2013-09-08 21:22:55 +04:00
|
|
|
#include <map>
|
2009-02-05 20:37:09 +03:00
|
|
|
#include "StaticData.h"
|
|
|
|
#include "WordLattice.h"
|
|
|
|
#include "PCNTools.h"
|
|
|
|
#include "Util.h"
|
|
|
|
#include "FloydWarshall.h"
|
2013-10-02 21:42:56 +04:00
|
|
|
#include "TranslationOptionCollectionLattice.h"
|
2013-10-02 23:02:05 +04:00
|
|
|
#include "TranslationOptionCollectionConfusionNet.h"
|
2013-06-05 03:50:24 +04:00
|
|
|
#include "moses/FF/InputFeature.h"
|
2009-02-05 20:37:09 +03:00
|
|
|
|
|
|
|
namespace Moses
|
|
|
|
{
|
2015-03-21 18:22:12 +03:00
|
|
|
WordLattice::WordLattice() : ConfusionNet()
|
2013-09-24 01:20:58 +04:00
|
|
|
{
|
2013-12-07 04:21:06 +04:00
|
|
|
UTIL_THROW_IF2(&InputFeature::Instance() == NULL,
|
2014-01-15 19:42:02 +04:00
|
|
|
"Input feature must be specified");
|
2013-09-24 01:20:58 +04:00
|
|
|
}
|
2009-02-05 20:37:09 +03:00
|
|
|
|
|
|
|
size_t WordLattice::GetColumnIncrement(size_t i, size_t j) const
|
|
|
|
{
|
2011-02-24 16:14:42 +03:00
|
|
|
return next_nodes[i][j];
|
2009-02-05 20:37:09 +03:00
|
|
|
}
|
|
|
|
|
2011-02-24 16:14:42 +03:00
|
|
|
void WordLattice::Print(std::ostream& out) const
|
|
|
|
{
|
|
|
|
out<<"word lattice: "<<data.size()<<"\n";
|
|
|
|
for(size_t i=0; i<data.size(); ++i) {
|
|
|
|
out<<i<<" -- ";
|
|
|
|
for(size_t j=0; j<data[i].size(); ++j) {
|
|
|
|
out<<"("<<data[i][j].first.ToString()<<", ";
|
2013-09-08 21:22:55 +04:00
|
|
|
|
|
|
|
// dense
|
|
|
|
std::vector<float>::const_iterator iterDense;
|
|
|
|
for(iterDense = data[i][j].second.denseScores.begin(); iterDense < data[i][j].second.denseScores.end(); ++iterDense) {
|
|
|
|
out<<", "<<*iterDense;
|
|
|
|
}
|
|
|
|
|
|
|
|
// sparse
|
|
|
|
std::map<StringPiece, float>::const_iterator iterSparse;
|
|
|
|
for(iterSparse = data[i][j].second.sparseScores.begin(); iterSparse != data[i][j].second.sparseScores.end(); ++iterSparse) {
|
|
|
|
out << ", " << iterSparse->first << "=" << iterSparse->second;
|
2011-02-24 16:14:42 +03:00
|
|
|
}
|
2013-09-08 21:22:55 +04:00
|
|
|
|
2011-02-24 16:14:42 +03:00
|
|
|
out << GetColumnIncrement(i,j) << ") ";
|
|
|
|
}
|
|
|
|
|
|
|
|
out<<"\n";
|
|
|
|
}
|
|
|
|
out<<"\n\n";
|
2009-02-05 20:37:09 +03:00
|
|
|
}
|
|
|
|
|
2014-05-19 17:34:27 +04:00
|
|
|
int
|
2014-04-08 15:44:21 +04:00
|
|
|
WordLattice::
|
|
|
|
InitializeFromPCNDataType
|
2014-05-19 17:34:27 +04:00
|
|
|
(const PCN::CN& cn,
|
|
|
|
const std::vector<FactorType>& factorOrder,
|
2014-04-08 15:44:21 +04:00
|
|
|
const std::string& debug_line)
|
2009-02-05 20:37:09 +03:00
|
|
|
{
|
2014-06-12 21:32:23 +04:00
|
|
|
// const StaticData &staticData = StaticData::Instance();
|
2013-12-07 04:21:06 +04:00
|
|
|
const InputFeature &inputFeature = InputFeature::Instance();
|
|
|
|
size_t numInputScores = inputFeature.GetNumInputScores();
|
|
|
|
size_t numRealWordCount = inputFeature.GetNumRealWordsInInput();
|
2013-06-05 03:50:24 +04:00
|
|
|
|
|
|
|
size_t maxSizePhrase = StaticData::Instance().GetMaxPhraseLength();
|
|
|
|
|
2012-12-06 00:21:33 +04:00
|
|
|
bool addRealWordCount = (numRealWordCount > 0);
