moses-smt/mosesdecoder
1
1
Fork 0
mirror of https://github.com/moses-smt/mosesdecoder.git synced 2024-12-27 22:14:57 +03:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' of ssh://github.com/moses-smt/mosesdecoder

Browse Source
This commit is contained in:
Prashant Mathur 2016-06-15 05:33:42 -07:00
commit e31bc247ea
71 changed files with 3281 additions and 536 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
compile.sh
View File

@ -3,6 +3,6 @@
# you can install all 3rd-party dependencies by running make -f contrib/Makefiles/install-dependencies.gmake
set -e -o pipefail
OPT=${OPT:-$(pwd)/OPT}
OPT=${OPT:-$(pwd)/opt}
./bjam --with-irstlm=$OPT/irstlm-5.80.08 --with-boost=$OPT --with-cmph=$OPT --with-xmlrpc-c=$OPT --with-mm --with-probing-pt -j$(getconf _NPROCESSORS_ONLN) $@

20
contrib/other-builds/moses/.project
View File

@ -1105,6 +1105,16 @@
<type>1</type>
<locationURI>PARENT-3-PROJECT_LOC/moses/FF/ControlRecombination.h</locationURI>
</link>
<link>
<name>FF/CorrectionPattern.cpp</name>
<type>1</type>
<locationURI>PARENT-3-PROJECT_LOC/moses/FF/CorrectionPattern.cpp</locationURI>
</link>
<link>
<name>FF/CorrectionPattern.h</name>
<type>1</type>
<locationURI>PARENT-3-PROJECT_LOC/moses/FF/CorrectionPattern.h</locationURI>
</link>
<link>
<name>FF/CountNonTerms.cpp</name>
<type>1</type>
@ -1170,6 +1180,16 @@
<type>1</type>
<locationURI>PARENT-3-PROJECT_LOC/moses/FF/DynamicCacheBasedLanguageModel.h</locationURI>
</link>
<link>
<name>FF/EditOps.cpp</name>
<type>1</type>
<locationURI>PARENT-3-PROJECT_LOC/moses/FF/EditOps.cpp</locationURI>
</link>
<link>
<name>FF/EditOps.h</name>
<type>1</type>
<locationURI>PARENT-3-PROJECT_LOC/moses/FF/EditOps.h</locationURI>
</link>
<link>
<name>FF/FFState.cpp</name>
<type>1</type>

2
mert/Jamfile
View File

@ -31,6 +31,8 @@ Point.cpp
PerScorer.cpp
HwcmScorer.cpp
InternalTree.cpp
M2.cpp
M2Scorer.cpp
Scorer.cpp
ScorerFactory.cpp
Optimizer.cpp

61
mert/M2.cpp Normal file
View File

@ -0,0 +1,61 @@
#include <boost/algorithm/string.hpp>
#include "M2.h"
namespace MosesTuning
{
namespace M2
{
bool Annot::lowercase = true;
std::string Annot::transform(const std::string& e)
{
std::string temp = e;
if(lowercase) {
boost::erase_all(temp, " ");
return ToLower(temp);
} else
return e;
}
const std::string ToLower(const std::string& str)
{
std::string lc(str);
std::transform(lc.begin(), lc.end(), lc.begin(), (int(*)(int))std::tolower);
return lc;
}
Edit operator+(Edit& e1, Edit& e2)
{
std::string edit;
if(e1.edit.size() > 0 && e2.edit.size() > 0)
edit = e1.edit + " " + e2.edit;
else if(e1.edit.size() > 0)
edit = e1.edit;
else if(e2.edit.size() > 0)
edit = e2.edit;
return Edit(e1.cost + e2.cost, e1.changed + e2.changed, e1.unchanged + e2.unchanged, edit);
}
Edge operator+(Edge e1, Edge e2)
{
return Edge(e1.v, e2.u, e1.edit + e2.edit);
}
std::ostream& operator<<(std::ostream& o, Sentence s)
{
for(Sentence::iterator it = s.begin(); it != s.end(); it++)
o << *it << " ";
return o;
}
}
}

480
mert/M2.h Normal file
View File

@ -0,0 +1,480 @@
#pragma once
#include <cmath>
#include <string>
#include <vector>
#include <set>
#include <map>
#include <queue>
#include <iostream>
#include <fstream>
#include <iterator>
#include <algorithm>
#include <limits>
#include <sstream>
#include <boost/algorithm/string.hpp>
namespace MosesTuning
{
namespace M2
{
typedef std::vector<float> Stats;
typedef std::vector<std::string> Sentence;
std::ostream& operator<<(std::ostream& o, Sentence s);
const std::string ToLower(const std::string& str);
struct Annot {
size_t i;
size_t j;
std::string type;
std::string edit;
size_t annotator;
bool operator<(Annot a) const {
return i < a.i || (i == a.i && j < a.j)
|| (i == a.i && j == a.j && annotator < a.annotator)
|| (i == a.i && j == a.j && annotator == a.annotator && transform(edit) < transform(a.edit));
}
bool operator==(Annot a) const {
return (!(*this < a) && !(a < *this));
}
static std::string transform(const std::string& e);
static bool lowercase;
};
typedef std::set<Annot> Annots;
typedef std::set<size_t> Users;
struct Unit {
Sentence first;
Annots second;
Users third;
};
typedef std::vector<Unit> M2File;
struct Edit {
Edit(float c = 1.0, size_t ch = 0, size_t unch = 1, std::string e = "")
: cost(c), changed(ch), unchanged(unch), edit(e) {}
float cost;
size_t changed;
size_t unchanged;
std::string edit;
};
Edit operator+(Edit& e1, Edit& e2);
struct Vertex {
Vertex(size_t a = 0, size_t b = 0) : i(a), j(b) {}
bool operator<(const Vertex &v) const {
return i < v.i || (i == v.i && j < v.j);
}
bool operator==(const Vertex &v) const {
return i == v.i && j == v.j;
}
size_t i;
size_t j;
};
struct Edge {
Edge(Vertex vv = Vertex(), Vertex uu = Vertex(), Edit editt = Edit())
: v(vv), u(uu), edit(editt) {}
bool operator<(const Edge &e) const {
return v < e.v || (v == e.v && u < e.u);
}
Vertex v;
Vertex u;
Edit edit;
};
Edge operator+(Edge e1, Edge e2);
typedef std::vector<size_t> Row;
typedef std::vector<Row> Matrix;
struct Info {
Info(Vertex vv = Vertex(), Edit editt = Edit())
: v(vv), edit(editt) {}
bool operator<(const Info &i) const {
return v < i.v;
}
Vertex v;
Edit edit;
};
typedef std::set<Info> Track;
typedef std::vector<Track> TrackRow;
typedef std::vector<TrackRow> TrackMatrix;
typedef std::set<Vertex> Vertices;
typedef std::set<Edge> Edges;
class M2
{
private:
M2File m_m2;
size_t m_max_unchanged;
float m_beta;
bool m_lowercase;
bool m_verbose;
public:
M2() : m_max_unchanged(2), m_beta(0.5), m_lowercase(true), m_verbose(false) { }
M2(size_t max_unchanged, float beta, bool truecase, bool verbose = false)
: m_max_unchanged(max_unchanged), m_beta(beta), m_lowercase(!truecase), m_verbose(verbose) {
if(!m_lowercase) {
Annot::lowercase = false;
}
}
float Beta() {
return m_beta;
}
void ReadM2(const std::string& filename) {
std::ifstream m2file(filename.c_str());
std::string line;
Unit unit;
bool first = true;
while(std::getline(m2file, line)) {
if(line.size() > 2) {
if(line.substr(0, 2) == "S ") {
if(!first) {
if(unit.third.empty())
unit.third.insert(0);
m_m2.push_back(unit);
}
first = false;
unit.first = Sentence();
unit.second = Annots();
std::string sentenceLine = line.substr(2);
boost::split(unit.first, sentenceLine, boost::is_any_of(" "), boost::token_compress_on);
}
if(line.substr(0, 2) == "A ") {
std::string annotLine = line.substr(2);
std::vector<std::string> annot;
boost::iter_split(annot, annotLine, boost::algorithm::first_finder("|||"));
if(annot[1] != "noop") {
Annot a;
std::stringstream rangeStr(annot[0]);
rangeStr >> a.i >> a.j;
a.type = annot[1];
a.edit = annot[2];
std::stringstream annotStr(annot[5]);
annotStr >> a.annotator;
unit.third.insert(a.annotator);
unit.second.insert(a);
} else {
std::stringstream annotStr(annot[5]);
size_t annotator;
annotStr >> annotator;
unit.third.insert(annotator);
}
}
}
}
if(unit.third.empty())
unit.third.insert(0);
m_m2.push_back(unit);
}
size_t LevenshteinMatrix(const Sentence &s1, const Sentence &s2, Matrix &d, TrackMatrix &bt) {
size_t n = s1.size();
size_t m = s2.size();
if (n == 0)
return m;
if (m == 0)
return n;
d.resize(n + 1, Row(m + 1, 0));
bt.resize(n + 1, TrackRow(m + 1));
for(size_t i = 0; i <= n; ++i) {
d[i][0] = i;
if(i > 0)
bt[i][0].insert(Info(Vertex(i - 1, 0), Edit(1, 1, 0, "")));
}
for(size_t j = 0; j <= m; ++j) {
d[0][j] = j;
if(j > 0)
bt[0][j].insert(Info(Vertex(0, j - 1), Edit(1, 1, 0, s2[j - 1])));
}
int cost;
for(size_t i = 1; i <= n; ++i) {
for(size_t j = 1; j <= m; ++j) {
if(Annot::transform(s1[i-1]) == Annot::transform(s2[j-1]))
cost = 0;
else
cost = 2;
size_t left = d[i][j - 1] + 1;
size_t down = d[i - 1][j] + 1;
size_t diag = d[i - 1][j - 1] + cost;
d[i][j] = std::min(left, std::min(down, diag));
if(d[i][j] == left)
bt[i][j].insert(Info(Vertex(i, j - 1), Edit(1, 1, 0, s2[j - 1])));
if(d[i][j] == down)
bt[i][j].insert(Info(Vertex(i - 1, j), Edit(1, 1, 0, "")));
if(d[i][j] == diag)
bt[i][j].insert(Info(Vertex(i - 1, j - 1), cost ? Edit(1, 1, 0, s2[j - 1]) : Edit(1, 0, 1, s2[j - 1]) ));
}
}
return d[n][m];
}
void BuildGraph(const TrackMatrix &bt, Vertices &V, Edges &E) {
Vertex start(bt.size() - 1, bt[0].size() - 1);
std::queue<Vertex> Q;
Q.push(start);
while(!Q.empty()) {
Vertex v = Q.front();
Q.pop();
if(V.count(v) > 0)
continue;
V.insert(v);
for(Track::iterator it = bt[v.i][v.j].begin();
it != bt[v.i][v.j].end(); ++it) {
Edge e(it->v, v, it->edit);
E.insert(e);
if(V.count(e.v) == 0)
Q.push(e.v);
}
}
Edges newE;
do {
newE.clear();
for(Edges::iterator it1 = E.begin(); it1 != E.end(); ++it1) {
for(Edges::iterator it2 = E.begin(); it2 != E.end(); ++it2) {
if(it1->u == it2->v) {
Edge e = *it1 + *it2;
if(e.edit.changed > 0 &&
e.edit.unchanged <= m_max_unchanged &&
E.count(e) == 0)
newE.insert(e);
}
}
}
E.insert(newE.begin(), newE.end());
} while(newE.size() > 0);
}
void AddWeights(Edges &E, const Unit &u, size_t aid) {
for(Edges::iterator it1 = E.begin(); it1 != E.end(); ++it1) {
if(it1->edit.changed > 0) {
const_cast<float&>(it1->edit.cost) += 0.001;
for(Annots::iterator it2 = u.second.begin(); it2 != u.second.end(); ++it2) {
// if matches an annotator
if(it1->v.i == it2->i && it1->u.i == it2->j
&& Annot::transform(it1->edit.edit) == Annot::transform(it2->edit)
&& it2->annotator == aid) {
int newWeight = -(m_max_unchanged + 1) * E.size();
const_cast<float&>(it1->edit.cost) = newWeight;
}
}
}
}
}
void BellmanFord(Vertices &V, Edges &E) {
Vertex source(0, 0);
std::map<Vertex, float> distance;
std::map<Vertex, Vertex> predecessor;
for(Vertices::iterator it = V.begin(); it != V.end(); ++it) {
if(*it == source)
distance[*it] = 0;
else {
distance[*it] = std::numeric_limits<float>::infinity();
}
}
for(size_t i = 1; i < V.size(); ++i) {
for(Edges::iterator it = E.begin(); it != E.end(); ++it) {
if(distance[it->v] + it->edit.cost < distance[it->u]) {
distance[it->u] = distance[it->v] + it->edit.cost;
predecessor[it->u] = it->v;
}
}
}
Edges newE;
Vertex v = *V.rbegin();
while(true) {
//std::cout << predecessor[v] << " -> " << v << std::endl;
Edges::iterator it = E.find(Edge(predecessor[v], v));
if(it != E.end()) {
Edge f = *it;
//std::cout << f << std::endl;
newE.insert(f);
v = predecessor[v];
if(v == source)
break;
} else {
std::cout << "Error" << std::endl;
break;
}
}
E.clear();
E.insert(newE.begin(), newE.end());
}
void AddStats(const std::vector<Edges> &Es, const Unit &u, Stats &stats, size_t line) {
std::map<size_t, Stats> statsPerAnnotator;
for(std::set<size_t>::iterator it = u.third.begin();
it != u.third.end(); ++it) {
statsPerAnnotator[*it] = Stats(4, 0);
}
for(Annots::iterator it = u.second.begin(); it != u.second.end(); it++)
statsPerAnnotator[it->annotator][2]++;
for(std::set<size_t>::iterator ait = u.third.begin();
ait != u.third.end(); ++ait) {
for(Edges::iterator eit = Es[*ait].begin(); eit != Es[*ait].end(); ++eit) {
if(eit->edit.changed > 0) {
statsPerAnnotator[*ait][1]++;
Annot f;
f.i = eit->v.i;
f.j = eit->u.i;
f.annotator = *ait;
f.edit = eit->edit.edit;
for(Annots::iterator fit = u.second.begin(); fit != u.second.end(); fit++) {
if(f == *fit)
statsPerAnnotator[*ait][0]++;
}
}
}
}
size_t bestAnnot = 0;
float bestF = -1;
for(std::set<size_t>::iterator it = u.third.begin();
it != u.third.end(); ++it) {
Stats localStats = stats;
localStats[0] += statsPerAnnotator[*it][0];
localStats[1] += statsPerAnnotator[*it][1];
localStats[2] += statsPerAnnotator[*it][2];
if(m_verbose)
std::cerr << *it << " : " << localStats[0] << " " << localStats[1] << " " << localStats[2] << std::endl;
float f = FScore(localStats);
if(m_verbose)
std::cerr << f << std::endl;
if(f > bestF) {
bestF = f;
bestAnnot = *it;
}
}
if(m_verbose)
std::cerr << ">> Chosen Annotator for line " << line + 1 << " : " << bestAnnot << std::endl;
stats[0] += statsPerAnnotator[bestAnnot][0];
stats[1] += statsPerAnnotator[bestAnnot][1];
stats[2] += statsPerAnnotator[bestAnnot][2];
}
void SufStats(const std::string &sStr, size_t i, Stats &stats) {
std::string temp = sStr;
Sentence s;
boost::split(s, temp, boost::is_any_of(" "), boost::token_compress_on);
Unit &unit = m_m2[i];
Matrix d;
TrackMatrix bt;
size_t distance = LevenshteinMatrix(unit.first, s, d, bt);
std::vector<Vertices> Vs(unit.third.size());
std::vector<Edges> Es(unit.third.size());
if(distance > unit.first.size()) {
std::cerr << "Levenshtein distance is greater than source size." << std::endl;
stats[0] = 0;
stats[1] = distance;
stats[2] = 0;
stats[3] = unit.first.size();
return;
} else if(distance > 0) {
for(size_t j = 0; j < unit.third.size(); j++) {
BuildGraph(bt, Vs[j], Es[j]);
AddWeights(Es[j], unit, j);
BellmanFord(Vs[j], Es[j]);
}
}
AddStats(Es, unit, stats, i);
stats[3] = unit.first.size();
}
float FScore(const Stats& stats) {
float p = 1.0;
if(stats[1] != 0)
p = (float)stats[0] / (float)stats[1];
float r = 1.0;
if(stats[2] != 0)
r = (float)stats[0] / (float)stats[2];
float denom = (m_beta * m_beta * p + r);
float f = 0.0;
if(denom != 0)
f = ((1 + m_beta * m_beta) * p * r) / denom;
return f;
}
void FScore(const Stats& stats, float &p, float &r, float &f) {
p = 1.0;
if(stats[1] != 0)
p = (float)stats[0] / (float)stats[1];
r = 1.0;
if(stats[2] != 0)
r = (float)stats[0] / (float)stats[2];
float denom = (m_beta * m_beta * p + r);
f = 0.0;
if(denom != 0)
f = ((1 + m_beta * m_beta) * p * r) / denom;
}
};
}
}

