mosesdecoder/moses/Hypothesis.cpp

468 lines
14 KiB
C++

// -*- mode: c++; indent-tabs-mode: nil; tab-width:2 -*-
// vim:tabstop=2
/***********************************************************************
Moses - factored phrase-based language decoder
Copyright (C) 2006 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 <iostream>
#include <limits>
#include <vector>
#include <algorithm>
#include "TranslationOption.h"
#include "TranslationOptionCollection.h"
#include "Hypothesis.h"
#include "Util.h"
#include "SquareMatrix.h"
#include "StaticData.h"
#include "InputType.h"
#include "Manager.h"
#include "IOWrapper.h"
#include "moses/FF/FFState.h"
#include "moses/FF/StatefulFeatureFunction.h"
#include "moses/FF/StatelessFeatureFunction.h"
#include <boost/foreach.hpp>
using namespace std;
namespace Moses
{
//size_t g_numHypos = 0;
Hypothesis::
Hypothesis(Manager& manager, InputType const& source, const TranslationOption &initialTransOpt, const Bitmap &bitmap, int id)
: m_prevHypo(NULL)
, m_sourceCompleted(bitmap)
, m_sourceInput(source)
, m_currSourceWordsRange(
m_sourceCompleted.GetFirstGapPos()>0 ? 0 : NOT_FOUND,
m_sourceCompleted.GetFirstGapPos()>0 ? m_sourceCompleted.GetFirstGapPos()-1 : NOT_FOUND)
, m_currTargetWordsRange(NOT_FOUND, NOT_FOUND)
, m_wordDeleted(false)
, m_futureScore(0.0f)
, m_estimatedScore(0.0f)
, m_ffStates(StatefulFeatureFunction::GetStatefulFeatureFunctions().size())
, m_arcList(NULL)
, m_transOpt(initialTransOpt)
, m_manager(manager)
, m_id(id)
{
// ++g_numHypos;
// used for initial seeding of trans process
// initialize scores
//_hash_computed = false;
//s_HypothesesCreated = 1;
const vector<const StatefulFeatureFunction*>& ffs = StatefulFeatureFunction::GetStatefulFeatureFunctions();
for (unsigned i = 0; i < ffs.size(); ++i)
m_ffStates[i] = ffs[i]->EmptyHypothesisState(source);
}
/***
* continue prevHypo by appending the phrases in transOpt
*/
Hypothesis::
Hypothesis(const Hypothesis &prevHypo, const TranslationOption &transOpt, const Bitmap &bitmap, int id)
: m_prevHypo(&prevHypo)
, m_sourceCompleted(bitmap)
, m_sourceInput(prevHypo.m_sourceInput)
, m_currSourceWordsRange(transOpt.GetSourceWordsRange())
, m_currTargetWordsRange(prevHypo.m_currTargetWordsRange.GetEndPos() + 1,
prevHypo.m_currTargetWordsRange.GetEndPos()
+ transOpt.GetTargetPhrase().GetSize())
, m_wordDeleted(false)
, m_futureScore(0.0f)
, m_estimatedScore(0.0f)
, m_ffStates(prevHypo.m_ffStates.size())
, m_arcList(NULL)
, m_transOpt(transOpt)
, m_manager(prevHypo.GetManager())
, m_id(id)
{
// ++g_numHypos;
m_currScoreBreakdown.PlusEquals(transOpt.GetScoreBreakdown());
m_wordDeleted = transOpt.IsDeletionOption();
}
Hypothesis::
~Hypothesis()
{
for (unsigned i = 0; i < m_ffStates.size(); ++i)
delete m_ffStates[i];
if (m_arcList) {
ArcList::iterator iter;
for (iter = m_arcList->begin() ; iter != m_arcList->end() ; ++iter) {
delete *iter;
}
m_arcList->clear();
delete m_arcList;
m_arcList = NULL;
}
}
void
Hypothesis::
AddArc(Hypothesis *loserHypo)
{
if (!m_arcList) {
if (loserHypo->m_arcList) { // we don't have an arcList, but loser does
this->m_arcList = loserHypo->m_arcList; // take ownership, we'll delete
loserHypo->m_arcList = 0; // prevent a double deletion
} else {
this->m_arcList = new ArcList();
}
} else {
if (loserHypo->m_arcList) { // both have an arc list: merge. delete loser
size_t my_size = m_arcList->size();
size_t add_size = loserHypo->m_arcList->size();
this->m_arcList->resize(my_size + add_size, 0);
std::memcpy(&(*m_arcList)[0] + my_size, &(*loserHypo->m_arcList)[0], add_size * sizeof(Hypothesis *));
