollection);
std::ostream& operator<<(std::ostream& out, const TranslationOptionCollection& coll)
{
size_t size = coll.m_source.GetSize();
for (size_t startPos = 0 ; startPos < size ; ++startPos) {
size_t maxSize = size - startPos;
for (size_t sPos = 0 ; sPos < size ; ++sPos) {
size_t maxSize = size - sPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t endPos = startPos ; endPos < startPos + maxSize ; ++endPos) {
const TranslationOptionList& fullList = coll.GetTranslationOptionList(startPos, endPos);
size_t sizeFull = fullList.size();
for (size_t ePos = sPos ; ePos < sPos + maxSize ; ++ePos) {
const TranslationOptionList* fullList
= coll.GetTranslationOptionList(sPos, ePos);
if (!fullList) break;
size_t sizeFull = fullList->size();
for (size_t i = 0; i < sizeFull; i++) {
out << *fullList.Get(i) << std::endl;
out << *fullList->Get(i) << std::endl;
}
}
}
//std::vector< std::vector< TranslationOptionList > >::const_iterator i = coll.m_collection.begin();
//size_t j = 0;
//for (; i!=coll.m_collection.end(); ++i) {
//out << "s[" << j++ << "].size=" << i->size() << std::endl;
//}
return out;
}
void TranslationOptionCollection::CacheLexReordering()
void
TranslationOptionCollection::
CacheLexReordering()
{
size_t size = m_source.GetSize();
@ -665,68 +618,67 @@ void TranslationOptionCollection::CacheLexReordering()
const StatefulFeatureFunction &ff = **iter;
if (typeid(ff) == typeid(LexicalReordering)) {
const LexicalReordering &lexreordering = static_cast<const LexicalReordering&>(ff);
for (size_t startPos = 0 ; startPos < size ; startPos++) {
size_t maxSize = size - startPos;
for (size_t sPos = 0 ; sPos < size ; sPos++) {
size_t maxSize = size - sPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
for (size_t endPos = startPos ; endPos < startPos + maxSize; endPos++) {
TranslationOptionList &transOptList = GetTranslationOptionList( startPos, endPos);
TranslationOptionList::iterator iterTransOpt;
for(iterTransOpt = transOptList.begin() ; iterTransOpt != transOptList.end() ; ++iterTransOpt) {
TranslationOption &transOpt = **iterTransOpt;
//Phrase sourcePhrase = m_source.GetSubString(WordsRange(startPos,endPos));
const Phrase &sourcePhrase = transOpt.GetInputPath().GetPhrase();
Scores score = lexreordering.GetProb(sourcePhrase
, transOpt.GetTargetPhrase());
if (!score.empty())
transOpt.CacheLexReorderingScores(lexreordering, score);
} // for(iterTransOpt
} // for (size_t endPos = startPos ; endPos < startPos + maxSize; endPos++) {
} // for (size_t startPos = 0 ; startPos < size ; startPos++) {
BOOST_FOREACH(TranslationOptionList& tol, m_collection)
{
TranslationOptionList::iterator it;
for(it = tol.begin(); it != tol.end(); ++it)
{
TranslationOption &o = **it;
const Phrase &srcPhrase = o.GetInputPath().GetPhrase();
Scores score = lexreordering.GetProb(srcPhrase, o.GetTargetPhrase());
if (!score.empty())
transOpt.CacheLexReorderingScores(lexreordering, score);
} // for(iterTransOpt
} // for (size_t ePos = sPos ; ePos < sPos + maxSize; ePos++) {
} // for (size_t sPos = 0 ; sPos < size ; sPos++) {
} // if (typeid(ff) == typeid(LexicalReordering)) {
} // for (iter = ffs.begin(); iter != ffs.end(); ++iter) {
}
//! list of trans opt for a particular span
TranslationOptionList &TranslationOptionCollection::GetTranslationOptionList(size_t startPos, size_t endPos)
TranslationOptionList*
TranslationOptionCollection::
GetTranslationOptionList(size_t const sPos, size_t const ePos)
{
size_t maxSize = endPos - startPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
UTIL_THROW_IF2(maxSize >= m_collection[startPos].size(),
"Out of bound access: " << maxSize);
return m_collection[startPos][maxSize];
}
const TranslationOptionList &TranslationOptionCollection::GetTranslationOptionList(size_t startPos, size_t endPos) const
{
size_t maxSize = endPos - startPos;
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
maxSize = std::min(maxSize, maxSizePhrase);
UTIL_THROW_IF2(maxSize >= m_collection[startPos].size(),
"Out of bound access: " << maxSize);
return m_collection[startPos][maxSize];
UTIL_THROW_IF2(sPos >= m_collection.size(), "Out of bound access.");
vector<TranslationOptionList>& tol = m_collection[sPos];
size_t idx = ePos - sPos;
return idx < tol.size() ? &tol[idx] : NULL;
}
void TranslationOptionCollection::GetTargetPhraseCollectionBatch()
TranslationOptionList const*
TranslationOptionCollection::
GetTranslationOptionList(size_t sPos, size_t ePos) const
{
const vector <DecodeGraph*> &decodeGraphList = StaticData::Instance().GetDecodeGraphs();
for (size_t graphInd = 0 ; graphInd < decodeGraphList.size() ; graphInd++) {
const DecodeGraph &decodeGraph = *decodeGraphList[graphInd];
UTIL_THROW_IF2(sPos >= m_collection.size(), "Out of bound access.");
vector<TranslationOptionList>& tol = m_collection[sPos];
size_t idx = ePos - sPos;
return idx < tol.size() ? &tol[idx] : NULL;
}
list <const DecodeStep* >::const_iterator iterStep;
for (iterStep = decodeGraph.begin(); iterStep != decodeGraph.end() ; ++iterStep) {
const DecodeStep &decodeStep = **iterStep;
const DecodeStepTranslation *transStep = dynamic_cast<const DecodeStepTranslation *>(&decodeStep);
if (transStep) {
const PhraseDictionary &phraseDictionary = *transStep->GetPhraseDictionaryFeature();
phraseDictionary.GetTargetPhraseCollectionBatch(m_inputPathQueue);
}
void
TranslationOptionCollection::
GetTargetPhraseCollectionBatch()
{
const vector <DecodeGraph*> &dgl = StaticData::Instance().GetDecodeGraphs();
BOOST_FOREACH(DecodeGraph const& dg, dgl)
{
typedef list <const DecodeStep* >::const_iterator dsiter;
for (dsiter i = dg.begin(); i != dg.end() ; ++i)
{
const DecodeStepTranslation* tstep = dynamic_cast<const DecodeStepTranslation *>(*i);
if (tstep)
{
const PhraseDictionary &pdict = *tstep->GetPhraseDictionaryFeature();
pdict.GetTargetPhraseCollectionBatch(m_inputPathQueue);
}
}
}
}
}
} // namespace