137
mert/M2Scorer.cpp Normal file
View File

@ -0,0 +1,137 @@
#include "M2Scorer.h"
#include <algorithm>
#include <fstream>
#include <stdexcept>
#include <sstream>
#include <cstdlib>
#include <boost/lexical_cast.hpp>
using namespace std;
namespace MosesTuning
{
M2Scorer::M2Scorer(const string& config)
: StatisticsBasedScorer("M2Scorer", config),
beta_(Scan<float>(getConfig("beta", "0.5"))),
max_unchanged_words_(Scan<int>(getConfig("max_unchanged_words", "2"))),
truecase_(Scan<bool>(getConfig("truecase", "false"))),
verbose_(Scan<bool>(getConfig("verbose", "false"))),
m2_(max_unchanged_words_, beta_, truecase_)
{}
void M2Scorer::setReferenceFiles(const vector<string>& referenceFiles)
{
for(size_t i = 0; i < referenceFiles.size(); ++i) {
m2_.ReadM2(referenceFiles[i]);
break;
}
}
void M2Scorer::prepareStats(size_t sid, const string& text, ScoreStats& entry)
{
string sentence = trimStr(this->preprocessSentence(text));
std::vector<ScoreStatsType> stats(4, 0);
m2_.SufStats(sentence, sid, stats);
entry.set(stats);
}
float M2Scorer::calculateScore(const vector<ScoreStatsType>& comps) const
{
if (comps.size() != NumberOfScores()) {
throw runtime_error("Size of stat vector for M2Scorer is not " + NumberOfScores());
}
float beta = beta_;
float p = 0.0;
float r = 0.0;
float f = 0.0;
if(comps[1] != 0)
p = comps[0] / (double)comps[1];
else
p = 1.0;
if(comps[2] != 0)
r = comps[0] / (double)comps[2];
else
r = 1.0;
float denom = beta * beta * p + r;
if(denom != 0)
f = (1.0 + beta * beta) * p * r / denom;
else
f = 0.0;
if(verbose_)
std::cerr << comps[0] << " " << comps[1] << " " << comps[2] << std::endl;
if(verbose_)
std::cerr << p << " " << r << " " << f << std::endl;
return f;
}
float M2Scorer::getReferenceLength(const vector<ScoreStatsType>& comps) const
{
return comps[3];
}
std::vector<ScoreStatsType> randomStats(float decay, int max)
{
int gold = rand() % max;
int prop = rand() % max;
int corr = 0.0;
if(std::min(prop, gold) > 0)
corr = rand() % std::min(prop, gold);
//std::cerr << corr << " " << prop << " " << gold << std::endl;
std::vector<ScoreStatsType> stats(3, 0.0);
stats[0] = corr * decay;
stats[1] = prop * decay;
stats[2] = gold * decay;
return stats;
}
float sentenceM2(const std::vector<ScoreStatsType>& stats)
{
float beta = 0.5;
std::vector<ScoreStatsType> smoothStats(3, 0.0); // = randomStats(0.001, 5);
smoothStats[0] += stats[0];
smoothStats[1] += stats[1];
smoothStats[2] += stats[2];
float p = 0.0;
float r = 0.0;
float f = 0.0;
if(smoothStats[1] != 0)
p = smoothStats[0] / smoothStats[1];
else
p = 1.0;
if(smoothStats[2] != 0)
r = smoothStats[0] / smoothStats[2];
else
r = 1.0;
float denom = beta * beta * p + r;
if(denom != 0)
f = (1.0 + beta * beta) * p * r / denom;
else
f = 0.0;
return f;
}
}

52
mert/M2Scorer.h Normal file
View File

@ -0,0 +1,52 @@
#ifndef MERT_M2_SCORER_H_
#define MERT_M2_SCORER_H_
#include <string>
#include <vector>
#include <functional>
#include "Types.h"
#include "Util.h"
#include "StatisticsBasedScorer.h"
#include "M2.h"
namespace MosesTuning
{
/**
* M2Scorer class can compute CoNLL m2 F-score.
*/
class M2Scorer: public StatisticsBasedScorer
{
public:
explicit M2Scorer(const std::string& config);
virtual void setReferenceFiles(const std::vector<std::string>& referenceFiles);
virtual void prepareStats(std::size_t sid, const std::string& text, ScoreStats& entry);
virtual std::size_t NumberOfScores() const {
return 4;
}
virtual float calculateScore(const std::vector<ScoreStatsType>& comps) const;
virtual float getReferenceLength(const std::vector<ScoreStatsType>& comps) const;
private:
float beta_;
int max_unchanged_words_;
bool truecase_;
bool verbose_;
M2::M2 m2_;
std::map<std::pair<size_t, std::string>, std::vector<ScoreStatsType> > seen_;
// no copying allowed
M2Scorer(const M2Scorer&);
M2Scorer& operator=(const M2Scorer&);
};
float sentenceM2 (const std::vector<ScoreStatsType>& stats);
}
#endif // MERT_M2_SCORER_H_

4
mert/ScorerFactory.cpp
View File

@ -11,6 +11,7 @@
#include "SemposScorer.h"
#include "PermutationScorer.h"
#include "MeteorScorer.h"
#include "M2Scorer.h"
#include "HwcmScorer.h"
#include "Reference.h"
@ -34,6 +35,7 @@ vector<string> ScorerFactory::getTypes()
types.push_back(string("LRSCORE"));
types.push_back(string("METEOR"));
types.push_back(string("HWCM"));
types.push_back(string("M2SCORER"));
return types;
}
@ -54,6 +56,8 @@ Scorer* ScorerFactory::getScorer(const string& type, const string& config)
return new CderScorer(config, false);
} else if (type == "SEMPOS") {
return new SemposScorer(config);
} else if (type == "M2SCORER") {
return new M2Scorer(config);
} else if ((type == "HAMMING") || (type == "KENDALL")) {
return (PermutationScorer*) new PermutationScorer(type, config);
} else if (type == "METEOR") {

354
moses/FF/CorrectionPattern.cpp Normal file
View File

@ -0,0 +1,354 @@
#include <sstream>
#include "CorrectionPattern.h"
#include "moses/Phrase.h"
#include "moses/TargetPhrase.h"
#include "moses/InputPath.h"
#include "moses/Hypothesis.h"
#include "moses/ChartHypothesis.h"
#include "moses/ScoreComponentCollection.h"
#include "moses/TranslationOption.h"
#include "util/string_piece_hash.hh"
#include "util/exception.hh"
#include <functional>
#include <algorithm>
#include <boost/foreach.hpp>
#include <boost/algorithm/string.hpp>
#include "Diffs.h"
namespace Moses
{
using namespace std;
std::string MakePair(const std::string &s1, const std::string &s2, bool general)
{
std::vector<std::string> sourceList;
std::vector<std::string> targetList;
if(general) {
Diffs diffs = CreateDiff(s1, s2);
size_t i = 0, j = 0;
char lastType = 'm';
std::string source, target;
std::string match;
int count = 1;
BOOST_FOREACH(Diff type, diffs) {
if(type == 'm') {
if(lastType != 'm') {
sourceList.push_back(source);
targetList.push_back(target);
}
source.clear();
target.clear();
if(s1[i] == '+') {
if(match.size() >= 3) {
sourceList.push_back("(\\w{3,})·");
std::string temp = "1";
sprintf((char*)temp.c_str(), "%d", count);
targetList.push_back("\\" + temp + "·");
count++;
} else {
sourceList.push_back(match + "·");
targetList.push_back(match + "·");
}
match.clear();
} else
match.push_back(s1[i]);
i++;
j++;
} else if(type == 'd') {
if(s1[i] == '+')
source += "·";
else
source.push_back(s1[i]);
i++;
} else if(type == 'i') {
if(s2[j] == '+')
target += "·";
else
target.push_back(s2[j]);
j++;
}
if(type != 'm' && !match.empty()) {
if(match.size() >= 3) {
sourceList.push_back("(\\w{3,})");
std::string temp = "1";
sprintf((char*)temp.c_str(), "%d", count);
targetList.push_back("\\" + temp);
count++;
} else {
sourceList.push_back(match);
targetList.push_back(match);
}
match.clear();
}
lastType = type;
}
if(lastType != 'm') {
sourceList.push_back(source);
targetList.push_back(target);
}
if(!match.empty()) {
if(match.size() >= 3) {
sourceList.push_back("(\\w{3,})");
std::string temp = "1";
sprintf((char*)temp.c_str(), "%d", count);
targetList.push_back("\\"+ temp);
count++;
} else {
sourceList.push_back(match);
targetList.push_back(match);
}
}
match.clear();
} else {
std::string cs1 = s1;
std::string cs2 = s2;
boost::replace_all(cs1, "+", "·");
boost::replace_all(cs2, "+", "·");
sourceList.push_back(cs1);
targetList.push_back(cs2);
}
std::stringstream out;
out << "sub(«";
out << boost::join(sourceList, "");
out << "»,«";
out << boost::join(targetList, "");
out << "»)";
return out.str();
}
std::string CorrectionPattern::CreateSinglePattern(const Tokens &s1, const Tokens &s2) const
{
std::stringstream out;
if(s1.empty()) {
out << "ins(«" << boost::join(s2, "·") << "»)";
return out.str();
} else if(s2.empty()) {
out << "del(«" << boost::join(s1, "·") << "»)";
return out.str();
} else {
typename Tokens::value_type v1 = boost::join(s1, "+");
typename Tokens::value_type v2 = boost::join(s2, "+");
out << MakePair(v1, v2, m_general);
return out.str();
}
}
std::vector<std::string> GetContext(size_t pos,
size_t len,
size_t window,
const InputType &input,
const InputPath &inputPath,
const std::vector<FactorType>& factorTypes,
bool isRight)
{
const Sentence& sentence = static_cast<const Sentence&>(input);
const Range& range = inputPath.GetWordsRange();
int leftPos = range.GetStartPos() + pos - len - 1;
int rightPos = range.GetStartPos() + pos;
std::vector<std::string> contexts;
for(int length = 1; length <= (int)window; ++length) {
std::vector<std::string> current;
if(!isRight) {
for(int i = 0; i < length; i++) {
if(leftPos - i >= 0) {
current.push_back(sentence.GetWord(leftPos - i).GetString(factorTypes, false));
} else {
current.push_back("<s>");
}
}
if(current.back() == "<s>" && current.size() >= 2 && current[current.size()-2] == "<s>")
continue;
std::reverse(current.begin(), current.end());
contexts.push_back("left(«" + boost::join(current, "·") + "»)_");
}
if(isRight) {
for(int i = 0; i < length; i++) {
if(rightPos + i < (int)sentence.GetSize()) {
current.push_back(sentence.GetWord(rightPos + i).GetString(factorTypes, false));
} else {
current.push_back("</s>");
}
}
if(current.back() == "</s>" && current.size() >= 2 && current[current.size()-2] == "</s>")
continue;
contexts.push_back("_right(«" + boost::join(current, "·") + "»)");
}
}
return contexts;
}
std::vector<std::string>
CorrectionPattern::CreatePattern(const Tokens &s1,
const Tokens &s2,
const InputType &input,
const InputPath &inputPath) const
{
Diffs diffs = CreateDiff(s1, s2);
size_t i = 0, j = 0;
char lastType = 'm';
std::vector<std::string> patternList;
Tokens source, target;
BOOST_FOREACH(Diff type, diffs) {
if(type == 'm') {
if(lastType != 'm') {
std::string pattern = CreateSinglePattern(source, target);
patternList.push_back(pattern);
if(m_context > 0) {
std::vector<std::string> leftContexts = GetContext(i, source.size(), m_context, input, inputPath, m_contextFactors, false);
std::vector<std::string> rightContexts = GetContext(i, source.size(), m_context, input, inputPath, m_contextFactors, true);
BOOST_FOREACH(std::string left, leftContexts)
patternList.push_back(left + pattern);
BOOST_FOREACH(std::string right, rightContexts)
patternList.push_back(pattern + right);
BOOST_FOREACH(std::string left, leftContexts)
BOOST_FOREACH(std::string right, rightContexts)
patternList.push_back(left + pattern + right);
}
}
source.clear();
target.clear();
if(s1[i] != s2[j]) {
source.push_back(s1[i]);
target.push_back(s2[j]);
}
i++;
j++;
} else if(type == 'd') {
source.push_back(s1[i]);
i++;
} else if(type == 'i') {
target.push_back(s2[j]);
j++;
}
lastType = type;
}
if(lastType != 'm') {
std::string pattern = CreateSinglePattern(source, target);
patternList.push_back(pattern);
if(m_context > 0) {
std::vector<std::string> leftContexts = GetContext(i, source.size(), m_context, input, inputPath, m_contextFactors, false);
std::vector<std::string> rightContexts = GetContext(i, source.size(), m_context, input, inputPath, m_contextFactors, true);
BOOST_FOREACH(std::string left, leftContexts)
patternList.push_back(left + pattern);
BOOST_FOREACH(std::string right, rightContexts)
patternList.push_back(pattern + right);
BOOST_FOREACH(std::string left, leftContexts)
BOOST_FOREACH(std::string right, rightContexts)
patternList.push_back(left + pattern + right);
}
}
return patternList;
}
CorrectionPattern::CorrectionPattern(const std::string &line)
: StatelessFeatureFunction(0, line), m_factors(1, 0), m_general(false),
m_context(0), m_contextFactors(1, 0)
{
std::cerr << "Initializing correction pattern feature.." << std::endl;
ReadParameters();
}
void CorrectionPattern::SetParameter(const std::string& key, const std::string& value)
{
if (key == "factor") {
m_factors = std::vector<FactorType>(1, Scan<FactorType>(value));
} else if (key == "context-factor") {
m_contextFactors = std::vector<FactorType>(1, Scan<FactorType>(value));
} else if (key == "general") {
m_general = Scan<bool>(value);
} else if (key == "context") {
m_context = Scan<size_t>(value);
} else {
StatelessFeatureFunction::SetParameter(key, value);
}
}
void CorrectionPattern::EvaluateWithSourceContext(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, const StackVec *stackVec
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection *estimatedFutureScore) const
{
ComputeFeatures(input, inputPath, targetPhrase, &scoreBreakdown);
}
void CorrectionPattern::ComputeFeatures(
const InputType &input,
const InputPath &inputPath,
const TargetPhrase& target,
ScoreComponentCollection* accumulator) const
{
const Phrase &source = inputPath.GetPhrase();
std::vector<std::string> sourceTokens;
for(size_t i = 0; i < source.GetSize(); ++i)
sourceTokens.push_back(source.GetWord(i).GetString(m_factors, false));
std::vector<std::string> targetTokens;
for(size_t i = 0; i < target.GetSize(); ++i)
targetTokens.push_back(target.GetWord(i).GetString(m_factors, false));
std::vector<std::string> patternList = CreatePattern(sourceTokens, targetTokens, input, inputPath);
for(size_t i = 0; i < patternList.size(); ++i)
accumulator->PlusEquals(this, patternList[i], 1);
/*
BOOST_FOREACH(std::string w, sourceTokens)
std::cerr << w << " ";
std::cerr << std::endl;
BOOST_FOREACH(std::string w, targetTokens)
std::cerr << w << " ";
std::cerr << std::endl;
BOOST_FOREACH(std::string w, patternList)
std::cerr << w << " ";
std::cerr << std::endl << std::endl;
*/
}
bool CorrectionPattern::IsUseable(const FactorMask &mask) const
{
bool ret = true;
for(size_t i = 0; i < m_factors.size(); ++i)
ret = ret && mask[m_factors[i]];
for(size_t i = 0; i < m_contextFactors.size(); ++i)
ret = ret && mask[m_contextFactors[i]];
return ret;
}
}

73
moses/FF/CorrectionPattern.h Normal file
View File

@ -0,0 +1,73 @@
#ifndef moses_CorrectionPattern_h
#define moses_CorrectionPattern_h
#include <string>
#include <boost/unordered_set.hpp>
#include "StatelessFeatureFunction.h"
#include "moses/FactorCollection.h"
#include "moses/AlignmentInfo.h"
namespace Moses
{
typedef std::vector<std::string> Tokens;
/** Sets the features for length of source phrase, target phrase, both.
*/
class CorrectionPattern : public StatelessFeatureFunction
{
private:
std::vector<FactorType> m_factors;
bool m_general;
size_t m_context;
std::vector<FactorType> m_contextFactors;
public:
CorrectionPattern(const std::string &line);
bool IsUseable(const FactorMask &mask) const;
void EvaluateInIsolation(const Phrase &source
, const TargetPhrase &targetPhrase
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection &estimatedFutureScore) const
{}
virtual void EvaluateWithSourceContext(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, const StackVec *stackVec
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection *estimatedFutureScore = NULL) const;
void EvaluateTranslationOptionListWithSourceContext(const InputType &input
, const TranslationOptionList &translationOptionList) const
{}
void EvaluateWhenApplied(const Hypothesis& hypo,
ScoreComponentCollection* accumulator) const
{}
void EvaluateWhenApplied(const ChartHypothesis &hypo,
ScoreComponentCollection* accumulator) const
{}
void ComputeFeatures(const InputType &input,
const InputPath &inputPath,
const TargetPhrase& targetPhrase,
ScoreComponentCollection* accumulator) const;
void SetParameter(const std::string& key, const std::string& value);
std::vector<std::string> CreatePattern(const Tokens &s1,
const Tokens &s2,
const InputType &input,
const InputPath &inputPath) const;
std::string CreateSinglePattern(const Tokens &s1, const Tokens &s2) const;
};
}
#endif // moses_CorrectionPattern_h

150
moses/FF/Diffs.h Normal file
View File

@ -0,0 +1,150 @@
#ifndef moses_Diffs_h
#define moses_Diffs_h
#include <cmath>
namespace Moses
{
typedef char Diff;
typedef std::vector<Diff> Diffs;
template <class Sequence, class Pred>
void CreateDiffRec(size_t** c,
const Sequence &s1,
const Sequence &s2,
size_t start,
size_t i,
size_t j,
Diffs& diffs,
Pred pred)
{
if(i > 0 && j > 0 && pred(s1[i - 1 + start], s2[j - 1 + start])) {
CreateDiffRec(c, s1, s2, start, i - 1, j - 1, diffs, pred);
diffs.push_back(Diff('m'));
} else if(j > 0 && (i == 0 || c[i][j-1] >= c[i-1][j])) {
CreateDiffRec(c, s1, s2, start, i, j-1, diffs, pred);
diffs.push_back(Diff('i'));
} else if(i > 0 && (j == 0 || c[i][j-1] < c[i-1][j])) {
CreateDiffRec(c, s1, s2, start, i-1, j, diffs, pred);
diffs.push_back(Diff('d'));
}
}
template <class Sequence, class Pred>
Diffs CreateDiff(const Sequence& s1,
const Sequence& s2,
Pred pred)
{
Diffs diffs;
size_t n = s2.size();
int start = 0;
int m_end = s1.size() - 1;
int n_end = s2.size() - 1;
while(start <= m_end && start <= n_end && pred(s1[start], s2[start])) {
diffs.push_back(Diff('m'));
start++;
}
while(start <= m_end && start <= n_end && pred(s1[m_end], s2[n_end])) {
m_end--;
n_end--;
}
size_t m_new = m_end - start + 1;
size_t n_new = n_end - start + 1;
size_t** c = new size_t*[m_new + 1];
for(size_t i = 0; i <= m_new; ++i) {
c[i] = new size_t[n_new + 1];
c[i][0] = 0;
}
for(size_t j = 0; j <= n_new; ++j)
c[0][j] = 0;
for(size_t i = 1; i <= m_new; ++i)
for(size_t j = 1; j <= n_new; ++j)
if(pred(s1[i - 1 + start], s2[j - 1 + start]))
c[i][j] = c[i-1][j-1] + 1;
else
c[i][j] = c[i][j-1] > c[i-1][j] ? c[i][j-1] : c[i-1][j];
CreateDiffRec(c, s1, s2, start, m_new, n_new, diffs, pred);
for(size_t i = 0; i <= m_new; ++i)
delete[] c[i];
delete[] c;
for (size_t i = n_end + 1; i < n; ++i)
diffs.push_back(Diff('m'));
return diffs;
}
template <class Sequence>
Diffs CreateDiff(const Sequence& s1, const Sequence& s2)
{
return CreateDiff(s1, s2, std::equal_to<typename Sequence::value_type>());
}
template <class Sequence, class Sig, class Stats>
void AddStats(const Sequence& s1, const Sequence& s2, const Sig& sig, Stats& stats)
{
if(sig.size() != stats.size())
throw "Signature size differs from score array size.";
size_t m = 0, d = 0, i = 0, s = 0;
Diffs diff = CreateDiff(s1, s2);
for(int j = 0; j < (int)diff.size(); ++j) {
if(diff[j] == 'm')
m++;
else if(diff[j] == 'd') {
d++;
int k = 0;
while(j - k >= 0 && j + 1 + k < (int)diff.size() &&
diff[j - k] == 'd' && diff[j + 1 + k] == 'i') {
d--;
s++;
k++;
}
j += k;
} else if(diff[j] == 'i')
i++;
}
for(size_t j = 0; j < sig.size(); ++j) {
switch (sig[j]) {
case 'l':
stats[j] += d + i + s;
break;
case 'm':
stats[j] += m;
break;
case 'd':
stats[j] += d;
break;
case 'i':
stats[j] += i;
break;
case 's':
stats[j] += s;
break;
case 'r':
float macc = 1;
if (d + i + s + m)
macc = 1.0 - (float)(d + i + s)/(float)(d + i + s + m);
if(macc > 0)
stats[j] += log(macc);
else
stats[j] += log(1.0/(float)(d + i + s + m + 1));
break;
}
}
}
}
#endif