delete loserHypo->m_arcList;
loserHypo->m_arcList = 0;
} else { // loserHypo doesn't have any arcs
// DO NOTHING
}
}
m_arcList->push_back(loserHypo);
}
/***
* calculate the logarithm of our total translation score (sum up components)
*/
void
Hypothesis::
EvaluateWhenApplied(float estimatedScore)
{
const StaticData &staticData = StaticData::Instance();
// some stateless score producers cache their values in the translation
// option: add these here
// language model scores for n-grams completely contained within a target
// phrase are also included here
// compute values of stateless feature functions that were not
// cached in the translation option
const vector<const StatelessFeatureFunction*>& sfs =
StatelessFeatureFunction::GetStatelessFeatureFunctions();
for (unsigned i = 0; i < sfs.size(); ++i) {
const StatelessFeatureFunction &ff = *sfs[i];
if(!staticData.IsFeatureFunctionIgnored(ff)) {
ff.EvaluateWhenApplied(*this, &m_currScoreBreakdown);
}
}
const vector<const StatefulFeatureFunction*>& ffs =
StatefulFeatureFunction::GetStatefulFeatureFunctions();
for (unsigned i = 0; i < ffs.size(); ++i) {
const StatefulFeatureFunction &ff = *ffs[i];
if(!staticData.IsFeatureFunctionIgnored(ff)) {
FFState const* s = m_prevHypo ? m_prevHypo->m_ffStates[i] : NULL;
m_ffStates[i] = ff.EvaluateWhenApplied(*this, s, &m_currScoreBreakdown);
}
}
// FUTURE COST
m_estimatedScore = estimatedScore;
// TOTAL
m_futureScore = m_currScoreBreakdown.GetWeightedScore() + m_estimatedScore;
if (m_prevHypo) m_futureScore += m_prevHypo->GetScore();
}
const Hypothesis* Hypothesis::GetPrevHypo()const
{
return m_prevHypo;
}
/**
* print hypothesis information for pharaoh-style logging
*/
void
Hypothesis::
PrintHypothesis() const
{
if (!m_prevHypo) {
TRACE_ERR(endl << "NULL hypo" << endl);
return;
}
TRACE_ERR(endl << "creating hypothesis "<< m_id <<" from "<< m_prevHypo->m_id<<" ( ");
int end = (int)(m_prevHypo->GetCurrTargetPhrase().GetSize()-1);
int start = end-1;
if ( start < 0 ) start = 0;
if ( m_prevHypo->m_currTargetWordsRange.GetStartPos() == NOT_FOUND ) {
TRACE_ERR( "<s> ");
} else {
TRACE_ERR( "... ");
}
if (end>=0) {
Range range(start, end);
TRACE_ERR( m_prevHypo->GetCurrTargetPhrase().GetSubString(range) << " ");
}
TRACE_ERR( ")"<<endl);
TRACE_ERR( "\tbase score "<< (m_prevHypo->m_futureScore - m_prevHypo->m_estimatedScore) <<endl);
TRACE_ERR( "\tcovering "<<m_currSourceWordsRange.GetStartPos()<<"-"<<m_currSourceWordsRange.GetEndPos()
<<": " << m_transOpt.GetInputPath().GetPhrase() << endl);
TRACE_ERR( "\ttranslated as: "<<(Phrase&) GetCurrTargetPhrase()<<endl); // <<" => translation cost "<<m_score[ScoreType::PhraseTrans];
if (m_wordDeleted) TRACE_ERR( "\tword deleted"<<endl);
// TRACE_ERR( "\tdistance: "<<GetCurrSourceWordsRange().CalcDistortion(m_prevHypo->GetCurrSourceWordsRange())); // << " => distortion cost "<<(m_score[ScoreType::Distortion]*weightDistortion)<<endl;
// TRACE_ERR( "\tlanguage model cost "); // <<m_score[ScoreType::LanguageModelScore]<<endl;
// TRACE_ERR( "\tword penalty "); // <<(m_score[ScoreType::WordPenalty]*weightWordPenalty)<<endl;
TRACE_ERR( "\tscore "<<m_futureScore - m_estimatedScore<<" + future cost "<<m_estimatedScore<<" = "<<m_futureScore<<endl);
TRACE_ERR( "\tunweighted feature scores: " << m_currScoreBreakdown << endl);
//PrintLMScores();
}
void
Hypothesis::
CleanupArcList(size_t nBestSize, bool distinctNBest)
{
// point this hypo's main hypo to itself
SetWinningHypo(this);
if (!m_arcList) return;
/* keep only number of arcs we need to create all n-best paths.