|
|
|
|
|
2011-02-24 16:14:42 +03:00
|
|
|
//when we have one more weight than params, we add a word count feature
|
|
|
|
data.resize(cn.size());
|
|
|
|
next_nodes.resize(cn.size());
|
|
|
|
for(size_t i=0; i<cn.size(); ++i) {
|
|
|
|
const PCN::CNCol& col = cn[i];
|
|
|
|
if (col.empty()) return false;
|
|
|
|
data[i].resize(col.size());
|
|
|
|
next_nodes[i].resize(col.size());
|
|
|
|
for (size_t j=0; j<col.size(); ++j) {
|
|
|
|
const PCN::CNAlt& alt = col[j];
|
|
|
|
|
|
|
|
//check for correct number of link parameters
|
2013-09-06 01:34:22 +04:00
|
|
|
if (alt.m_denseFeatures.size() != numInputScores) {
|
2014-05-19 17:34:27 +04:00
|
|
|
TRACE_ERR("ERROR: need " << numInputScores
|
|
|
|
<< " link parameters, found "
|
|
|
|
<< alt.m_denseFeatures.size()
|
|
|
|
<< " while reading column " << i
|
|
|
|
<< " from " << debug_line << "\n");
|
2011-02-24 16:14:42 +03:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
//check each element for bounds
|
|
|
|
std::vector<float>::const_iterator probsIterator;
|
|
|
|
data[i][j].second = std::vector<float>(0);
|
2014-05-19 17:34:27 +04:00
|
|
|
for(probsIterator = alt.m_denseFeatures.begin();
|
|
|
|
probsIterator < alt.m_denseFeatures.end();
|
|
|
|
probsIterator++) {
|
2011-02-24 16:14:42 +03:00
|
|
|
IFVERBOSE(1) {
|
|
|
|
if (*probsIterator < 0.0f) {
|
|
|
|
TRACE_ERR("WARN: neg probability: " << *probsIterator << "\n");
|
|
|
|
//*probsIterator = 0.0f;
|
|
|
|
}
|
|
|
|
if (*probsIterator > 1.0f) {
|
|
|
|
TRACE_ERR("WARN: probability > 1: " << *probsIterator << "\n");
|
|
|
|
//*probsIterator = 1.0f;
|
|
|
|
}
|
|
|
|
}
|
2013-09-08 21:22:55 +04:00
|
|
|
|
|
|
|
float score = std::max(static_cast<float>(log(*probsIterator)), LOWEST_SCORE);
|
|
|
|
ScorePair &scorePair = data[i][j].second;
|
|
|
|
scorePair.denseScores.push_back(score);
|
2011-02-24 16:14:42 +03:00
|
|
|
}
|
|
|
|
//store 'real' word count in last feature if we have one more weight than we do arc scores and not epsilon
|
|
|
|
if (addRealWordCount) {
|
|
|
|
//only add count if not epsilon
|
2013-09-06 01:34:22 +04:00
|
|
|
float value = (alt.m_word=="" || alt.m_word==EPSILON) ? 0.0f : -1.0f;
|
2013-09-08 21:22:55 +04:00
|
|
|
data[i][j].second.denseScores.push_back(value);
|
2011-02-24 16:14:42 +03:00
|
|
|
}
|
2014-04-08 15:44:21 +04:00
|
|
|
Word& w = data[i][j].first;
|
|
|
|
w.CreateFromString(Input,factorOrder,StringPiece(alt.m_word),false);
|
|
|
|
// String2Word(alt.m_word, data[i][j]. first, factorOrder);
|
2013-09-06 01:34:22 +04:00
|
|
|
next_nodes[i][j] = alt.m_next;
|
2011-02-24 16:14:42 +03:00
|
|
|
|
|
|
|
if(next_nodes[i][j] > maxSizePhrase) {
|
|
|
|
TRACE_ERR("ERROR: Jump length " << next_nodes[i][j] << " in word lattice exceeds maximum phrase length " << maxSizePhrase << ".\n");
|
|
|
|
TRACE_ERR("ERROR: Increase max-phrase-length to process this lattice.\n");
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (!cn.empty()) {
|
|
|
|
std::vector<std::vector<bool> > edges(0);
|
|
|
|
this->GetAsEdgeMatrix(edges);
|
|
|
|
floyd_warshall(edges,distances);