119
moses/FF/EditOps.cpp Normal file
View File

@ -0,0 +1,119 @@
#include <sstream>
#include "EditOps.h"
#include "moses/Phrase.h"
#include "moses/TargetPhrase.h"
#include "moses/Hypothesis.h"
#include "moses/ChartHypothesis.h"
#include "moses/ScoreComponentCollection.h"
#include "moses/TranslationOption.h"
#include "util/string_piece_hash.hh"
#include "util/exception.hh"
#include <functional>
#include <boost/foreach.hpp>
#include <boost/algorithm/string.hpp>
#include "Diffs.h"
namespace Moses
{
using namespace std;
std::string ParseScores(const std::string &line, const std::string& defaultScores)
{
std::vector<std::string> toks = Tokenize(line);
UTIL_THROW_IF2(toks.empty(), "Empty line");
for (size_t i = 1; i < toks.size(); ++i) {
std::vector<std::string> args = TokenizeFirstOnly(toks[i], "=");
UTIL_THROW_IF2(args.size() != 2,
"Incorrect format for feature function arg: " << toks[i]);
if (args[0] == "scores") {
return args[1];
}
}
return defaultScores;
}
EditOps::EditOps(const std::string &line)
: StatelessFeatureFunction(ParseScores(line, "dis").size(), line)
, m_factorType(0), m_chars(false), m_scores(ParseScores(line, "dis"))
{
std::cerr << "Initializing EditOps feature.." << std::endl;
ReadParameters();
}
void EditOps::SetParameter(const std::string& key, const std::string& value)
{
if (key == "factor") {
m_factorType = Scan<FactorType>(value);
} else if (key == "chars") {
m_chars = Scan<bool>(value);
} else if (key == "scores") {
m_scores = value;
} else {
StatelessFeatureFunction::SetParameter(key, value);
}
}
void EditOps::Load()
{ }
void EditOps::EvaluateInIsolation(const Phrase &source
, const TargetPhrase &target
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection &estimatedFutureScore) const
{
ComputeFeatures(source, target, &scoreBreakdown);
}
void EditOps::ComputeFeatures(
const Phrase &source,
const TargetPhrase& target,
ScoreComponentCollection* accumulator) const
{
std::vector<float> ops(GetNumScoreComponents(), 0);
if(m_chars) {
std::vector<FactorType> factors;
factors.push_back(m_factorType);
std::string sourceStr = source.GetStringRep(factors);
std::string targetStr = target.GetStringRep(factors);
AddStats(sourceStr, targetStr, m_scores, ops);
} else {
std::vector<std::string> sourceTokens;
//std::cerr << "Ed src: ";
for(size_t i = 0; i < source.GetSize(); ++i) {
if(!source.GetWord(i).IsNonTerminal())
sourceTokens.push_back(source.GetWord(i).GetFactor(m_factorType)->GetString().as_string());
//std::cerr << sourceTokens.back() << " ";
}
//std::cerr << std::endl;
std::vector<std::string> targetTokens;
//std::cerr << "Ed trg: ";
for(size_t i = 0; i < target.GetSize(); ++i) {
if(!target.GetWord(i).IsNonTerminal())
targetTokens.push_back(target.GetWord(i).GetFactor(m_factorType)->GetString().as_string());
//std::cerr << targetTokens.back() << " ";
}
//std::cerr << std::endl;
AddStats(sourceTokens, targetTokens, m_scores, ops);
}
accumulator->PlusEquals(this, ops);
}
bool EditOps::IsUseable(const FactorMask &mask) const
{
bool ret = mask[m_factorType];
return ret;
}
}

64
moses/FF/EditOps.h Normal file
View File

@ -0,0 +1,64 @@
#ifndef moses_EditOps_h
#define moses_EditOps_h
#include <string>
#include <boost/unordered_set.hpp>
#include "StatelessFeatureFunction.h"
#include "moses/FactorCollection.h"
#include "moses/AlignmentInfo.h"
namespace Moses
{
typedef std::vector<std::string> Tokens;
/** Calculates string edit operations that transform source phrase into target
* phrase using the LCS algorithm. Potentially usefule for monolingual tasks
* like paraphrasing, summarization, correction.
*/
class EditOps : public StatelessFeatureFunction
{
private:
FactorType m_factorType;
bool m_chars;
std::string m_scores;
public:
EditOps(const std::string &line);
bool IsUseable(const FactorMask &mask) const;
void Load();
virtual void EvaluateInIsolation(const Phrase &source
, const TargetPhrase &targetPhrase
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection &estimatedFutureScore) const;
void EvaluateWithSourceContext(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, const StackVec *stackVec
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection *estimatedFutureScore = NULL) const
{}
void EvaluateWhenApplied(const Hypothesis& hypo,
ScoreComponentCollection* accumulator) const
{}
void EvaluateWhenApplied(const ChartHypothesis &hypo,
ScoreComponentCollection* accumulator) const
{}
void EvaluateTranslationOptionListWithSourceContext(const InputType &input
, const TranslationOptionList &translationOptionList) const
{}
void ComputeFeatures(const Phrase &source,
const TargetPhrase& targetPhrase,
ScoreComponentCollection* accumulator) const;
void SetParameter(const std::string& key, const std::string& value);
};
}
#endif // moses_CorrectionPattern_h

12
moses/FF/Factory.cpp
View File

@ -73,8 +73,14 @@
#include "moses/Syntax/InputWeightFF.h"
#include "moses/Syntax/RuleTableFF.h"
#include "moses/FF/EditOps.h"
#include "moses/FF/CorrectionPattern.h"
#ifdef HAVE_VW
#include "moses/FF/VW/VW.h"
#include "moses/FF/VW/VWFeatureContextBigrams.h"
#include "moses/FF/VW/VWFeatureContextBilingual.h"
#include "moses/FF/VW/VWFeatureContextWindow.h"
#include "moses/FF/VW/VWFeatureSourceBagOfWords.h"
#include "moses/FF/VW/VWFeatureSourceBigrams.h"
#include "moses/FF/VW/VWFeatureSourceIndicator.h"
@ -294,8 +300,14 @@ FeatureRegistry::FeatureRegistry()
MOSES_FNAME(SkeletonTranslationOptionListFeature);
MOSES_FNAME(SkeletonPT);
MOSES_FNAME(EditOps);
MOSES_FNAME(CorrectionPattern);
#ifdef HAVE_VW
MOSES_FNAME(VW);
MOSES_FNAME(VWFeatureContextBigrams);
MOSES_FNAME(VWFeatureContextBilingual);
MOSES_FNAME(VWFeatureContextWindow);
MOSES_FNAME(VWFeatureSourceBagOfWords);
MOSES_FNAME(VWFeatureSourceBigrams);
MOSES_FNAME(VWFeatureSourceIndicator);

3
moses/FF/OSM-Feature/KenOSM.cpp
View File

@ -3,10 +3,11 @@
namespace Moses
{
OSMLM* ConstructOSMLM(const char *file)
OSMLM* ConstructOSMLM(const char *file, util::LoadMethod load_method)
{
lm::ngram::ModelType model_type;
lm::ngram::Config config;
config.load_method = load_method;
if (lm::ngram::RecognizeBinary(file, model_type)) {
switch(model_type) {
case lm::ngram::PROBING:

2
moses/FF/OSM-Feature/KenOSM.h
View File

@ -47,7 +47,7 @@ private:
typedef KenOSMBase OSMLM;
OSMLM* ConstructOSMLM(const char *file);
OSMLM* ConstructOSMLM(const char *file, util::LoadMethod load_method);
} // namespace

17
moses/FF/OSM-Feature/OpSequenceModel.cpp
View File

@ -17,6 +17,7 @@ OpSequenceModel::OpSequenceModel(const std::string &line)
tFactor = 0;
numFeatures = 5;
ReadParameters();
load_method = util::READ;
}
OpSequenceModel::~OpSequenceModel()
@ -27,7 +28,7 @@ OpSequenceModel::~OpSequenceModel()
void OpSequenceModel :: readLanguageModel(const char *lmFile)
{
string unkOp = "_TRANS_SLF_";
OSM = ConstructOSMLM(m_lmPath.c_str());
OSM = ConstructOSMLM(m_lmPath.c_str(), load_method);
State startState = OSM->NullContextState();
State endState;
@ -248,6 +249,20 @@ void OpSequenceModel::SetParameter(const std::string& key, const std::string& va
sFactor = Scan<int>(value);
} else if (key == "output-factor") {
tFactor = Scan<int>(value);
} else if (key == "load") {
if (value == "lazy") {
load_method = util::LAZY;
} else if (value == "populate_or_lazy") {
load_method = util::POPULATE_OR_LAZY;
} else if (value == "populate_or_read" || value == "populate") {
load_method = util::POPULATE_OR_READ;
} else if (value == "read") {
load_method = util::READ;
} else if (value == "parallel_read") {
load_method = util::PARALLEL_READ;
} else {
UTIL_THROW2("Unknown KenLM load method " << value);
}
} else {
StatefulFeatureFunction::SetParameter(key, value);
}

1
moses/FF/OSM-Feature/OpSequenceModel.h
View File

@ -20,6 +20,7 @@ public:
int sFactor; // Source Factor ...
int tFactor; // Target Factor ...
int numFeatures; // Number of features used ...
util::LoadMethod load_method; // method to load model
OpSequenceModel(const std::string &line);
~OpSequenceModel();

40
moses/FF/VW/AlignmentConstraint.h Normal file
View File

@ -0,0 +1,40 @@
#pragma once
namespace Moses
{
/**
* Helper class for storing alignment constraints.
*/
class AlignmentConstraint
{
public:
AlignmentConstraint() : m_min(std::numeric_limits<int>::max()), m_max(-1) {}
AlignmentConstraint(int min, int max) : m_min(min), m_max(max) {}
/**
* We are aligned to point => our min cannot be larger, our max cannot be smaller.
*/
void Update(int point) {
if (m_min > point) m_min = point;
if (m_max < point) m_max = point;
}
bool IsSet() const {
return m_max != -1;
}
int GetMin() const {
return m_min;
}
int GetMax() const {
return m_max;
}
private:
int m_min, m_max;
};
}