* However, may not be enough if only unique candidates are needed,
* so we'll keep all of arc list if nedd distinct n-best list
*/
if (!distinctNBest && m_arcList->size() > nBestSize * 5) {
// prune arc list only if there too many arcs
NTH_ELEMENT4(m_arcList->begin(), m_arcList->begin() + nBestSize - 1,
m_arcList->end(), CompareHypothesisTotalScore());
// delete bad ones
ArcList::iterator i = m_arcList->begin() + nBestSize;
while (i != m_arcList->end()) delete *i++;
m_arcList->erase(m_arcList->begin() + nBestSize, m_arcList->end());
}
// set all arc's main hypo variable to this hypo
ArcList::iterator iter = m_arcList->begin();
for (; iter != m_arcList->end() ; ++iter) {
Hypothesis *arc = *iter;
arc->SetWinningHypo(this);
}
}
TargetPhrase const&
Hypothesis::
GetCurrTargetPhrase() const
{
return m_transOpt.GetTargetPhrase();
}
void
Hypothesis::
GetOutputPhrase(Phrase &out) const
{
if (m_prevHypo != NULL)
m_prevHypo->GetOutputPhrase(out);
out.Append(GetCurrTargetPhrase());
}
TO_STRING_BODY(Hypothesis)
// friend
ostream& operator<<(ostream& out, const Hypothesis& hypo)
{
hypo.ToStream(out);
// words bitmap
out << "[" << hypo.m_sourceCompleted << "] ";
// scores
out << " [total=" << hypo.GetFutureScore() << "]";
out << " " << hypo.GetScoreBreakdown();
// alignment
out << " " << hypo.GetCurrTargetPhrase().GetAlignNonTerm();
return out;
}
std::string
Hypothesis::
GetSourcePhraseStringRep(const vector<FactorType> factorsToPrint) const
{
return m_transOpt.GetInputPath().GetPhrase().GetStringRep(factorsToPrint);
}
std::string
Hypothesis::
GetTargetPhraseStringRep(const vector<FactorType> factorsToPrint) const
{
return (m_prevHypo
? GetCurrTargetPhrase().GetStringRep(factorsToPrint)
: "");
}
std::string
Hypothesis::
GetSourcePhraseStringRep() const
{
vector<FactorType> allFactors(MAX_NUM_FACTORS);
for(size_t i=0; i < MAX_NUM_FACTORS; i++)
allFactors[i] = i;
return GetSourcePhraseStringRep(allFactors);
}
std::string
Hypothesis::
GetTargetPhraseStringRep() const
{
vector<FactorType> allFactors(MAX_NUM_FACTORS);
for(size_t i=0; i < MAX_NUM_FACTORS; i++)
allFactors[i] = i;
return GetTargetPhraseStringRep(allFactors);
}
size_t
Hypothesis::
OutputAlignment(std::ostream &out, bool recursive=true) const
{
WordAlignmentSort const& waso = m_manager.options()->output.WA_SortOrder;
TargetPhrase const& tp = GetCurrTargetPhrase();
// call with head recursion to output things in the right order
size_t trg_off = recursive && m_prevHypo ? m_prevHypo->OutputAlignment(out) : 0;
size_t src_off = GetCurrSourceWordsRange().GetStartPos();
typedef std::pair<size_t,size_t> const* entry;
std::vector<entry> alnvec = tp.GetAlignTerm().GetSortedAlignments(waso);
BOOST_FOREACH(entry e, alnvec)
out << e->first + src_off << "-" << e->second + trg_off << " ";
return trg_off + tp.GetSize();
}
void
Hypothesis::