|
|
|
|
|
|
|
|
IFVERBOSE(2) {
|
|
|
|
TRACE_ERR("Shortest paths:\n");
|
|
|
|
for (size_t i=0; i<edges.size(); ++i) {
|
|
|
|
for (size_t j=0; j<edges.size(); ++j) {
|
|
|
|
int d = distances[i][j];
|
|
|
|
if (d > 99999) {
|
|
|
|
d=-1;
|
|
|
|
}
|
|
|
|
TRACE_ERR("\t" << d);
|
|
|
|
}
|
|
|
|
TRACE_ERR("\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return !cn.empty();
|
2009-02-05 20:37:09 +03:00
|
|
|
}
|
|
|
|
|
2011-01-25 23:08:29 +03:00
|
|
|
int WordLattice::Read(std::istream& in,const std::vector<FactorType>& factorOrder)
|
|
|
|
{
|
2011-02-24 16:14:42 +03:00
|
|
|
Clear();
|
|
|
|
std::string line;
|
|
|
|
if(!getline(in,line)) return 0;
|
|
|
|
std::map<std::string, std::string> meta=ProcessAndStripSGML(line);
|
|
|
|
if (meta.find("id") != meta.end()) {
|
|
|
|
this->SetTranslationId(atol(meta["id"].c_str()));
|
|
|
|
}
|
|
|
|
|
|
|
|
PCN::CN cn = PCN::parsePCN(line);
|
|
|
|
return InitializeFromPCNDataType(cn, factorOrder, line);
|
2011-01-25 23:08:29 +03:00
|
|
|
}
|
|
|
|
|
2009-02-05 20:37:09 +03:00
|
|
|
void WordLattice::GetAsEdgeMatrix(std::vector<std::vector<bool> >& edges) const
|
|
|
|
{
|
|
|
|
edges.resize(data.size()+1,std::vector<bool>(data.size()+1, false));
|
2011-02-24 16:14:42 +03:00
|
|
|
for (size_t i=0; i<data.size(); ++i) {
|
|
|
|
for (size_t j=0; j<data[i].size(); ++j) {
|
2009-02-05 20:37:09 +03:00
|
|
|
edges[i][i+next_nodes[i][j]] = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
int WordLattice::ComputeDistortionDistance(const WordsRange& prev, const WordsRange& current) const
|
|
|
|
{
|
2011-02-24 16:14:42 +03:00
|
|
|
int result;
|
|
|
|
|
|
|
|
if (prev.GetStartPos() == NOT_FOUND && current.GetStartPos() == 0) {
|
|
|
|
result = 0;
|
|
|
|
|
|
|
|
VERBOSE(4, "Word lattice distortion: monotonic initial step\n");
|
|
|
|
} else if (prev.GetEndPos()+1 == current.GetStartPos()) {
|
|
|
|
result = 0;
|
|
|
|
|
|
|
|
VERBOSE(4, "Word lattice distortion: monotonic step from " << prev.GetEndPos() << " to " << current.GetStartPos() << "\n");
|
|
|
|
} else if (prev.GetStartPos() == NOT_FOUND) {
|
|
|
|
result = distances[0][current.GetStartPos()];
|
|
|
|
|
|
|
|
VERBOSE(4, "Word lattice distortion: initial step from 0 to " << current.GetStartPos() << " of length " << result << "\n");
|
|
|
|
if (result < 0 || result > 99999) {
|
|
|
|
TRACE_ERR("prev: " << prev << "\ncurrent: " << current << "\n");
|
|
|
|
TRACE_ERR("A: got a weird distance from 0 to " << (current.GetStartPos()+1) << " of " << result << "\n");
|
|
|
|
}
|
|
|
|
} else if (prev.GetEndPos() > current.GetStartPos()) {
|
|
|
|
result = distances[current.GetStartPos()][prev.GetEndPos() + 1];
|
|
|
|
|
|
|
|
VERBOSE(4, "Word lattice distortion: backward step from " << (prev.GetEndPos()+1) << " to " << current.GetStartPos() << " of length " << result << "\n");
|
|
|
|
if (result < 0 || result > 99999) {
|
|
|
|
TRACE_ERR("prev: " << prev << "\ncurrent: " << current << "\n");
|
|
|
|
TRACE_ERR("B: got a weird distance from "<< current.GetStartPos() << " to " << prev.GetEndPos()+1 << " of " << result << "\n");