637
moses/FF/VW/VW.cpp Normal file
View File

@ -0,0 +1,637 @@
#include <string>
#include <map>
#include <limits>
#include <vector>
#include <boost/unordered_map.hpp>
#include <boost/functional/hash.hpp>
#include "moses/FF/StatefulFeatureFunction.h"
#include "moses/PP/CountsPhraseProperty.h"
#include "moses/TranslationOptionList.h"
#include "moses/TranslationOption.h"
#include "moses/Util.h"
#include "moses/TypeDef.h"
#include "moses/StaticData.h"
#include "moses/Phrase.h"
#include "moses/AlignmentInfo.h"
#include "moses/AlignmentInfoCollection.h"
#include "moses/Word.h"
#include "moses/FactorCollection.h"
#include "Normalizer.h"
#include "Classifier.h"
#include "VWFeatureBase.h"
#include "TabbedSentence.h"
#include "ThreadLocalByFeatureStorage.h"
#include "TrainingLoss.h"
#include "VWTargetSentence.h"
#include "VWState.h"
#include "VW.h"
namespace Moses
{
VW::VW(const std::string &line)
: StatefulFeatureFunction(1, line)
, TLSTargetSentence(this)
, m_train(false)
, m_sentenceStartWord(Word())
{
ReadParameters();
Discriminative::ClassifierFactory *classifierFactory = m_train
? new Discriminative::ClassifierFactory(m_modelPath)
: new Discriminative::ClassifierFactory(m_modelPath, m_vwOptions);
m_tlsClassifier = new TLSClassifier(this, *classifierFactory);
m_tlsFutureScores = new TLSFloatHashMap(this);
m_tlsComputedStateExtensions = new TLSStateExtensions(this);
m_tlsTranslationOptionFeatures = new TLSFeatureVectorMap(this);
m_tlsTargetContextFeatures = new TLSFeatureVectorMap(this);
if (! m_normalizer) {
VERBOSE(1, "VW :: No loss function specified, assuming logistic loss.\n");
m_normalizer = (Discriminative::Normalizer *) new Discriminative::LogisticLossNormalizer();
}
if (! m_trainingLoss) {
VERBOSE(1, "VW :: Using basic 1/0 loss calculation in training.\n");
m_trainingLoss = (TrainingLoss *) new TrainingLossBasic();
}
// create a virtual beginning-of-sentence word with all factors replaced by <S>
const Factor *bosFactor = FactorCollection::Instance().AddFactor(BOS_);
for (size_t i = 0; i < MAX_NUM_FACTORS; i++)
m_sentenceStartWord.SetFactor(i, bosFactor);
}
VW::~VW()
{
delete m_tlsClassifier;
delete m_normalizer;
// TODO delete more stuff
}
FFState* VW::EvaluateWhenApplied(
const Hypothesis& curHypo,
const FFState* prevState,
ScoreComponentCollection* accumulator) const
{
VERBOSE(3, "VW :: Evaluating translation options\n");
const VWState& prevVWState = *static_cast<const VWState *>(prevState);
const std::vector<VWFeatureBase*>& contextFeatures =
VWFeatureBase::GetTargetContextFeatures(GetScoreProducerDescription());
if (contextFeatures.empty()) {
// no target context features => we already evaluated everything in
// EvaluateTranslationOptionListWithSourceContext(). Nothing to do now,
// no state information to track.
return new VWState();
}
size_t spanStart = curHypo.GetTranslationOption().GetStartPos();
size_t spanEnd = curHypo.GetTranslationOption().GetEndPos();
// compute our current key
size_t cacheKey = MakeCacheKey(prevState, spanStart, spanEnd);
boost::unordered_map<size_t, FloatHashMap> &computedStateExtensions
= *m_tlsComputedStateExtensions->GetStored();
if (computedStateExtensions.find(cacheKey) == computedStateExtensions.end()) {
// we have not computed this set of translation options yet
const TranslationOptionList *topts =
curHypo.GetManager().getSntTranslationOptions()->GetTranslationOptionList(spanStart, spanEnd);
const InputType& input = curHypo.GetManager().GetSource();
Discriminative::Classifier &classifier = *m_tlsClassifier->GetStored();
// extract target context features
size_t contextHash = prevVWState.hash();
FeatureVectorMap &contextFeaturesCache = *m_tlsTargetContextFeatures->GetStored();
FeatureVectorMap::const_iterator contextIt = contextFeaturesCache.find(contextHash);
if (contextIt == contextFeaturesCache.end()) {
// we have not extracted features for this context yet
const Phrase &targetContext = prevVWState.GetPhrase();
Discriminative::FeatureVector contextVector;
const AlignmentInfo *alignInfo = TransformAlignmentInfo(curHypo, targetContext.GetSize());
for(size_t i = 0; i < contextFeatures.size(); ++i)
(*contextFeatures[i])(input, targetContext, *alignInfo, classifier, contextVector);
contextFeaturesCache[contextHash] = contextVector;
VERBOSE(3, "VW :: context cache miss\n");
} else {
// context already in cache, simply put feature IDs in the classifier object
classifier.AddLabelIndependentFeatureVector(contextIt->second);
VERBOSE(3, "VW :: context cache hit\n");
}
std::vector<float> losses(topts->size());
for (size_t toptIdx = 0; toptIdx < topts->size(); toptIdx++) {
const TranslationOption *topt = topts->Get(toptIdx);
const TargetPhrase &targetPhrase = topt->GetTargetPhrase();
size_t toptHash = hash_value(*topt);
// start with pre-computed source-context-only VW scores
losses[toptIdx] = m_tlsFutureScores->GetStored()->find(toptHash)->second;
// add all features associated with this translation option
// (pre-computed when evaluated with source context)
const Discriminative::FeatureVector &targetFeatureVector =
m_tlsTranslationOptionFeatures->GetStored()->find(toptHash)->second;
classifier.AddLabelDependentFeatureVector(targetFeatureVector);
// add classifier score with context+target features only to the total loss
losses[toptIdx] += classifier.Predict(MakeTargetLabel(targetPhrase));
}
// normalize classifier scores to get a probability distribution
(*m_normalizer)(losses);
// fill our cache with the results
FloatHashMap &toptScores = computedStateExtensions[cacheKey];
for (size_t toptIdx = 0; toptIdx < topts->size(); toptIdx++) {
const TranslationOption *topt = topts->Get(toptIdx);
size_t toptHash = hash_value(*topt);
toptScores[toptHash] = FloorScore(TransformScore(losses[toptIdx]));
}
VERBOSE(3, "VW :: cache miss\n");
} else {
VERBOSE(3, "VW :: cache hit\n");
}
// now our cache is guaranteed to contain the required score, simply look it up
std::vector<float> newScores(m_numScoreComponents);
size_t toptHash = hash_value(curHypo.GetTranslationOption());
newScores[0] = computedStateExtensions[cacheKey][toptHash];
VERBOSE(3, "VW :: adding score: " << newScores[0] << "\n");
accumulator->PlusEquals(this, newScores);
return new VWState(prevVWState, curHypo);
}
const FFState* VW::EmptyHypothesisState(const InputType &input) const
{
size_t maxContextSize = VWFeatureBase::GetMaximumContextSize(GetScoreProducerDescription());
Phrase initialPhrase;
for (size_t i = 0; i < maxContextSize; i++)
initialPhrase.AddWord(m_sentenceStartWord);
return new VWState(initialPhrase);
}
void VW::EvaluateTranslationOptionListWithSourceContext(const InputType &input
, const TranslationOptionList &translationOptionList) const
{
Discriminative::Classifier &classifier = *m_tlsClassifier->GetStored();
if (translationOptionList.size() == 0)
return; // nothing to do
VERBOSE(3, "VW :: Evaluating translation options\n");
// which feature functions do we use (on the source and target side)
const std::vector<VWFeatureBase*>& sourceFeatures =
VWFeatureBase::GetSourceFeatures(GetScoreProducerDescription());
const std::vector<VWFeatureBase*>& contextFeatures =
VWFeatureBase::GetTargetContextFeatures(GetScoreProducerDescription());
const std::vector<VWFeatureBase*>& targetFeatures =
VWFeatureBase::GetTargetFeatures(GetScoreProducerDescription());
size_t maxContextSize = VWFeatureBase::GetMaximumContextSize(GetScoreProducerDescription());
// only use stateful score computation when needed
bool haveTargetContextFeatures = ! contextFeatures.empty();
const Range &sourceRange = translationOptionList.Get(0)->GetSourceWordsRange();
if (m_train) {
//
// extract features for training the classifier (only call this when using vwtrainer, not in Moses!)
//
// find which topts are correct
std::vector<bool> correct(translationOptionList.size());
std::vector<int> startsAt(translationOptionList.size());
std::set<int> uncoveredStartingPositions;
for (size_t i = 0; i < translationOptionList.size(); i++) {
std::pair<bool, int> isCorrect = IsCorrectTranslationOption(* translationOptionList.Get(i));
correct[i] = isCorrect.first;
startsAt[i] = isCorrect.second;
if (isCorrect.first) {
uncoveredStartingPositions.insert(isCorrect.second);
}
}
// optionally update translation options using leave-one-out
std::vector<bool> keep = (m_leaveOneOut.size() > 0)
? LeaveOneOut(translationOptionList, correct)
: std::vector<bool>(translationOptionList.size(), true);
while (! uncoveredStartingPositions.empty()) {
int currentStart = *uncoveredStartingPositions.begin();
uncoveredStartingPositions.erase(uncoveredStartingPositions.begin());
// check whether we (still) have some correct translation
int firstCorrect = -1;
for (size_t i = 0; i < translationOptionList.size(); i++) {
if (keep[i] && correct[i] && startsAt[i] == currentStart) {
firstCorrect = i;
break;
}
}
// do not train if there are no positive examples
if (firstCorrect == -1) {
VERBOSE(3, "VW :: skipping topt collection, no correct translation for span at current tgt start position\n");
continue;
}
// the first correct topt can be used by some loss functions
const TargetPhrase &correctPhrase = translationOptionList.Get(firstCorrect)->GetTargetPhrase();
// feature extraction *at prediction time* outputs feature hashes which can be cached;
// this is training time, simply store everything in this dummyVector
Discriminative::FeatureVector dummyVector;
// extract source side features
for(size_t i = 0; i < sourceFeatures.size(); ++i)
(*sourceFeatures[i])(input, sourceRange, classifier, dummyVector);
// build target-side context
Phrase targetContext;
for (size_t i = 0; i < maxContextSize; i++)
targetContext.AddWord(m_sentenceStartWord);
const Phrase *targetSent = GetStored()->m_sentence;
// word alignment info shifted by context size
AlignmentInfo contextAlignment = TransformAlignmentInfo(*GetStored()->m_alignment, maxContextSize, currentStart);
if (currentStart > 0)
targetContext.Append(targetSent->GetSubString(Range(0, currentStart - 1)));
// extract target-context features
for(size_t i = 0; i < contextFeatures.size(); ++i)
(*contextFeatures[i])(input, targetContext, contextAlignment, classifier, dummyVector);
// go over topts, extract target side features and train the classifier
for (size_t toptIdx = 0; toptIdx < translationOptionList.size(); toptIdx++) {
// this topt was discarded by leaving one out
if (! keep[toptIdx])
continue;
// extract target-side features for each topt
const TargetPhrase &targetPhrase = translationOptionList.Get(toptIdx)->GetTargetPhrase();
for(size_t i = 0; i < targetFeatures.size(); ++i)
(*targetFeatures[i])(input, targetPhrase, classifier, dummyVector);
bool isCorrect = correct[toptIdx] && startsAt[toptIdx] == currentStart;
float loss = (*m_trainingLoss)(targetPhrase, correctPhrase, isCorrect);
// train classifier on current example
classifier.Train(MakeTargetLabel(targetPhrase), loss);
}
}
} else {
//
// predict using a trained classifier, use this in decoding (=at test time)
//
std::vector<float> losses(translationOptionList.size());
Discriminative::FeatureVector outFeaturesSourceNamespace;
// extract source side features
for(size_t i = 0; i < sourceFeatures.size(); ++i)
(*sourceFeatures[i])(input, sourceRange, classifier, outFeaturesSourceNamespace);
for (size_t toptIdx = 0; toptIdx < translationOptionList.size(); toptIdx++) {
const TranslationOption *topt = translationOptionList.Get(toptIdx);
const TargetPhrase &targetPhrase = topt->GetTargetPhrase();
Discriminative::FeatureVector outFeaturesTargetNamespace;
// extract target-side features for each topt
for(size_t i = 0; i < targetFeatures.size(); ++i)
(*targetFeatures[i])(input, targetPhrase, classifier, outFeaturesTargetNamespace);
// cache the extracted target features (i.e. features associated with given topt)
// for future use at decoding time
size_t toptHash = hash_value(*topt);
m_tlsTranslationOptionFeatures->GetStored()->insert(
std::make_pair(toptHash, outFeaturesTargetNamespace));
// get classifier score
losses[toptIdx] = classifier.Predict(MakeTargetLabel(targetPhrase));
}
// normalize classifier scores to get a probability distribution
std::vector<float> rawLosses = losses;
(*m_normalizer)(losses);
// update scores of topts
for (size_t toptIdx = 0; toptIdx < translationOptionList.size(); toptIdx++) {
TranslationOption *topt = *(translationOptionList.begin() + toptIdx);
if (! haveTargetContextFeatures) {
// no target context features; evaluate the FF now
std::vector<float> newScores(m_numScoreComponents);
newScores[0] = FloorScore(TransformScore(losses[toptIdx]));
ScoreComponentCollection &scoreBreakDown = topt->GetScoreBreakdown();
scoreBreakDown.PlusEquals(this, newScores);
topt->UpdateScore();
} else {
// We have target context features => this is just a partial score,
// do not add it to the score component collection.
size_t toptHash = hash_value(*topt);
// Subtract the score contribution of target-only features, otherwise it would
// be included twice.
Discriminative::FeatureVector emptySource;
const Discriminative::FeatureVector &targetFeatureVector =
m_tlsTranslationOptionFeatures->GetStored()->find(toptHash)->second;
classifier.AddLabelIndependentFeatureVector(emptySource);
classifier.AddLabelDependentFeatureVector(targetFeatureVector);
float targetOnlyLoss = classifier.Predict(VW_DUMMY_LABEL);
float futureScore = rawLosses[toptIdx] - targetOnlyLoss;
m_tlsFutureScores->GetStored()->insert(std::make_pair(toptHash, futureScore));
}
}
}
}
void VW::SetParameter(const std::string& key, const std::string& value)
{
if (key == "train") {
m_train = Scan<bool>(value);
} else if (key == "path") {
m_modelPath = value;
} else if (key == "vw-options") {
m_vwOptions = value;
} else if (key == "leave-one-out-from") {
m_leaveOneOut = value;
} else if (key == "training-loss") {
// which type of loss to use for training
if (value == "basic") {
m_trainingLoss = (TrainingLoss *) new TrainingLossBasic();
} else if (value == "bleu") {
m_trainingLoss = (TrainingLoss *) new TrainingLossBLEU();
} else {
UTIL_THROW2("Unknown training loss type:" << value);
}
} else if (key == "loss") {
// which normalizer to use (theoretically depends on the loss function used for training the
// classifier (squared/logistic/hinge/...), hence the name "loss"
if (value == "logistic") {
m_normalizer = (Discriminative::Normalizer *) new Discriminative::LogisticLossNormalizer();
} else if (value == "squared") {
m_normalizer = (Discriminative::Normalizer *) new Discriminative::SquaredLossNormalizer();
} else {
UTIL_THROW2("Unknown loss type:" << value);
}
} else {
StatefulFeatureFunction::SetParameter(key, value);
}
}
void VW::InitializeForInput(ttasksptr const& ttask)
{
// do not keep future cost estimates across sentences!
m_tlsFutureScores->GetStored()->clear();
// invalidate our caches after each sentence
m_tlsComputedStateExtensions->GetStored()->clear();
// it's not certain that we should clear these caches; we do it
// because they shouldn't be allowed to grow indefinitely large but
// target contexts and translation options will have identical features
// the next time we extract them...
m_tlsTargetContextFeatures->GetStored()->clear();
m_tlsTranslationOptionFeatures->GetStored()->clear();
InputType const& source = *(ttask->GetSource().get());
// tabbed sentence is assumed only in training
if (! m_train)
return;
UTIL_THROW_IF2(source.GetType() != TabbedSentenceInput,
"This feature function requires the TabbedSentence input type");
const TabbedSentence& tabbedSentence = static_cast<const TabbedSentence&>(source);
UTIL_THROW_IF2(tabbedSentence.GetColumns().size() < 2,
"TabbedSentence must contain target<tab>alignment");
// target sentence represented as a phrase
Phrase *target = new Phrase();
target->CreateFromString(
Output
, StaticData::Instance().options()->output.factor_order
, tabbedSentence.GetColumns()[0]
, NULL);
// word alignment between source and target sentence
// we don't store alignment info in AlignmentInfoCollection because we keep alignments of whole
// sentences, not phrases
AlignmentInfo *alignment = new AlignmentInfo(tabbedSentence.GetColumns()[1]);
VWTargetSentence &targetSent = *GetStored();
targetSent.Clear();
targetSent.m_sentence = target;
targetSent.m_alignment = alignment;
// pre-compute max- and min- aligned points for faster translation option checking
targetSent.SetConstraints(source.GetSize());
}
/*************************************************************************************
* private methods
************************************************************************************/
const AlignmentInfo *VW::TransformAlignmentInfo(const Hypothesis &curHypo, size_t contextSize) const
{
std::set<std::pair<size_t, size_t> > alignmentPoints;
const Hypothesis *contextHypo = curHypo.GetPrevHypo();
int idxInContext = contextSize - 1;
int processedWordsInHypo = 0;
while (idxInContext >= 0 && contextHypo) {
int idxInHypo = contextHypo->GetCurrTargetLength() - 1 - processedWordsInHypo;
if (idxInHypo >= 0) {
const AlignmentInfo &hypoAlign = contextHypo->GetCurrTargetPhrase().GetAlignTerm();
std::set<size_t> alignedToTgt = hypoAlign.GetAlignmentsForTarget(idxInHypo);
size_t srcOffset = contextHypo->GetCurrSourceWordsRange().GetStartPos();
BOOST_FOREACH(size_t srcIdx, alignedToTgt) {
alignmentPoints.insert(std::make_pair(srcOffset + srcIdx, idxInContext));
}
processedWordsInHypo++;
idxInContext--;
} else {
processedWordsInHypo = 0;
contextHypo = contextHypo->GetPrevHypo();
}
}
return AlignmentInfoCollection::Instance().Add(alignmentPoints);
}
AlignmentInfo VW::TransformAlignmentInfo(const AlignmentInfo &alignInfo, size_t contextSize, int currentStart) const
{
std::set<std::pair<size_t, size_t> > alignmentPoints;
for (int i = std::max(0, currentStart - (int)contextSize); i < currentStart; i++) {
std::set<size_t> alignedToTgt = alignInfo.GetAlignmentsForTarget(i);
BOOST_FOREACH(size_t srcIdx, alignedToTgt) {
alignmentPoints.insert(std::make_pair(srcIdx, i + contextSize));
}
}
return AlignmentInfo(alignmentPoints);
}
std::pair<bool, int> VW::IsCorrectTranslationOption(const TranslationOption &topt) const
{
//std::cerr << topt.GetSourceWordsRange() << std::endl;
int sourceStart = topt.GetSourceWordsRange().GetStartPos();
int sourceEnd = topt.GetSourceWordsRange().GetEndPos();
const VWTargetSentence &targetSentence = *GetStored();
// [targetStart, targetEnd] spans aligned target words
int targetStart = targetSentence.m_sentence->GetSize();
int targetEnd = -1;
// get the left-most and right-most alignment point within source span
for(int i = sourceStart; i <= sourceEnd; ++i) {
if(targetSentence.m_sourceConstraints[i].IsSet()) {
if(targetStart > targetSentence.m_sourceConstraints[i].GetMin())
targetStart = targetSentence.m_sourceConstraints[i].GetMin();
if(targetEnd < targetSentence.m_sourceConstraints[i].GetMax())
targetEnd = targetSentence.m_sourceConstraints[i].GetMax();
}
}
// there was no alignment
if(targetEnd == -1)
return std::make_pair(false, -1);
//std::cerr << "Shorter: " << targetStart << " " << targetEnd << std::endl;
// [targetStart2, targetEnd2] spans unaligned words left and right of [targetStart, targetEnd]
int targetStart2 = targetStart;
for(int i = targetStart2; i >= 0 && !targetSentence.m_targetConstraints[i].IsSet(); --i)
targetStart2 = i;
int targetEnd2 = targetEnd;
for(int i = targetEnd2;
i < targetSentence.m_sentence->GetSize() && !targetSentence.m_targetConstraints[i].IsSet();
++i)
targetEnd2 = i;
//std::cerr << "Longer: " << targetStart2 << " " << targetEnd2 << std::endl;
const TargetPhrase &tphrase = topt.GetTargetPhrase();
//std::cerr << tphrase << std::endl;
// if target phrase is shorter than inner span return false
if(tphrase.GetSize() < targetEnd - targetStart + 1)
return std::make_pair(false, -1);
// if target phrase is longer than outer span return false
if(tphrase.GetSize() > targetEnd2 - targetStart2 + 1)
return std::make_pair(false, -1);
// for each possible starting point
for(int tempStart = targetStart2; tempStart <= targetStart; tempStart++) {
bool found = true;
// check if the target phrase is within longer span
for(int i = tempStart; i <= targetEnd2 && i < tphrase.GetSize() + tempStart; ++i) {
if(tphrase.GetWord(i - tempStart) != targetSentence.m_sentence->GetWord(i)) {
found = false;
break;
}
}
// return true if there was a match
if(found) {
//std::cerr << "Found" << std::endl;
return std::make_pair(true, tempStart);
}
}
return std::make_pair(false, -1);
}
std::vector<bool> VW::LeaveOneOut(const TranslationOptionList &topts, const std::vector<bool> &correct) const
{
UTIL_THROW_IF2(m_leaveOneOut.size() == 0 || ! m_train, "LeaveOneOut called in wrong setting!");
float sourceRawCount = 0.0;
const float ONE = 1.0001; // I don't understand floating point numbers
std::vector<bool> keepOpt;
for (size_t i = 0; i < topts.size(); i++) {
TranslationOption *topt = *(topts.begin() + i);
const TargetPhrase &targetPhrase = topt->GetTargetPhrase();
// extract raw counts from phrase-table property
const CountsPhraseProperty *property =
static_cast<const CountsPhraseProperty *>(targetPhrase.GetProperty("Counts"));
if (! property) {
VERBOSE(2, "VW :: Counts not found for topt! Is this an OOV?\n");
// keep all translation opts without updating, this is either OOV or bad usage...
keepOpt.assign(topts.size(), true);
return keepOpt;
}
if (sourceRawCount == 0.0) {
sourceRawCount = property->GetSourceMarginal() - ONE; // discount one occurrence of the source phrase
if (sourceRawCount <= 0) {
// no translation options survived, source phrase was a singleton
keepOpt.assign(topts.size(), false);
return keepOpt;
}
}
float discount = correct[i] ? ONE : 0.0;
float target = property->GetTargetMarginal() - discount;
float joint = property->GetJointCount() - discount;
if (discount != 0.0) VERBOSE(3, "VW :: leaving one out!\n");
if (joint > 0) {
// topt survived leaving one out, update its scores
const FeatureFunction *feature = &FindFeatureFunction(m_leaveOneOut);
std::vector<float> scores = targetPhrase.GetScoreBreakdown().GetScoresForProducer(feature);
UTIL_THROW_IF2(scores.size() != 4, "Unexpected number of scores in feature " << m_leaveOneOut);
scores[0] = TransformScore(joint / target); // P(f|e)
scores[2] = TransformScore(joint / sourceRawCount); // P(e|f)
ScoreComponentCollection &scoreBreakDown = topt->GetScoreBreakdown();
scoreBreakDown.Assign(feature, scores);
topt->UpdateScore();
keepOpt.push_back(true);
} else {
// they only occurred together once, discard topt
VERBOSE(2, "VW :: discarded topt when leaving one out\n");
keepOpt.push_back(false);
}
}
return keepOpt;
}
} // namespace Moses