OutputInput(std::vector<const Phrase*>& map, const Hypothesis* hypo)
{
if (!hypo->GetPrevHypo()) return;
OutputInput(map, hypo->GetPrevHypo());
map[hypo->GetCurrSourceWordsRange().GetStartPos()]
= &hypo->GetTranslationOption().GetInputPath().GetPhrase();
}
void
Hypothesis::
OutputInput(std::ostream& os) const
{
size_t len = this->GetInput().GetSize();
std::vector<const Phrase*> inp_phrases(len, 0);
OutputInput(inp_phrases, this);
for (size_t i=0; i<len; ++i)
if (inp_phrases[i]) os << *inp_phrases[i];
}
std::map<size_t, const Factor*>
Hypothesis::
GetPlaceholders(const Hypothesis &hypo, FactorType placeholderFactor) const
{
const InputPath &inputPath = hypo.GetTranslationOption().GetInputPath();
const Phrase &inputPhrase = inputPath.GetPhrase();
std::map<size_t, const Factor*> ret;
for (size_t sourcePos = 0; sourcePos < inputPhrase.GetSize(); ++sourcePos) {
const Factor *factor = inputPhrase.GetFactor(sourcePos, placeholderFactor);
if (factor) {
std::set<size_t> targetPos = hypo.GetTranslationOption().GetTargetPhrase().GetAlignTerm().GetAlignmentsForSource(sourcePos);
UTIL_THROW_IF2(targetPos.size() != 1,
"Placeholder should be aligned to 1, and only 1, word");
ret[*targetPos.begin()] = factor;
}
}
return ret;
}
size_t Hypothesis::hash() const
{
size_t seed;
// coverage NOTE from Hieu - we could make bitmap comparison here
// and in operator== compare the pointers since the bitmaps come
// from a factory. Same coverage is guaranteed to have the same
// bitmap. However, this make the decoding algorithm
// non-deterministic as the order of hypo extension can be
// different. This causes several regression tests to break. Since
// the speedup is minimal, I'm gonna leave it comparing the actual
// bitmaps
seed = m_sourceCompleted.hash();
// states
for (size_t i = 0; i < m_ffStates.size(); ++i) {
const FFState *state = m_ffStates[i];
if (state) {
size_t hash = state->hash();
boost::hash_combine(seed, hash);
}
}
return seed;
}
bool Hypothesis::operator==(const Hypothesis& other) const
{
// coverage
if (&m_sourceCompleted != &other.m_sourceCompleted) {
return false;
}
// states
for (size_t i = 0; i < m_ffStates.size(); ++i) {
const FFState *thisState = m_ffStates[i];
if (thisState) {
const FFState *otherState = other.m_ffStates[i];
assert(otherState);
if ((*thisState) != (*otherState)) {
return false;
}
}
}
return true;
}
bool
Hypothesis::
beats(Hypothesis const& b) const
{
if (m_futureScore != b.m_futureScore)
return m_futureScore > b.m_futureScore;
else if (m_estimatedScore != b.m_estimatedScore)
return m_estimatedScore > b.m_estimatedScore;
else if (m_prevHypo)
return b.m_prevHypo ? m_prevHypo->beats(*b.m_prevHypo) : true;
else return false;
// TO DO: add more tie breaking here
// results. We should compare other property of the hypos here.
// On the other hand, how likely is this going to happen?
}
}