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
result = distances[prev.GetEndPos() + 1][current.GetStartPos()];
|
|
|
|
|
|
|
|
VERBOSE(4, "Word lattice distortion: forward step from " << (prev.GetEndPos()+1) << " to " << current.GetStartPos() << " of length " << result << "\n");
|
|
|
|
if (result < 0 || result > 99999) {
|
|
|
|
TRACE_ERR("prev: " << prev << "\ncurrent: " << current << "\n");
|
|
|
|
TRACE_ERR("C: got a weird distance from "<< prev.GetEndPos()+1 << " to " << current.GetStartPos() << " of " << result << "\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return result;
|
2009-02-05 20:37:09 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
bool WordLattice::CanIGetFromAToB(size_t start, size_t end) const
|
|
|
|
{
|
2011-02-24 16:14:42 +03:00
|
|
|
// std::cerr << "CanIgetFromAToB(" << start << "," << end << ")=" << distances[start][end] << std::endl;
|
|
|
|
return distances[start][end] < 100000;
|
2009-02-05 20:37:09 +03:00
|
|
|
}
|
|
|
|
|
2013-10-02 21:42:56 +04:00
|
|
|
TranslationOptionCollection*
|
|
|
|
WordLattice::CreateTranslationOptionCollection() const
|
|
|
|
{
|
|
|
|
size_t maxNoTransOptPerCoverage = StaticData::Instance().GetMaxNoTransOptPerCoverage();
|
|
|
|
float translationOptionThreshold = StaticData::Instance().GetTranslationOptionThreshold();
|
2013-10-03 15:03:09 +04:00
|
|
|
|
|
|
|
TranslationOptionCollection *rv = NULL;
|
2013-10-04 16:08:14 +04:00
|
|
|
//rv = new TranslationOptionCollectionConfusionNet(*this, maxNoTransOptPerCoverage, translationOptionThreshold);
|
|
|
|
|
2013-10-03 15:03:09 +04:00
|
|
|
if (StaticData::Instance().GetUseLegacyPT()) {
|
|
|
|
rv = new TranslationOptionCollectionConfusionNet(*this, maxNoTransOptPerCoverage, translationOptionThreshold);
|
2014-01-15 19:42:02 +04:00
|
|
|
} else {
|
|
|
|
rv = new TranslationOptionCollectionLattice(*this, maxNoTransOptPerCoverage, translationOptionThreshold);
|
2013-10-03 15:03:09 +04:00
|
|
|
}
|
2013-10-04 16:08:14 +04:00
|
|
|
|
2013-11-19 22:52:15 +04:00
|
|
|
assert(rv);
|
2013-10-02 21:42:56 +04:00
|
|
|
return rv;
|
|
|
|
}
|
2009-02-05 20:37:09 +03:00
|
|
|
|
2013-10-02 23:02:05 +04:00
|
|
|
|
2013-10-02 20:43:59 +04:00
|
|
|
std::ostream& operator<<(std::ostream &out, const WordLattice &obj)
|
|
|
|
{
|
2013-10-03 14:33:48 +04:00
|
|
|
out << "next_nodes=";
|
|
|
|
for (size_t i = 0; i < obj.next_nodes.size(); ++i) {
|
|
|
|
out << i << ":";
|
|
|
|
|
|
|
|
const std::vector<size_t> &inner = obj.next_nodes[i];
|
|
|
|
for (size_t j = 0; j < inner.size(); ++j) {
|
|
|
|
out << inner[j] << " ";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
out << "distances=";
|
|
|
|
for (size_t i = 0; i < obj.distances.size(); ++i) {
|
|
|
|
out << i << ":";
|
|
|
|
|
|
|
|
const std::vector<int> &inner = obj.distances[i];
|
|
|
|
for (size_t j = 0; j < inner.size(); ++j) {
|
|
|
|
out << inner[j] << " ";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return out;
|
2009-02-05 20:37:09 +03:00
|
|
|
}
|
|
|
|
|
2013-10-02 20:43:59 +04:00
|
|
|
} // namespace
|
|
|
|
|