522
moses/FF/VW/VW.h
View File

@ -3,8 +3,12 @@
#include <string>
#include <map>
#include <limits>
#include <vector>
#include "moses/FF/StatelessFeatureFunction.h"
#include <boost/unordered_map.hpp>
#include <boost/functional/hash.hpp>
#include "moses/FF/StatefulFeatureFunction.h"
#include "moses/PP/CountsPhraseProperty.h"
#include "moses/TranslationOptionList.h"
#include "moses/TranslationOption.h"
@ -13,6 +17,8 @@
#include "moses/StaticData.h"
#include "moses/Phrase.h"
#include "moses/AlignmentInfo.h"
#include "moses/Word.h"
#include "moses/FactorCollection.h"
#include "Normalizer.h"
#include "Classifier.h"
@ -20,119 +26,50 @@
#include "TabbedSentence.h"
#include "ThreadLocalByFeatureStorage.h"
#include "TrainingLoss.h"
#include "VWTargetSentence.h"
/*
* VW classifier feature. See vw/README.md for further information.
*
* TODO: say which paper to cite.
*/
namespace Moses
{
const std::string VW_DUMMY_LABEL = "1111"; // VW does not use the actual label, other classifiers might
/**
* Helper class for storing alignment constraints.
*/
class Constraint
{
public:
Constraint() : m_min(std::numeric_limits<int>::max()), m_max(-1) {}
Constraint(int min, int max) : m_min(min), m_max(max) {}
/**
* We are aligned to point => our min cannot be larger, our max cannot be smaller.
*/
void Update(int point) {
if (m_min > point) m_min = point;
if (m_max < point) m_max = point;
}
bool IsSet() const {
return m_max != -1;
}
int GetMin() const {
return m_min;
}
int GetMax() const {
return m_max;
}
private:
int m_min, m_max;
};
/**
* VW thread-specific data about target sentence.
*/
struct VWTargetSentence {
VWTargetSentence() : m_sentence(NULL), m_alignment(NULL) {}
void Clear() {
if (m_sentence) delete m_sentence;
if (m_alignment) delete m_alignment;
}
~VWTargetSentence() {
Clear();
}
void SetConstraints(size_t sourceSize) {
// initialize to unconstrained
m_sourceConstraints.assign(sourceSize, Constraint());
m_targetConstraints.assign(m_sentence->GetSize(), Constraint());
// set constraints according to alignment points
AlignmentInfo::const_iterator it;
for (it = m_alignment->begin(); it != m_alignment->end(); it++) {
int src = it->first;
int tgt = it->second;
if (src >= m_sourceConstraints.size() || tgt >= m_targetConstraints.size()) {
UTIL_THROW2("VW :: alignment point out of bounds: " << src << "-" << tgt);
}
m_sourceConstraints[src].Update(tgt);
m_targetConstraints[tgt].Update(src);
}
}
Phrase *m_sentence;
AlignmentInfo *m_alignment;
std::vector<Constraint> m_sourceConstraints, m_targetConstraints;
};
// dummy class label; VW does not use the actual label, other classifiers might
const std::string VW_DUMMY_LABEL = "1111";
// thread-specific classifier instance
typedef ThreadLocalByFeatureStorage<Discriminative::Classifier, Discriminative::ClassifierFactory &> TLSClassifier;
// current target sentence, used in VW training (vwtrainer), not in decoding (prediction time)
typedef ThreadLocalByFeatureStorage<VWTargetSentence> TLSTargetSentence;
class VW : public StatelessFeatureFunction, public TLSTargetSentence
// hash table of feature vectors
typedef boost::unordered_map<size_t, Discriminative::FeatureVector> FeatureVectorMap;
// thread-specific feature vector hash
typedef ThreadLocalByFeatureStorage<FeatureVectorMap> TLSFeatureVectorMap;
// hash table of partial scores
typedef boost::unordered_map<size_t, float> FloatHashMap;
// thread-specific score hash table, used for caching
typedef ThreadLocalByFeatureStorage<FloatHashMap> TLSFloatHashMap;
// thread-specific hash tablei for caching full classifier outputs
typedef ThreadLocalByFeatureStorage<boost::unordered_map<size_t, FloatHashMap> > TLSStateExtensions;
/*
* VW feature function. A discriminative classifier with source and target context features.
*/
class VW : public StatefulFeatureFunction, public TLSTargetSentence
{
public:
VW(const std::string &line)
: StatelessFeatureFunction(1, line)
, TLSTargetSentence(this)
, m_train(false) {
ReadParameters();
Discriminative::ClassifierFactory *classifierFactory = m_train
? new Discriminative::ClassifierFactory(m_modelPath)
: new Discriminative::ClassifierFactory(m_modelPath, m_vwOptions);
VW(const std::string &line);
m_tlsClassifier = new TLSClassifier(this, *classifierFactory);
if (! m_normalizer) {
VERBOSE(1, "VW :: No loss function specified, assuming logistic loss.\n");
m_normalizer = (Discriminative::Normalizer *) new Discriminative::LogisticLossNormalizer();
}
if (! m_trainingLoss) {
VERBOSE(1, "VW :: Using basic 1/0 loss calculation in training.\n");
m_trainingLoss = (TrainingLoss *) new TrainingLossBasic();
}
}
virtual ~VW() {
delete m_tlsClassifier;
delete m_normalizer;
}
virtual ~VW();
bool IsUseable(const FactorMask &mask) const {
return true;
@ -152,335 +89,89 @@ public:
, ScoreComponentCollection *estimatedFutureScore = NULL) const {
}
void EvaluateTranslationOptionListWithSourceContext(const InputType &input
, const TranslationOptionList &translationOptionList) const {
Discriminative::Classifier &classifier = *m_tlsClassifier->GetStored();
if (translationOptionList.size() == 0)
return; // nothing to do
VERBOSE(2, "VW :: Evaluating translation options\n");
// which feature functions do we use (on the source and target side)
const std::vector<VWFeatureBase*>& sourceFeatures =
VWFeatureBase::GetSourceFeatures(GetScoreProducerDescription());
const std::vector<VWFeatureBase*>& targetFeatures =
VWFeatureBase::GetTargetFeatures(GetScoreProducerDescription());
const Range &sourceRange = translationOptionList.Get(0)->GetSourceWordsRange();
const InputPath &inputPath = translationOptionList.Get(0)->GetInputPath();
if (m_train) {
// This behavior of this method depends on whether it's called during VW
// training (feature extraction) by vwtrainer or during decoding (prediction
// time) by Moses.
//
// extract features for training the classifier (only call this when using vwtrainer, not in Moses!)
// When predicting, it evaluates all translation options with the VW model;
// if no target-context features are defined, this is the final score and it
// is added directly to the TranslationOption score. If there are target
// context features, the score is a partial score and it is only stored in
// cache; the final score is computed based on target context in
// EvaluateWhenApplied().
//
// This method is also used in training by vwtrainer in which case features
// are written to a file, no classifier predictions take place. Target-side
// context is constant at training time (we know the true target sentence),
// so target-context features are extracted here as well.
virtual void EvaluateTranslationOptionListWithSourceContext(const InputType &input
, const TranslationOptionList &translationOptionList) const;
// find which topts are correct
std::vector<bool> correct(translationOptionList.size());
for (size_t i = 0; i < translationOptionList.size(); i++)
correct[i] = IsCorrectTranslationOption(* translationOptionList.Get(i));
// Evaluate VW during decoding. This is only used at prediction time (not in training).
// When no target-context features are defined, VW predictions were already fully calculated
// in EvaluateTranslationOptionListWithSourceContext() and the scores were added to the model.
// If there are target-context features, we compute the context-dependent part of the
// classifier score and combine it with the source-context only partial score which was computed
// in EvaluateTranslationOptionListWithSourceContext(). Various caches are used to make this
// method more efficient.
virtual FFState* EvaluateWhenApplied(
const Hypothesis& curHypo,
const FFState* prevState,
ScoreComponentCollection* accumulator) const;
// optionally update translation options using leave-one-out
std::vector<bool> keep = (m_leaveOneOut.size() > 0)
? LeaveOneOut(translationOptionList, correct)
: std::vector<bool>(translationOptionList.size(), true);
// check whether we (still) have some correct translation
int firstCorrect = -1;
for (size_t i = 0; i < translationOptionList.size(); i++) {
if (keep[i] && correct[i]) {
firstCorrect = i;
break;
}
}
// do not train if there are no positive examples
if (firstCorrect == -1) {
VERBOSE(2, "VW :: skipping topt collection, no correct translation for span\n");
return;
}
// the first correct topt can be used by some loss functions
const TargetPhrase &correctPhrase = translationOptionList.Get(firstCorrect)->GetTargetPhrase();
// extract source side features
for(size_t i = 0; i < sourceFeatures.size(); ++i)
(*sourceFeatures[i])(input, inputPath, sourceRange, classifier);
// go over topts, extract target side features and train the classifier
for (size_t toptIdx = 0; toptIdx < translationOptionList.size(); toptIdx++) {
// this topt was discarded by leaving one out
if (! keep[toptIdx])
continue;
// extract target-side features for each topt
const TargetPhrase &targetPhrase = translationOptionList.Get(toptIdx)->GetTargetPhrase();
for(size_t i = 0; i < targetFeatures.size(); ++i)
(*targetFeatures[i])(input, inputPath, targetPhrase, classifier);
float loss = (*m_trainingLoss)(targetPhrase, correctPhrase, correct[toptIdx]);
// train classifier on current example
classifier.Train(MakeTargetLabel(targetPhrase), loss);
}
} else {
//
// predict using a trained classifier, use this in decoding (=at test time)
//
std::vector<float> losses(translationOptionList.size());
// extract source side features
for(size_t i = 0; i < sourceFeatures.size(); ++i)
(*sourceFeatures[i])(input, inputPath, sourceRange, classifier);
for (size_t toptIdx = 0; toptIdx < translationOptionList.size(); toptIdx++) {
const TranslationOption *topt = translationOptionList.Get(toptIdx);
const TargetPhrase &targetPhrase = topt->GetTargetPhrase();
// extract target-side features for each topt
for(size_t i = 0; i < targetFeatures.size(); ++i)
(*targetFeatures[i])(input, inputPath, targetPhrase, classifier);
// get classifier score
losses[toptIdx] = classifier.Predict(MakeTargetLabel(targetPhrase));
}
// normalize classifier scores to get a probability distribution
(*m_normalizer)(losses);
// update scores of topts
for (size_t toptIdx = 0; toptIdx < translationOptionList.size(); toptIdx++) {
TranslationOption *topt = *(translationOptionList.begin() + toptIdx);
std::vector<float> newScores(m_numScoreComponents);
newScores[0] = FloorScore(TransformScore(losses[toptIdx]));
ScoreComponentCollection &scoreBreakDown = topt->GetScoreBreakdown();
scoreBreakDown.PlusEquals(this, newScores);
topt->UpdateScore();
}
}
}
void EvaluateWhenApplied(const Hypothesis& hypo,
virtual FFState* EvaluateWhenApplied(
const ChartHypothesis&,
int,
ScoreComponentCollection* accumulator) const {
throw new std::logic_error("hiearchical/syntax not supported");
}
void EvaluateWhenApplied(const ChartHypothesis &hypo,
ScoreComponentCollection* accumulator) const {
}
// Initial VW state; contains unaligned BOS symbols.
const FFState* EmptyHypothesisState(const InputType &input) const;
void SetParameter(const std::string& key, const std::string& value) {
if (key == "train") {
m_train = Scan<bool>(value);
} else if (key == "path") {
m_modelPath = value;
} else if (key == "vw-options") {
m_vwOptions = value;
} else if (key == "leave-one-out-from") {
m_leaveOneOut = value;
} else if (key == "training-loss") {
// which type of loss to use for training
if (value == "basic") {
m_trainingLoss = (TrainingLoss *) new TrainingLossBasic();
} else if (value == "bleu") {
m_trainingLoss = (TrainingLoss *) new TrainingLossBLEU();
} else {
UTIL_THROW2("Unknown training loss type:" << value);
}
} else if (key == "loss") {
// which normalizer to use (theoretically depends on the loss function used for training the
// classifier (squared/logistic/hinge/...), hence the name "loss"
if (value == "logistic") {
m_normalizer = (Discriminative::Normalizer *) new Discriminative::LogisticLossNormalizer();
} else if (value == "squared") {
m_normalizer = (Discriminative::Normalizer *) new Discriminative::SquaredLossNormalizer();
} else {
UTIL_THROW2("Unknown loss type:" << value);
}
} else {
StatelessFeatureFunction::SetParameter(key, value);
}
}
virtual void InitializeForInput(ttasksptr const& ttask) {
InputType const& source = *(ttask->GetSource().get());
// tabbed sentence is assumed only in training
if (! m_train)
return;
UTIL_THROW_IF2(source.GetType() != TabbedSentenceInput,
"This feature function requires the TabbedSentence input type");
const TabbedSentence& tabbedSentence = static_cast<const TabbedSentence&>(source);
UTIL_THROW_IF2(tabbedSentence.GetColumns().size() < 2,
"TabbedSentence must contain target<tab>alignment");
// target sentence represented as a phrase
Phrase *target = new Phrase();
target->CreateFromString(
Output
, StaticData::Instance().options()->output.factor_order
, tabbedSentence.GetColumns()[0]
, NULL);
// word alignment between source and target sentence
// we don't store alignment info in AlignmentInfoCollection because we keep alignments of whole
// sentences, not phrases
AlignmentInfo *alignment = new AlignmentInfo(tabbedSentence.GetColumns()[1]);
VWTargetSentence &targetSent = *GetStored();
targetSent.Clear();
targetSent.m_sentence = target;
targetSent.m_alignment = alignment;
// pre-compute max- and min- aligned points for faster translation option checking
targetSent.SetConstraints(source.GetSize());
}
void SetParameter(const std::string& key, const std::string& value);
// At prediction time, this clears our caches. At training time, we load the next sentence, its
// translation and word alignment.
virtual void InitializeForInput(ttasksptr const& ttask);
private:
std::string MakeTargetLabel(const TargetPhrase &targetPhrase) const {
return VW_DUMMY_LABEL;
inline std::string MakeTargetLabel(const TargetPhrase &targetPhrase) const {
return VW_DUMMY_LABEL; // VW does not care about class labels in our setting (--csoaa_ldf mc).
}
bool IsCorrectTranslationOption(const TranslationOption &topt) const {
//std::cerr << topt.GetSourceWordsRange() << std::endl;
int sourceStart = topt.GetSourceWordsRange().GetStartPos();
int sourceEnd = topt.GetSourceWordsRange().GetEndPos();
const VWTargetSentence &targetSentence = *GetStored();
// [targetStart, targetEnd] spans aligned target words
int targetStart = targetSentence.m_sentence->GetSize();
int targetEnd = -1;
// get the left-most and right-most alignment point within source span
for(int i = sourceStart; i <= sourceEnd; ++i) {
if(targetSentence.m_sourceConstraints[i].IsSet()) {
if(targetStart > targetSentence.m_sourceConstraints[i].GetMin())
targetStart = targetSentence.m_sourceConstraints[i].GetMin();
if(targetEnd < targetSentence.m_sourceConstraints[i].GetMax())
targetEnd = targetSentence.m_sourceConstraints[i].GetMax();
}
}
// there was no alignment
if(targetEnd == -1)
return false;
//std::cerr << "Shorter: " << targetStart << " " << targetEnd << std::endl;
// [targetStart2, targetEnd2] spans unaligned words left and right of [targetStart, targetEnd]
int targetStart2 = targetStart;
for(int i = targetStart2; i >= 0 && !targetSentence.m_targetConstraints[i].IsSet(); --i)
targetStart2 = i;
int targetEnd2 = targetEnd;
for(int i = targetEnd2;
i < targetSentence.m_sentence->GetSize() && !targetSentence.m_targetConstraints[i].IsSet();
++i)
targetEnd2 = i;
//std::cerr << "Longer: " << targetStart2 << " " << targetEnd2 << std::endl;
const TargetPhrase &tphrase = topt.GetTargetPhrase();
//std::cerr << tphrase << std::endl;
// if target phrase is shorter than inner span return false
if(tphrase.GetSize() < targetEnd - targetStart + 1)
return false;
// if target phrase is longer than outer span return false
if(tphrase.GetSize() > targetEnd2 - targetStart2 + 1)
return false;
// for each possible starting point
for(int tempStart = targetStart2; tempStart <= targetStart; tempStart++) {
bool found = true;
// check if the target phrase is within longer span
for(int i = tempStart; i <= targetEnd2 && i < tphrase.GetSize() + tempStart; ++i) {
if(tphrase.GetWord(i - tempStart) != targetSentence.m_sentence->GetWord(i)) {
found = false;
break;
}
}
// return true if there was a match
if(found) {
//std::cerr << "Found" << std::endl;
return true;
}
inline size_t MakeCacheKey(const FFState *prevState, size_t spanStart, size_t spanEnd) const {
size_t key = 0;
boost::hash_combine(key, prevState);
boost::hash_combine(key, spanStart);
boost::hash_combine(key, spanEnd);
return key;
}
return false;
}
// used in decoding to transform the global word alignment information into
// context-phrase internal alignment information (i.e., with target indices correspoding
// to positions in contextPhrase)
const AlignmentInfo *TransformAlignmentInfo(const Hypothesis &curHypo, size_t contextSize) const;
std::vector<bool> LeaveOneOut(const TranslationOptionList &topts, const std::vector<bool> &correct) const {
UTIL_THROW_IF2(m_leaveOneOut.size() == 0 || ! m_train, "LeaveOneOut called in wrong setting!");
// used during training to extract relevant alignment points from the full sentence alignment
// and shift them by target context size
AlignmentInfo TransformAlignmentInfo(const AlignmentInfo &alignInfo, size_t contextSize, int currentStart) const;
float sourceRawCount = 0.0;
const float ONE = 1.0001; // I don't understand floating point numbers
// At training time, determine whether a translation option is correct for the current target sentence
// based on word alignment. This is a bit complicated because we need to handle various corner-cases
// where some word(s) on phrase borders are unaligned.
std::pair<bool, int> IsCorrectTranslationOption(const TranslationOption &topt) const;
std::vector<bool> keepOpt;
for (size_t i = 0; i < topts.size(); i++) {
TranslationOption *topt = *(topts.begin() + i);
const TargetPhrase &targetPhrase = topt->GetTargetPhrase();
// extract raw counts from phrase-table property
const CountsPhraseProperty *property =
static_cast<const CountsPhraseProperty *>(targetPhrase.GetProperty("Counts"));
if (! property) {
VERBOSE(1, "VW :: Counts not found for topt! Is this an OOV?\n");
// keep all translation opts without updating, this is either OOV or bad usage...
keepOpt.assign(topts.size(), true);
return keepOpt;
}
if (sourceRawCount == 0.0) {
sourceRawCount = property->GetSourceMarginal() - ONE; // discount one occurrence of the source phrase
if (sourceRawCount <= 0) {
// no translation options survived, source phrase was a singleton
keepOpt.assign(topts.size(), false);
return keepOpt;
}
}
float discount = correct[i] ? ONE : 0.0;
float target = property->GetTargetMarginal() - discount;
float joint = property->GetJointCount() - discount;
if (discount != 0.0) VERBOSE(2, "VW :: leaving one out!\n");
if (joint > 0) {
// topt survived leaving one out, update its scores
const FeatureFunction *feature = &FindFeatureFunction(m_leaveOneOut);
std::vector<float> scores = targetPhrase.GetScoreBreakdown().GetScoresForProducer(feature);
UTIL_THROW_IF2(scores.size() != 4, "Unexpected number of scores in feature " << m_leaveOneOut);
scores[0] = TransformScore(joint / target); // P(f|e)
scores[2] = TransformScore(joint / sourceRawCount); // P(e|f)
ScoreComponentCollection &scoreBreakDown = topt->GetScoreBreakdown();
scoreBreakDown.Assign(feature, scores);
topt->UpdateScore();
keepOpt.push_back(true);
} else {
// they only occurred together once, discard topt
VERBOSE(2, "VW :: discarded topt when leaving one out\n");
keepOpt.push_back(false);
}
}
return keepOpt;
}
// At training time, optionally discount occurrences of phrase pairs from the current sentence, helps prevent
// over-fitting.
std::vector<bool> LeaveOneOut(const TranslationOptionList &topts, const std::vector<bool> &correct) const;
bool m_train; // false means predict
std::string m_modelPath;
std::string m_vwOptions;
std::string m_modelPath; // path to the VW model file; at training time, this is where extracted features are stored
std::string m_vwOptions; // options for Vowpal Wabbit
// BOS token, all factors
Word m_sentenceStartWord;
// calculator of training loss
TrainingLoss *m_trainingLoss = NULL;
@ -488,9 +179,16 @@ private:
// optionally contains feature name of a phrase table where we recompute scores with leaving one out
std::string m_leaveOneOut;
// normalizer, typically this means softmax
Discriminative::Normalizer *m_normalizer = NULL;
// thread-specific classifier instance
TLSClassifier *m_tlsClassifier;
// caches for partial scores and feature vectors
TLSFloatHashMap *m_tlsFutureScores;
TLSStateExtensions *m_tlsComputedStateExtensions;
TLSFeatureVectorMap *m_tlsTranslationOptionFeatures, *m_tlsTargetContextFeatures;
};
}

15
moses/FF/VW/VWFeatureBase.cpp
View File

@ -2,11 +2,26 @@
#include <string>
#include "VWFeatureBase.h"
#include "VWFeatureContext.h"
namespace Moses
{
std::map<std::string, std::vector<VWFeatureBase*> > VWFeatureBase::s_features;
std::map<std::string, std::vector<VWFeatureBase*> > VWFeatureBase::s_sourceFeatures;
std::map<std::string, std::vector<VWFeatureBase*> > VWFeatureBase::s_targetContextFeatures;
std::map<std::string, std::vector<VWFeatureBase*> > VWFeatureBase::s_targetFeatures;
std::map<std::string, size_t> VWFeatureBase::s_targetContextLength;
void VWFeatureBase::UpdateContextSize(const std::string &usedBy)
{
// using the standard map behavior here: if the entry does not
// exist, it will be added and initialized to zero
size_t currentSize = s_targetContextLength[usedBy];
size_t newSize = static_cast<VWFeatureContext *const>(this)->GetContextSize();
s_targetContextLength[usedBy] = std::max(currentSize, newSize);
}
}

57
moses/FF/VW/VWFeatureBase.h
View File

@ -12,11 +12,17 @@
namespace Moses
{
enum VWFeatureType {
vwft_source,
vwft_target,
vwft_targetContext
};
class VWFeatureBase : public StatelessFeatureFunction
{
public:
VWFeatureBase(const std::string &line, bool isSource = true)
: StatelessFeatureFunction(0, line), m_usedBy(1, "VW0"), m_isSource(isSource) {
VWFeatureBase(const std::string &line, VWFeatureType featureType = vwft_source)
: StatelessFeatureFunction(0, line), m_usedBy(1, "VW0"), m_featureType(featureType) {
// defaults
m_sourceFactors.push_back(0);
m_targetFactors.push_back(0);
@ -71,26 +77,47 @@ public:
return s_sourceFeatures[name];
}
// Return only target-context classifier features
static const std::vector<VWFeatureBase*>& GetTargetContextFeatures(std::string name = "VW0") {
// don't throw an exception when there are no target-context features, this feature type is not mandatory
return s_targetContextFeatures[name];
}
// Return only target-dependent classifier features
static const std::vector<VWFeatureBase*>& GetTargetFeatures(std::string name = "VW0") {
UTIL_THROW_IF2(s_targetFeatures.count(name) == 0, "No target features registered for parent classifier: " + name);
return s_targetFeatures[name];
}
// Required length context (maximum context size of defined target-context features)
static size_t GetMaximumContextSize(std::string name = "VW0") {
return s_targetContextLength[name]; // 0 by default
}
// Overload to process source-dependent data, create features once for every
// source sentence word range.
virtual void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const = 0;
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const = 0;
// Overload to process target-dependent features, create features once for
// every target phrase. One source word range will have at leat one target
// every target phrase. One source word range will have at least one target
// phrase, but may have more.
virtual void operator()(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, Discriminative::Classifier &classifier) const = 0;
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const = 0;
// Overload to process target-context dependent features, these features are
// evaluated during decoding. For efficiency, features are not fed directly into
// the classifier object but instead output in the vector "features" and managed
// separately in VW.h.
virtual void operator()(const InputType &input
, const Phrase &contextPhrase
, const AlignmentInfo &alignmentInfo
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const = 0;
protected:
std::vector<FactorType> m_sourceFactors, m_targetFactors;
@ -99,12 +126,17 @@ protected:
for(std::vector<std::string>::const_iterator it = m_usedBy.begin();
it != m_usedBy.end(); it++) {
s_features[*it].push_back(this);
if(m_isSource)
if(m_featureType == vwft_source) {
s_sourceFeatures[*it].push_back(this);
else
} else if (m_featureType == vwft_targetContext) {
s_targetContextFeatures[*it].push_back(this);
UpdateContextSize(*it);
} else {
s_targetFeatures[*it].push_back(this);
}
}
}
private:
void ParseUsedBy(const std::string &usedBy) {
@ -112,11 +144,16 @@ private:
Tokenize(m_usedBy, usedBy, ",");
}
void UpdateContextSize(const std::string &usedBy);
std::vector<std::string> m_usedBy;
bool m_isSource;
VWFeatureType m_featureType;
static std::map<std::string, std::vector<VWFeatureBase*> > s_features;
static std::map<std::string, std::vector<VWFeatureBase*> > s_sourceFeatures;
static std::map<std::string, std::vector<VWFeatureBase*> > s_targetContextFeatures;
static std::map<std::string, std::vector<VWFeatureBase*> > s_targetFeatures;
static std::map<std::string, size_t> s_targetContextLength;
};
}

116
moses/FF/VW/VWFeatureContext.h Normal file
View File

@ -0,0 +1,116 @@
#pragma once
#include <string>
#include <boost/foreach.hpp>
#include "VWFeatureBase.h"
#include "moses/InputType.h"
#include "moses/TypeDef.h"
#include "moses/Word.h"
namespace Moses
{
// Inherit from this for source-dependent classifier features. They will
// automatically register with the classifier class named VW0 or one or more
// names specified by the used-by=name1,name2,... parameter.
//
// The classifier gets a full list by calling
// VWFeatureBase::GetTargetContextFeatures(GetScoreProducerDescription())
class VWFeatureContext : public VWFeatureBase
{
public:
VWFeatureContext(const std::string &line, size_t contextSize)
: VWFeatureBase(line, vwft_targetContext), m_contextSize(contextSize) {
}
// Gets its pure virtual functions from VWFeatureBase
virtual void operator()(const InputType &input
, const TargetPhrase &targetPhrase
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
}
virtual void operator()(const InputType &input
, const Range &sourceRange
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
}
virtual void SetParameter(const std::string& key, const std::string& value) {
if (key == "size") {
m_contextSize = Scan<size_t>(value);
} else if (key == "factor-positions") {
// factor positions: assuming a factor such as positional morphological tag, use this
// option to select only certain positions; this assumes that only a single
// target-side factor is defined
Tokenize<size_t>(m_factorPositions, value, ",");
} else {
VWFeatureBase::SetParameter(key, value);
}
}
size_t GetContextSize() {
return m_contextSize;
}
protected:
// Get word with the correct subset of factors as string. Because we're target
// context features, we look at a limited number of words to the left of the
// current translation. posFromEnd is interpreted like this:
// 0 = last word of the hypothesis
// 1 = next to last word
// ...etc.
inline std::string GetWord(const Phrase &phrase, size_t posFromEnd) const {
const Word &word = phrase.GetWord(phrase.GetSize() - posFromEnd - 1);
if (m_factorPositions.empty()) {
return word.GetString(m_targetFactors, false);
} else {
if (m_targetFactors.size() != 1)
UTIL_THROW2("You can only use factor-positions when a single target-side factor is defined.");
const std::string &fullFactor = word.GetFactor(m_targetFactors[0])->GetString().as_string();
// corner cases: at sentence beginning/end, we don't have the correct factors set up
// similarly for UNK
if (fullFactor == BOS_ || fullFactor == EOS_ || fullFactor == UNKNOWN_FACTOR)
return fullFactor;
std::string subFactor(m_factorPositions.size(), 'x'); // initialize string with correct size and placeholder chars
for (size_t i = 0; i < m_factorPositions.size(); i++)
subFactor[i] = fullFactor[m_factorPositions[i]];
return subFactor;
}
}
// some target-context feature functions also look at the source
inline std::string GetSourceWord(const InputType &input, size_t pos) const {
return input.GetWord(pos).GetString(m_sourceFactors, false);
}
// get source words aligned to a particular context word
std::vector<std::string> GetAlignedSourceWords(const Phrase &contextPhrase
, const InputType &input
, const AlignmentInfo &alignInfo
, size_t posFromEnd) const {
size_t idx = contextPhrase.GetSize() - posFromEnd - 1;
std::set<size_t> alignedToTarget = alignInfo.GetAlignmentsForTarget(idx);
std::vector<std::string> out;
out.reserve(alignedToTarget.size());
BOOST_FOREACH(size_t srcIdx, alignedToTarget) {
out.push_back(GetSourceWord(input, srcIdx));
}
return out;
}
// required context size
size_t m_contextSize;
// factor positions: assuming a factor such as positional morphological tag, use this
// option to select only certain positions
std::vector<size_t> m_factorPositions;
};
}

40
moses/FF/VW/VWFeatureContextBigrams.h Normal file
View File

@ -0,0 +1,40 @@
#pragma once
#include <string>
#include <algorithm>
#include "VWFeatureContext.h"
#include "moses/Util.h"
namespace Moses
{
class VWFeatureContextBigrams : public VWFeatureContext
{
public:
VWFeatureContextBigrams(const std::string &line)
: VWFeatureContext(line, DEFAULT_WINDOW_SIZE) {
ReadParameters();
// Call this last
VWFeatureBase::UpdateRegister();
}
virtual void operator()(const InputType &input
, const Phrase &contextPhrase
, const AlignmentInfo &alignmentInfo
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
for (size_t i = 1; i < m_contextSize; i++)
outFeatures.push_back(classifier.AddLabelIndependentFeature("tcbigram^-" + SPrint(i + 1)
+ "^" + GetWord(contextPhrase, i - 1) + "^" + GetWord(contextPhrase, i)));
}
virtual void SetParameter(const std::string& key, const std::string& value) {
VWFeatureContext::SetParameter(key, value);
}
private:
static const int DEFAULT_WINDOW_SIZE = 1;
};
}

45
moses/FF/VW/VWFeatureContextBilingual.h Normal file
View File

@ -0,0 +1,45 @@
#pragma once
#include <string>
#include <boost/foreach.hpp>
#include <algorithm>
#include "VWFeatureContext.h"
#include "moses/Util.h"
namespace Moses
{
class VWFeatureContextBilingual : public VWFeatureContext
{
public:
VWFeatureContextBilingual(const std::string &line)
: VWFeatureContext(line, DEFAULT_WINDOW_SIZE) {
ReadParameters();
// Call this last
VWFeatureBase::UpdateRegister();
}
virtual void operator()(const InputType &input
, const Phrase &contextPhrase
, const AlignmentInfo &alignmentInfo
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
for (size_t i = 0; i < m_contextSize; i++) {
std::string tgtWord = GetWord(contextPhrase, i);
std::vector<std::string> alignedTo = GetAlignedSourceWords(contextPhrase, input, alignmentInfo, i);
BOOST_FOREACH(const std::string &srcWord, alignedTo) {
outFeatures.push_back(classifier.AddLabelIndependentFeature("tcblng^-" + SPrint(i + 1) + "^" + tgtWord + "^" + srcWord));
}
}
}
virtual void SetParameter(const std::string& key, const std::string& value) {
VWFeatureContext::SetParameter(key, value);
}
private:
static const int DEFAULT_WINDOW_SIZE = 1;
};
}

39
moses/FF/VW/VWFeatureContextWindow.h Normal file
View File

@ -0,0 +1,39 @@
#pragma once
#include <string>
#include <algorithm>
#include "VWFeatureContext.h"
#include "moses/Util.h"
namespace Moses
{
class VWFeatureContextWindow : public VWFeatureContext
{
public:
VWFeatureContextWindow(const std::string &line)
: VWFeatureContext(line, DEFAULT_WINDOW_SIZE) {
ReadParameters();
// Call this last
VWFeatureBase::UpdateRegister();
}
virtual void operator()(const InputType &input
, const Phrase &contextPhrase
, const AlignmentInfo &alignmentInfo
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
for (size_t i = 0; i < m_contextSize; i++)
outFeatures.push_back(classifier.AddLabelIndependentFeature("tcwin^-" + SPrint(i + 1) + "^" + GetWord(contextPhrase, i)));
}
virtual void SetParameter(const std::string& key, const std::string& value) {
VWFeatureContext::SetParameter(key, value);
}
private:
static const int DEFAULT_WINDOW_SIZE = 1;
};
}

13
moses/FF/VW/VWFeatureSource.h
View File

@ -19,15 +19,22 @@ class VWFeatureSource : public VWFeatureBase
{
public:
VWFeatureSource(const std::string &line)
: VWFeatureBase(line, true) {
: VWFeatureBase(line, vwft_source) {
}
// Gets its pure virtual functions from VWFeatureBase
virtual void operator()(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
}
virtual void operator()(const InputType &input
, const Phrase &contextPhrase
, const AlignmentInfo &alignmentInfo
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
}
virtual void SetParameter(const std::string& key, const std::string& value) {

6
moses/FF/VW/VWFeatureSourceBagOfWords.h
View File

@ -18,11 +18,11 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
for (size_t i = 0; i < input.GetSize(); i++) {
classifier.AddLabelIndependentFeature("bow^" + GetWord(input, i));
outFeatures.push_back(classifier.AddLabelIndependentFeature("bow^" + GetWord(input, i)));
}
}

6
moses/FF/VW/VWFeatureSourceBigrams.h
View File

@ -18,11 +18,11 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
for (size_t i = 1; i < input.GetSize(); i++) {
classifier.AddLabelIndependentFeature("bigram^" + GetWord(input, i - 1) + "^" + GetWord(input, i));
outFeatures.push_back(classifier.AddLabelIndependentFeature("bigram^" + GetWord(input, i - 1) + "^" + GetWord(input, i)));
}
}

6
moses/FF/VW/VWFeatureSourceExternalFeatures.h
View File

@ -23,12 +23,12 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
const Features& features = *m_tls.GetStored();
for (size_t i = 0; i < features.size(); i++) {
classifier.AddLabelIndependentFeature("srcext^" + features[i]);
outFeatures.push_back(classifier.AddLabelIndependentFeature("srcext^" + features[i]));
}
}

6
moses/FF/VW/VWFeatureSourceIndicator.h
View File

@ -20,9 +20,9 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
size_t begin = sourceRange.GetStartPos();
size_t end = sourceRange.GetEndPos() + 1;
@ -31,7 +31,7 @@ public:
for (size_t i = 0; i < end - begin; i++)
words[i] = GetWord(input, begin + i);
classifier.AddLabelIndependentFeature("sind^" + Join(" ", words));
outFeatures.push_back(classifier.AddLabelIndependentFeature("sind^" + Join(" ", words)));
}
virtual void SetParameter(const std::string& key, const std::string& value) {

6
moses/FF/VW/VWFeatureSourcePhraseInternal.h
View File

@ -20,14 +20,14 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
size_t begin = sourceRange.GetStartPos();
size_t end = sourceRange.GetEndPos() + 1;
while (begin < end) {
classifier.AddLabelIndependentFeature("sin^" + GetWord(input, begin++));
outFeatures.push_back(classifier.AddLabelIndependentFeature("sin^" + GetWord(input, begin++)));
}
}

16
moses/FF/VW/VWFeatureSourceSenseWindow.h
View File

@ -51,9 +51,9 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
int begin = sourceRange.GetStartPos();
int end = sourceRange.GetEndPos() + 1;
int inputLen = input.GetSize();
@ -64,24 +64,24 @@ public:
// before current phrase
for (int i = std::max(0, begin - m_size); i < begin; i++) {
BOOST_FOREACH(const Sense &sense, senses[i]) {
classifier.AddLabelIndependentFeature("snsb^" + forms[i] + SPrint(i - begin) + "^" + sense.m_label, sense.m_prob);
classifier.AddLabelIndependentFeature("snsb^" + forms[i] + sense.m_label, sense.m_prob);
outFeatures.push_back(classifier.AddLabelIndependentFeature("snsb^" + forms[i] + SPrint(i - begin) + "^" + sense.m_label, sense.m_prob));
outFeatures.push_back(classifier.AddLabelIndependentFeature("snsb^" + forms[i] + sense.m_label, sense.m_prob));
}
}
// within current phrase
for (int i = begin; i < end; i++) {
BOOST_FOREACH(const Sense &sense, senses[i]) {
classifier.AddLabelIndependentFeature("snsin^" + forms[i] + SPrint(i - begin) + "^" + sense.m_label, sense.m_prob);
classifier.AddLabelIndependentFeature("snsin^" + forms[i] + sense.m_label, sense.m_prob);
outFeatures.push_back(classifier.AddLabelIndependentFeature("snsin^" + forms[i] + SPrint(i - begin) + "^" + sense.m_label, sense.m_prob));
outFeatures.push_back(classifier.AddLabelIndependentFeature("snsin^" + forms[i] + sense.m_label, sense.m_prob));
}
}
// after current phrase
for (int i = end; i < std::min(end + m_size, inputLen); i++) {
BOOST_FOREACH(const Sense &sense, senses[i]) {
classifier.AddLabelIndependentFeature("snsa^" + forms[i] + SPrint(i - begin) + "^" + sense.m_label, sense.m_prob);
classifier.AddLabelIndependentFeature("snsa^" + forms[i] + sense.m_label, sense.m_prob);
outFeatures.push_back(classifier.AddLabelIndependentFeature("snsa^" + forms[i] + SPrint(i - begin) + "^" + sense.m_label, sense.m_prob));
outFeatures.push_back(classifier.AddLabelIndependentFeature("snsa^" + forms[i] + sense.m_label, sense.m_prob));
}
}
}

8
moses/FF/VW/VWFeatureSourceWindow.h
View File

@ -20,19 +20,19 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
int begin = sourceRange.GetStartPos();
int end = sourceRange.GetEndPos() + 1;
int inputLen = input.GetSize();
for (int i = std::max(0, begin - m_size); i < begin; i++) {
classifier.AddLabelIndependentFeature("c^" + SPrint(i - begin) + "^" + GetWord(input, i));
outFeatures.push_back(classifier.AddLabelIndependentFeature("c^" + SPrint(i - begin) + "^" + GetWord(input, i)));
}
for (int i = end; i < std::min(end + m_size, inputLen); i++) {
classifier.AddLabelIndependentFeature("c^" + SPrint(i - end + 1) + "^" + GetWord(input, i));
outFeatures.push_back(classifier.AddLabelIndependentFeature("c^" + SPrint(i - end + 1) + "^" + GetWord(input, i)));
}
}

13
moses/FF/VW/VWFeatureTarget.h
View File

@ -17,15 +17,22 @@ class VWFeatureTarget : public VWFeatureBase
{
public:
VWFeatureTarget(const std::string &line)
: VWFeatureBase(line, false) {
: VWFeatureBase(line, vwft_target) {
}
// Gets its pure virtual functions from VWFeatureBase
virtual void operator()(const InputType &input
, const InputPath &inputPath
, const Range &sourceRange
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
}
virtual void operator()(const InputType &input
, const Phrase &contextPhrase
, const AlignmentInfo &alignmentInfo
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
}
virtual void SetParameter(const std::string& key, const std::string& value) {

6
moses/FF/VW/VWFeatureTargetBigrams.h
View File

@ -17,11 +17,11 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
for (size_t i = 1; i < targetPhrase.GetSize(); i++) {
classifier.AddLabelDependentFeature("tbigram^" + GetWord(targetPhrase, i - 1) + "^" + GetWord(targetPhrase, i));
outFeatures.push_back(classifier.AddLabelDependentFeature("tbigram^" + GetWord(targetPhrase, i - 1) + "^" + GetWord(targetPhrase, i)));
}
}

6
moses/FF/VW/VWFeatureTargetIndicator.h
View File

@ -17,10 +17,10 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, Discriminative::Classifier &classifier) const {
classifier.AddLabelDependentFeature("tind^" + targetPhrase.GetStringRep(m_targetFactors));
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
outFeatures.push_back(classifier.AddLabelDependentFeature("tind^" + targetPhrase.GetStringRep(m_targetFactors)));
}
virtual void SetParameter(const std::string& key, const std::string& value) {

6
moses/FF/VW/VWFeatureTargetPhraseInternal.h
View File

@ -17,11 +17,11 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
for (size_t i = 0; i < targetPhrase.GetSize(); i++) {
classifier.AddLabelDependentFeature("tin^" + GetWord(targetPhrase, i));
outFeatures.push_back(classifier.AddLabelDependentFeature("tin^" + GetWord(targetPhrase, i)));
}
}

6
moses/FF/VW/VWFeatureTargetPhraseScores.h
View File

@ -20,9 +20,9 @@ public:
}
void operator()(const InputType &input
, const InputPath &inputPath
, const TargetPhrase &targetPhrase
, Discriminative::Classifier &classifier) const {
, Discriminative::Classifier &classifier
, Discriminative::FeatureVector &outFeatures) const {
std::vector<FeatureFunction*> features = FeatureFunction::GetFeatureFunctions();
for (size_t i = 0; i < features.size(); i++) {
std::string fname = features[i]->GetScoreProducerDescription();
@ -31,7 +31,7 @@ public:
std::vector<float> scores = targetPhrase.GetScoreBreakdown().GetScoresForProducer(features[i]);
for(size_t j = 0; j < scores.size(); ++j)
classifier.AddLabelDependentFeature(fname + "^" + boost::lexical_cast<std::string>(j), scores[j]);
outFeatures.push_back(classifier.AddLabelDependentFeature(fname + "^" + boost::lexical_cast<std::string>(j), scores[j]));
}
}

77
moses/FF/VW/VWState.cpp Normal file
View File

@ -0,0 +1,77 @@
#include "VWState.h"
#include "moses/FF/FFState.h"
#include "moses/Phrase.h"
#include "moses/Hypothesis.h"
#include "moses/Util.h"
#include "moses/TypeDef.h"
#include "moses/StaticData.h"
#include "moses/TranslationOption.h"
#include <boost/functional/hash.hpp>
namespace Moses
{
VWState::VWState() : m_spanStart(0), m_spanEnd(0)
{
ComputeHash();
}
VWState::VWState(const Phrase &phrase)
: m_phrase(phrase), m_spanStart(0), m_spanEnd(0)
{
ComputeHash();
}
VWState::VWState(const VWState &prevState, const Hypothesis &curHypo)
{
VERBOSE(3, "VW :: updating state\n>> previous state: " << prevState << "\n");
// copy phrase from previous state
Phrase phrase = prevState.GetPhrase();
size_t contextSize = phrase.GetSize(); // identical to VWFeatureBase::GetMaximumContextSize()
// add words from current hypothesis
phrase.Append(curHypo.GetCurrTargetPhrase());
VERBOSE(3, ">> current hypo: " << curHypo.GetCurrTargetPhrase() << "\n");
// get a slice of appropriate length
Range range(phrase.GetSize() - contextSize, phrase.GetSize() - 1);
m_phrase = phrase.GetSubString(range);
// set current span start/end
m_spanStart = curHypo.GetTranslationOption().GetStartPos();
m_spanEnd = curHypo.GetTranslationOption().GetEndPos();
// compute our hash
ComputeHash();
VERBOSE(3, ">> updated state: " << *this << "\n");
}
bool VWState::operator==(const FFState& o) const
{
const VWState &other = static_cast<const VWState &>(o);
return m_phrase == other.GetPhrase()
&& m_spanStart == other.GetSpanStart()
&& m_spanEnd == other.GetSpanEnd();
}
void VWState::ComputeHash()
{
m_hash = 0;
boost::hash_combine(m_hash, m_phrase);
boost::hash_combine(m_hash, m_spanStart);
boost::hash_combine(m_hash, m_spanEnd);
}
std::ostream &operator<<(std::ostream &out, const VWState &state)
{
out << state.GetPhrase() << "::" << state.GetSpanStart() << "-" << state.GetSpanEnd();
return out;
}
}

56
moses/FF/VW/VWState.h Normal file
View File

@ -0,0 +1,56 @@
#pragma once
#include <ostream>
#include "moses/FF/FFState.h"
#include "moses/Phrase.h"
#include "moses/Hypothesis.h"
namespace Moses
{
/**
* VW state, used in decoding (when target context is enabled).
*/
class VWState : public FFState
{
public:
// empty state, used only when VWState is ignored
VWState();
// used for construction of the initial VW state
VWState(const Phrase &phrase);
// continue from previous VW state with a new hypothesis
VWState(const VWState &prevState, const Hypothesis &curHypo);
virtual bool operator==(const FFState& o) const;
inline virtual size_t hash() const {
return m_hash;
}
inline const Phrase &GetPhrase() const {
return m_phrase;
}
inline size_t GetSpanStart() const {
return m_spanStart;
}
inline size_t GetSpanEnd() const {
return m_spanEnd;
}
private:
void ComputeHash();
Phrase m_phrase;
size_t m_spanStart, m_spanEnd;
size_t m_hash;
};
// how to print a VW state
std::ostream &operator<<(std::ostream &out, const VWState &state);
}

55
moses/FF/VW/VWTargetSentence.h Normal file
View File

@ -0,0 +1,55 @@
#pragma once
#include <vector>
#include "moses/AlignmentInfo.h"
#include "moses/Phrase.h"
#include "AlignmentConstraint.h"
namespace Moses
{
/**
* VW thread-specific data about target sentence.
*/
class VWTargetSentence
{
public:
VWTargetSentence() : m_sentence(NULL), m_alignment(NULL) {}
void Clear() {
if (m_sentence) delete m_sentence;
if (m_alignment) delete m_alignment;
}
~VWTargetSentence() {
Clear();
}
void SetConstraints(size_t sourceSize) {
// initialize to unconstrained
m_sourceConstraints.assign(sourceSize, AlignmentConstraint());
m_targetConstraints.assign(m_sentence->GetSize(), AlignmentConstraint());
// set constraints according to alignment points
AlignmentInfo::const_iterator it;
for (it = m_alignment->begin(); it != m_alignment->end(); it++) {
int src = it->first;
int tgt = it->second;
if (src >= m_sourceConstraints.size() || tgt >= m_targetConstraints.size()) {
UTIL_THROW2("VW :: alignment point out of bounds: " << src << "-" << tgt);
}
m_sourceConstraints[src].Update(tgt);
m_targetConstraints[tgt].Update(src);
}
}
Phrase *m_sentence;
AlignmentInfo *m_alignment;
std::vector<AlignmentConstraint> m_sourceConstraints, m_targetConstraints;
};
}

1
moses/Parameter.cpp
View File

@ -59,6 +59,7 @@ Parameter::Parameter()
AddParam(main_opts,"version", "show version of Moses and libraries used");
AddParam(main_opts,"show-weights", "print feature weights and exit");
AddParam(main_opts,"time-out", "seconds after which is interrupted (-1=no time-out, default is -1)");
AddParam(main_opts,"segment-time-out", "seconds for single segment after which is interrupted (-1=no time-out, default is -1)");
///////////////////////////////////////////////////////////////////////////////////////
// factorization options

39
moses/ReorderingConstraint.cpp
View File

@ -54,8 +54,8 @@ void ReorderingConstraint::SetWall( size_t pos, bool value )
void ReorderingConstraint::FinalizeWalls()
{
for(size_t z = 0; z < m_zone.size(); z++ ) {
const size_t startZone = m_zone[z][0];
const size_t endZone = m_zone[z][1];// note: wall after endZone is not local
const size_t startZone = m_zone[z].first;
const size_t endZone = m_zone[z].second;// note: wall after endZone is not local
for( size_t pos = startZone; pos < endZone; pos++ ) {
if (m_wall[ pos ]) {
m_localWall[ pos ] = z;
@ -65,8 +65,8 @@ void ReorderingConstraint::FinalizeWalls()
// enforce that local walls only apply to innermost zone
else if (m_localWall[ pos ] != NOT_A_ZONE) {
size_t assigned_z = m_localWall[ pos ];
if ((m_zone[assigned_z][0] < startZone) ||
(m_zone[assigned_z][1] > endZone)) {
if ((m_zone[assigned_z].first < startZone) ||
(m_zone[assigned_z].second > endZone)) {
m_localWall[ pos ] = z;
}
}
@ -97,9 +97,9 @@ void ReorderingConstraint::SetMonotoneAtPunctuation( const Phrase &sentence )
void ReorderingConstraint::SetZone( size_t startPos, size_t endPos )
{
VERBOSE(3,"SETTING zone " << startPos << "-" << endPos << std::endl);
std::vector< size_t > newZone;
newZone.push_back( startPos );
newZone.push_back( endPos );
std::pair<size_t,size_t> newZone;
newZone.first = startPos;
newZone.second = endPos;
m_zone.push_back( newZone );
m_active = true;
}
@ -138,8 +138,8 @@ bool ReorderingConstraint::Check( const Bitmap &bitmap, size_t startPos, size_t
// check zones
for(size_t z = 0; z < m_zone.size(); z++ ) {
const size_t startZone = m_zone[z][0];
const size_t endZone = m_zone[z][1];
const size_t startZone = m_zone[z].first;
const size_t endZone = m_zone[z].second;
// fine, if translation has not reached zone yet and phrase outside zone
if (lastPos < startZone && ( endPos < startZone || startPos > endZone ) ) {
@ -236,4 +236,25 @@ bool ReorderingConstraint::Check( const Bitmap &bitmap, size_t startPos, size_t
return true;
}
std::ostream& operator<<(std::ostream& out, const ReorderingConstraint &obj)
{
out << "Zones:";
for (size_t i = 0; i < obj.m_zone.size(); ++i) {
const std::pair<size_t,size_t> &zone1 = obj.m_zone[i];
out << zone1.first << "-" << zone1.second << " ";
}
out << "Walls:";
for (size_t i = 0; i < obj.m_size; ++i) {
out << obj.m_wall[i];
}
out << " Local walls:";
for (size_t i = 0; i < obj.m_size; ++i) {
out << obj.m_localWall[i] << " ";
}
return out;
}
}

6
moses/ReorderingConstraint.h
View File

@ -45,13 +45,13 @@ class Bitmap;
*/
class ReorderingConstraint
{
friend std::ostream& operator<<(std::ostream& out, const ReorderingConstraint& reorderingConstraint);
friend std::ostream& operator<<(std::ostream& out, const ReorderingConstraint &obj);
protected:
// const size_t m_size; /**< number of words in sentence */
size_t m_size; /**< number of words in sentence */
bool *m_wall; /**< flag for each word if it is a wall */
size_t *m_localWall; /**< flag for each word if it is a local wall */
std::vector< std::vector< size_t > > m_zone; /** zones that limit reordering */
std::vector< std::pair<size_t,size_t> > m_zone; /** zones that limit reordering */
bool m_active; /**< flag indicating, if there are any active constraints */
int m_max_distortion;
public:
@ -93,7 +93,7 @@ public:
void SetZone( size_t startPos, size_t endPos );
//! returns the vector of zones
std::vector< std::vector< size_t > > & GetZones() {
std::vector< std::pair<size_t,size_t> > & GetZones() {
return m_zone;
}

19
moses/Search.cpp
View File

@ -17,21 +17,34 @@ Search::Search(Manager& manager)
, interrupted_flag(0)
{
m_initialTransOpt.SetInputPath(m_inputPath);
m_timer.start();
}
bool
Search::
out_of_time()
{
int const& timelimit = m_options.search.timeout;
if (!timelimit) return false;
if (timelimit > 0) {
double elapsed_time = GetUserTime();
if (elapsed_time <= timelimit) return false;
if (elapsed_time > timelimit) {
VERBOSE(1,"Decoding is out of time (" << elapsed_time << ","
<< timelimit << ")" << std::endl);
interrupted_flag = 1;
return true;
}
}
int const& segment_timelimit = m_options.search.segment_timeout;
if (segment_timelimit > 0) {
double elapsed_time = m_timer.get_elapsed_time();
if (elapsed_time > segment_timelimit) {
VERBOSE(1,"Decoding for segment is out of time (" << elapsed_time << ","
<< segment_timelimit << ")" << std::endl);
interrupted_flag = 1;
return true;
}
}
return false;
}
}

2
moses/Search.h
View File

@ -7,6 +7,7 @@
#include "Phrase.h"
#include "InputPath.h"
#include "Bitmaps.h"
#include "Timer.h"
namespace Moses
{
@ -48,6 +49,7 @@ protected:
/** flag indicating that decoder ran out of time (see switch -time-out) */
size_t interrupted_flag;
Timer m_timer;
bool out_of_time();
};

1
moses/SearchCubePruning.cpp
View File

@ -97,7 +97,6 @@ void SearchCubePruning::Decode()
// go through each stack
size_t stackNo = 1;
int timelimit = m_options.search.timeout;
std::vector < HypothesisStack* >::iterator iterStack;
for (iterStack = m_hypoStackColl.begin() + 1 ; iterStack != m_hypoStackColl.end() ; ++iterStack) {
// BOOST_FOREACH(HypothesisStack* hstack, m_hypoStackColl) {

4
moses/Sentence.cpp
View File

@ -155,7 +155,9 @@ aux_interpret_xml(std::string& line, std::vector<size_t> & xmlWalls,
m_xmlOptions,
m_reorderingConstraint,
xmlWalls, placeholders);
UTIL_THROW_IF2(!OK, "Unable to parse XML in line: " << line);
if (!OK) {
TRACE_ERR("Unable to parse XML in line: " << line);
}
}
}

5
moses/TranslationModel/CompactPT/CanonicalHuffman.h
View File

@ -76,8 +76,9 @@ private:
MinHeapSorter hs(A);
std::make_heap(A.begin(), A.begin() + n, hs);
size_t h = n;
size_t m1, m2;
// marked volatile to prevent the intel compiler from generating bad code
volatile size_t h = n;
volatile size_t m1, m2;
while(h > 1) {
m1 = A[0];
std::pop_heap(A.begin(), A.begin() + h, hs);

1
moses/parameters/SearchOptions.cpp
View File

@ -38,6 +38,7 @@ namespace Moses
param.SetParameter(early_discarding_threshold, "early-discarding-threshold",
DEFAULT_EARLY_DISCARDING_THRESHOLD);
param.SetParameter(timeout, "time-out", 0);
param.SetParameter(segment_timeout, "segment-time-out", 0);
param.SetParameter(max_phrase_length, "max-phrase-length",
DEFAULT_MAX_PHRASE_LENGTH);
param.SetParameter(trans_opt_threshold, "translation-option-threshold",

1
moses/parameters/SearchOptions.h
View File

@ -25,6 +25,7 @@ namespace Moses
float beam_width;
int timeout;
int segment_timeout;
bool consensus; //! Use Consensus decoding (DeNero et al 2009)

2
scripts/Transliteration/train-transliteration-module.pl
View File

@ -240,7 +240,7 @@ sub train_transliteration_module{
`$MOSES_SRC_DIR/scripts/ems/support/substitute-filtered-tables.perl $OUT_DIR/tuning/filtered/moses.ini < $OUT_DIR/model/moses.ini > $OUT_DIR/tuning/moses.filtered.ini`;
`$MOSES_SRC_DIR/scripts/training/mert-moses.pl $OUT_DIR/tuning/input $OUT_DIR/tuning/reference $DECODER $OUT_DIR/tuning/moses.filtered.ini --nbest 100 --working-dir $OUT_DIR/tuning/tmp --decoder-flags "-threads 16 -drop-unknown -v 0 -distortion-limit 0" --rootdir $MOSES_SRC_DIR/scripts -mertdir $MOSES_SRC_DIR/mert -threads=16 --no-filter-phrase-table`;
`$MOSES_SRC_DIR/scripts/training/mert-moses.pl $OUT_DIR/tuning/input $OUT_DIR/tuning/reference $DECODER $OUT_DIR/tuning/moses.filtered.ini --nbest 100 --working-dir $OUT_DIR/tuning/tmp --decoder-flags "-threads 16 -drop-unknown -v 0 -distortion-limit 0" --rootdir $MOSES_SRC_DIR/scripts -mertdir $MOSES_SRC_DIR/bin -threads=16 --no-filter-phrase-table`;
`cp $OUT_DIR/tuning/tmp/moses.ini $OUT_DIR/tuning/moses.ini`;

2
scripts/ems/example/config.basic
View File

@ -54,7 +54,7 @@ output-tokenizer = "$moses-script-dir/tokenizer/tokenizer.perl -a -l $output-ext
# For Arabic tokenizer try Farasa (download: http://qatsdemo.cloudapp.net/farasa/)
# Abdelali, Darwish, Durrani, Mubarak (NAACL demo 2016)
# "Farasa: A Fast and Furious Segmenter for Arabic"
input-tokenizer = "$farasa-dir/farasa_moses.sh"
#input-tokenizer = "$farasa-dir/farasa_moses.sh"
# truecasers - comment out if you do not use the truecaser

2
scripts/ems/example/config.factored
View File

@ -54,7 +54,7 @@ output-tokenizer = "$moses-script-dir/tokenizer/tokenizer.perl -a -l $output-ext
# For Arabic tokenizer try Farasa (download: http://qatsdemo.cloudapp.net/farasa/)
# Abdelali, Darwish, Durrani, Mubarak (NAACL demo 2016)
# "Farasa: A Fast and Furious Segmenter for Arabic"
input-tokenizer = "$farasa-dir/farasa_moses.sh"
#input-tokenizer = "$farasa-dir/farasa_moses.sh"
# truecasers - comment out if you do not use the truecaser
input-truecaser = $moses-script-dir/recaser/truecase.perl

2
scripts/ems/example/config.hierarchical
View File

@ -57,7 +57,7 @@ output-tokenizer = "$moses-script-dir/tokenizer/tokenizer.perl -a -l $output-ext
# For Arabic tokenizer try Farasa (download: http://qatsdemo.cloudapp.net/farasa/)
# Abdelali, Darwish, Durrani, Mubarak (NAACL demo 2016)
# "Farasa: A Fast and Furious Segmenter for Arabic"
input-tokenizer = "$farasa-dir/farasa_moses.sh"
#input-tokenizer = "$farasa-dir/farasa_moses.sh"
# truecasers - comment out if you do not use the truecaser
input-truecaser = $moses-script-dir/recaser/truecase.perl

2
scripts/ems/example/config.syntax
View File

@ -57,7 +57,7 @@ output-tokenizer = "$moses-script-dir/tokenizer/tokenizer.perl -a -l $output-ext
# For Arabic tokenizer try Farasa (download: http://qatsdemo.cloudapp.net/farasa/)
# Abdelali, Darwish, Durrani, Mubarak (NAACL demo 2016)
# "Farasa: A Fast and Furious Segmenter for Arabic"
input-tokenizer = "$farasa-dir/farasa_moses.sh"
#input-tokenizer = "$farasa-dir/farasa_moses.sh"
# truecasers - comment out if you do not use the truecaser
input-truecaser = $moses-script-dir/recaser/truecase.perl

2
scripts/ems/example/config.toy
View File

@ -54,7 +54,7 @@ output-tokenizer = "$moses-script-dir/tokenizer/tokenizer.perl -a -l $output-ext
# For Arabic tokenizer try Farasa (download: http://qatsdemo.cloudapp.net/farasa/)
# Abdelali, Darwish, Durrani, Mubarak (NAACL demo 2016)
# "Farasa: A Fast and Furious Segmenter for Arabic"
input-tokenizer = "$farasa-dir/farasa_moses.sh"
#input-tokenizer = "$farasa-dir/farasa_moses.sh"
# truecasers - comment out if you do not use the truecaser
input-truecaser = $moses-script-dir/recaser/truecase.perl

2
scripts/ems/example/config.toy.bilinguallm
View File

@ -54,7 +54,7 @@ output-tokenizer = "$moses-script-dir/tokenizer/tokenizer.perl -a -l $output-ext
# For Arabic tokenizer try Farasa (download: http://qatsdemo.cloudapp.net/farasa/)
# Abdelali, Darwish, Durrani, Mubarak (NAACL demo 2016)
# "Farasa: A Fast and Furious Segmenter for Arabic"
input-tokenizer = "$farasa-dir/farasa_moses.sh"
#input-tokenizer = "$farasa-dir/farasa_moses.sh"
# truecasers - comment out if you do not use the truecaser
input-truecaser = $moses-script-dir/recaser/truecase.perl

146
scripts/ems/experiment.meta
View File

@ -827,7 +827,7 @@ create-config
in: sigtest-filter-reordering-table sigtest-filter-phrase-translation-table transliteration-table generation-table-pruned sparse corpus-mml-prefilter=OR=corpus-mml-postfilter=OR=domains osm-model INTERPOLATED-LM:binlm LM:binlm
out: config
ignore-if: use-hiero thot
rerun-on-change: decoding-steps alignment-factors translation-factors reordering-factors generation-factors lexicalized-reordering training-options script decoding-graph-backoff score-settings additional-ini mmsapt no-glue-grammar dont-tune-glue-grammar use-syntax-input-weight-feature
rerun-on-change: decoding-steps alignment-factors translation-factors reordering-factors generation-factors lexicalized-reordering training-options script decoding-graph-backoff score-settings additional-ini mmsapt no-glue-grammar dont-tune-glue-grammar use-syntax-input-weight-feature operation-sequence-model-load-method
default-name: model/moses.ini
error: Unknown option
error: requires an argument
@ -1540,6 +1540,150 @@ analysis-precision
rerun-on-change: precision-by-coverage-base
final-model: yes
[QUALITY-ESTIMATION] single
tokenize-input
in: raw-input
out: tokenized-input
default-name: quality-estimation/input.tok
pass-unless: input-tokenizer
template: $input-tokenizer < IN > OUT
tokenize-input-devtest
in: raw-input-devtest
out: tokenized-input-devtest
default-name: quality-estimation/input.devtest.tok
pass-unless: input-tokenizer
template: $input-tokenizer < IN > OUT
lowercase-input
in: tokenized-input
out: truecased-input
default-name: quality-estimation/input.lc
pass-unless: input-lowercaser
ignore-if: input-truecaser
template: $input-lowercaser < IN > OUT
lowercase-input-devtest
in: tokenized-input-devtest
out: truecased-input-devtest
default-name: quality-estimation/input.devtest.lc
pass-unless: input-lowercaser
ignore-if: input-truecaser
template: $input-lowercaser < IN > OUT
truecase-input
in: tokenized-input TRUECASER:truecase-model
out: truecased-input
rerun-on-change: input-truecaser
default-name: quality-estimation/input.tc
ignore-unless: input-truecaser
template: $input-truecaser -model IN1.$input-extension < IN > OUT
truecase-input-devtest
in: tokenized-input-devtest TRUECASER:truecase-model
out: truecased-input-devtest
rerun-on-change: input-truecaser
ignore-unless: input-truecaser
default-name: quality-estimation/input.devtest.tc
template: $input-truecaser -model IN1.$input-extension < IN > OUT
split-input
in: truecased-input SPLITTER:splitter-model
out: split-input
rerun-on-change: input-splitter
default-name: quality-estimation/input.split
pass-unless: input-splitter
template: $input-splitter -model IN1.$input-extension < IN > OUT
split-input-devtest
in: truecased-input-devtest SPLITTER:splitter-model
out: split-input-devtest
rerun-on-change: input-splitter
default-name: quality-estimation/input.devtest.split
pass-unless: input-splitter
template: $input-splitter -model IN1.$input-extension < IN > OUT
tokenize-reference
in: raw-reference
out: tokenized-reference
default-name: quality-estimation/reference.tok
pass-unless: output-tokenizer
multiref: $moses-script-dir/ems/support/run-command-on-multiple-refsets.perl
template: $output-tokenizer < IN > OUT
tokenize-reference-devtest
in: raw-reference-devtest
out: tokenized-reference-devtest
default-name: quality-estimation/reference.devtest.tok
pass-unless: output-tokenizer
multiref: $moses-script-dir/ems/support/run-command-on-multiple-refsets.perl
template: $output-tokenizer < IN > OUT
lowercase-reference
in: tokenized-reference
out: truecased-reference
default-name: quality-estimation/reference.lc
pass-unless: output-lowercaser
ignore-if: output-truecaser
multiref: $moses-script-dir/ems/support/run-command-on-multiple-refsets.perl
template: $output-lowercaser < IN > OUT
lowercase-reference-devtest
in: tokenized-reference-devtest
out: truecased-reference-devtest
default-name: quality-estimation/reference.devtest.lc
pass-unless: output-lowercaser
ignore-if: output-truecaser
multiref: $moses-script-dir/ems/support/run-command-on-multiple-refsets.perl
template: $output-lowercaser < IN > OUT
truecase-reference
in: tokenized-reference TRUECASER:truecase-model
out: truecased-reference
rerun-on-change: output-truecaser
default-name: quality-estimation/reference.tc
ignore-unless: output-truecaser
multiref: $moses-script-dir/ems/support/run-command-on-multiple-refsets.perl
template: $output-truecaser -model IN1.$output-extension < IN > OUT
truecase-reference-devtest
in: tokenized-reference-devtest TRUECASER:truecase-model
out: truecased-reference-devtest
rerun-on-change: output-truecaser
default-name: quality-estimation/reference.devtest.tc
ignore-unless: output-truecaser
multiref: $moses-script-dir/ems/support/run-command-on-multiple-refsets.perl
template: $output-truecaser -model IN1.$output-extension < IN > OUT
decode
in: TUNING:config-with-reused-weights split-input
out: rich-output
default-name: quality-estimation/output
template: $decoder -v 0 -tt -f IN < IN1 > OUT
error: Translation was not performed correctly
not-error: trans: No such file or directory
decode-devtest
in: TUNING:config-with-reused-weights split-input-devtest
out: rich-output-devtest
default-name: quality-estimation/output-devtest
template: $decoder -v 0 -tt -f IN < IN1 > OUT
error: Translation was not performed correctly
not-error: trans: No such file or directory
remove-markup
in: rich-output
out: cleaned-output
default-name: quality-estimation/tokenized-output
template: $moses-script-dir/ems/support/remove-segmentation-markup.perl < IN > OUT
remove-markup-devtest
in: rich-output-devtest
out: cleaned-output-devtest
default-name: quality-estimation/tokenized-output-devtest
template: $moses-script-dir/ems/support/remove-segmentation-markup.perl < IN > OUT
score-output
in: cleaned-output truecased-reference
out: scored-output
default-name: quality-estimation/output-scored
tmp-name: quality-estimation/ter
template: mkdir TMP ; $moses-script-dir/ems/support/ter.perl $tercom IN IN1 TMP > OUT
score-output-devtest
in: cleaned-output-devtest truecased-reference-devtest
out: scored-output-devtest
default-name: quality-estimation/output-scored-devtest
tmp-name: quality-estimation/ter-devtest
template: mkdir TMP ; $moses-script-dir/ems/support/ter.perl $tercom IN IN1 TMP > OUT
train
in: input rich-output scored-output input-devtest rich-output-devtest scored-output-devtest
out: quality-estimation-model
default-name: quality-estimation/model
template: $trainer --train-rich IN1 --train-ter IN2 --eval-rich IN4 --eval-ter IN5 --model OUT
final-model: yes
[REPORTING] single
report
in: EVALUATION:nist-bleu-score EVALUATION:nist-bleu-c-score EVALUATION:bolt-bleu-score EVALUATION:bolt-bleu-c-score EVALUATION:multi-bleu-score EVALUATION:multi-bleu-c-score EVALUATION:multi-bleu-detok-score EVALUATION:multi-bleu-c-detok-score EVALUATION:meteor-score EVALUATION:ter-score EVALUATION:wer-score EVALUATION:ibm-bleu-score EVALUATION:ibm-bleu-c-score EVALUATION:analysis EVALUATION:analysis-coverage EVALUATION:analysis-prec TRAINING:biconcor-model EVALUATION:wade-analysis

6
scripts/ems/experiment.perl
View File

@ -2660,12 +2660,16 @@ sub define_training_create_config {
if ($osm) {
my $osm_settings = &get("TRAINING:operation-sequence-model-settings");
if ($osm_settings =~ /-factor *(\S+)/){
if ($osm_settings =~ /-factor *(\S+)/) {
$cmd .= "-osm-model $osm/ -osm-setting $1 ";
}
else {
$cmd .= "-osm-model $osm/operationLM.bin ";
}
my $osm_load_method = &get("TRAINING:operation-sequence-model-load-method");
if (defined($osm_load_method)) {
$cmd .= "-osm-load-method $osm_load_method ";
}
}
if (&get("TRAINING:phrase-orientation")) {

42
scripts/ems/support/create-xml.perl Executable file
View File

@ -0,0 +1,42 @@
#!/usr/bin/env perl
#
# This file is part of moses. Its use is licensed under the GNU Lesser General
# Public License version 2.1 or, at your option, any later version.
use warnings;
use strict;
my ($type) = @ARGV;
if ($type =~ /^s/i) {
print "<srcset setid=\"test\" srclang=\"any\">\n";
print "<doc docid=\"doc\">\n";
}
elsif ($type =~ /^t/i) {
print "<tstset setid=\"test\" tgtlang=\"any\" srclang=\"any\">\n";
print "<doc sysid=\"moses\" docid=\"doc\">\n";
}
elsif ($type =~ /^r/i) {
print "<refset setid=\"test\" tgtlang=\"any\" srclang=\"any\">\n";
print "<doc sysid=\"ref\" docid=\"doc\">\n";
}
else {
die("ERROR: specify source / target / ref");
}
my $i = 0;
while(<STDIN>) {
chomp;
print "<seg id=\"".(++$i)."\">$_</seg>\n";
}
print "</doc>\n";
if ($type =~ /^s/i) {
print "</srcset>\n";
}
elsif ($type =~ /^t/i) {
print "</tstset>\n";
}
elsif ($type =~ /^r/i) {
print "</refset>\n";
}

15
scripts/ems/support/remove-segmentation-markup.perl
View File

@ -9,7 +9,16 @@ use strict;
$|++;
while(<STDIN>) {
s/ \|\d+\-\d+\| / /g;
s/ \|\d+\-\d+\|$//;
print $_;
chop;
s/\|[^\|]+\|//g;
s/\s+/ /g;
s/^ //;
s/ $//;
print $_."\n";
}
#while(<STDIN>) {
# s/ \|\d+\-\d+\| / /g;
# s/ \|\d+\-\d+\|$//;
# print $_;
#}

15
scripts/ems/support/ter.perl Normal file
View File

@ -0,0 +1,15 @@
#!/usr/bin/env perl
#
# This file is part of moses. Its use is licensed under the GNU Lesser General
# Public License version 2.1 or, at your option, any later version.
use strict;
use FindBin qw($RealBin);
my ($jar, $hyp,$ref,$tmp) = @ARGV;
`mkdir -p $tmp`;
`$RealBin/create-xml.perl test < $hyp > $tmp/hyp`;
`$RealBin/create-xml.perl ref < $ref > $tmp/ref`;
`java -jar $jar -h $tmp/hyp -r $tmp/ref -o ter -n $tmp/out`;
print `cat $tmp/out.ter`;

14
scripts/training/train-model.perl
View File

@ -83,6 +83,7 @@ my($_EXTERNAL_BINDIR,
$_CONFIG,
$_OSM,
$_OSM_FACTORS,
$_OSM_LOAD_METHOD,
$_POST_DECODING_TRANSLIT,
$_TRANSLITERATION_PHRASE_TABLE,
$_HIERARCHICAL,
@ -238,6 +239,7 @@ $_HELP = 1
'config=s' => \$_CONFIG,
'osm-model=s' => \$_OSM,
'osm-setting=s' => \$_OSM_FACTORS,
'osm-load-method=s' => \$_OSM_LOAD_METHOD,
'post-decoding-translit=s' => \$_POST_DECODING_TRANSLIT,
'transliteration-phrase-table=s' => \$_TRANSLITERATION_PHRASE_TABLE,
'mmsapt' => \$_MMSAPT,
@ -2249,6 +2251,8 @@ sub create_ini {
if($_OSM)
{
my $load_method = "";
$load_method = " load=$_OSM_LOAD_METHOD" if defined($_OSM_LOAD_METHOD);
if (defined($_OSM_FACTORS))
{
my $count = 0;
@ -2258,11 +2262,11 @@ sub create_ini {
my ($factor_f,$factor_e) = split(/\-/,$factor_val);
if($count == 0){
$feature_spec .= "OpSequenceModel name=OpSequenceModel$count num-features=5 path=". $_OSM . $factor_val . "/operationLM.bin" . " input-factor=". $factor_f . " output-factor=". $factor_e . " support-features=yes \n";
$feature_spec .= "OpSequenceModel$load_method name=OpSequenceModel$count num-features=5 path=". $_OSM . $factor_val . "/operationLM.bin" . " input-factor=". $factor_f . " output-factor=". $factor_e . " support-features=yes \n";
$weight_spec .= "OpSequenceModel$count= 0.08 -0.02 0.02 -0.001 0.03\n";
}
else{
$feature_spec .= "OpSequenceModel name=OpSequenceModel$count num-features=1 path=". $_OSM . $factor_val . "/operationLM.bin" . " input-factor=". $factor_f . " output-factor=". $factor_e . " support-features=no \n";
$feature_spec .= "OpSequenceModel$load_method name=OpSequenceModel$count num-features=1 path=". $_OSM . $factor_val . "/operationLM.bin" . " input-factor=". $factor_f . " output-factor=". $factor_e . " support-features=no \n";
$weight_spec .= "OpSequenceModel$count= 0.08 \n";
}
@ -2271,7 +2275,7 @@ sub create_ini {
}
else
{
$feature_spec .= "OpSequenceModel name=OpSequenceModel0 num-features=5 path=". $_OSM . " \n";
$feature_spec .= "OpSequenceModel$load_method name=OpSequenceModel0 num-features=5 path=". $_OSM . " \n";
$weight_spec .= "OpSequenceModel0= 0.08 -0.02 0.02 -0.001 0.03\n";
}
}
@ -2292,7 +2296,9 @@ sub create_ini {
}
$type = "KENLM" unless defined $type; # default to KENLM if no type given
if ($type =~ /^\d+$/) {
if ($type =~ /^8-(.+)/) {
$type = "KENLM load=$1";
} elsif ($type =~ /^\d+$/) {
# backwards compatibility if the type is given not as string but as a number
if ($type == 0) {
$type = "SRILM";

38
vw/Classifier.h
View File

@ -24,6 +24,8 @@ class ezexample;
namespace Discriminative
{
typedef std::pair<uint32_t, float> FeatureType; // feature hash (=ID) and value
typedef std::vector<FeatureType> FeatureVector;
/**
* Abstract class to be implemented by classifiers.
@ -34,12 +36,22 @@ public:
/**
* Add a feature that does not depend on the class (label).
*/
virtual void AddLabelIndependentFeature(const StringPiece &name, float value) = 0;
virtual FeatureType AddLabelIndependentFeature(const StringPiece &name, float value) = 0;
/**
* Add a feature that is specific for the given class.
*/
virtual void AddLabelDependentFeature(const StringPiece &name, float value) = 0;
virtual FeatureType AddLabelDependentFeature(const StringPiece &name, float value) = 0;
/**
* Efficient addition of features when their IDs are already computed.
*/
virtual void AddLabelIndependentFeatureVector(const FeatureVector &features) = 0;
/**
* Efficient addition of features when their IDs are already computed.
*/
virtual void AddLabelDependentFeatureVector(const FeatureVector &features) = 0;
/**
* Train using current example. Use loss to distinguish positive and negative training examples.
@ -54,12 +66,12 @@ public:
virtual float Predict(const StringPiece &label) = 0;
// helper methods for indicator features
void AddLabelIndependentFeature(const StringPiece &name) {
AddLabelIndependentFeature(name, 1.0);
FeatureType AddLabelIndependentFeature(const StringPiece &name) {
return AddLabelIndependentFeature(name, 1.0);
}
void AddLabelDependentFeature(const StringPiece &name) {
AddLabelDependentFeature(name, 1.0);
FeatureType AddLabelDependentFeature(const StringPiece &name) {
return AddLabelDependentFeature(name, 1.0);
}
virtual ~Classifier() {}
@ -95,8 +107,10 @@ public:
VWTrainer(const std::string &outputFile);
virtual ~VWTrainer();
virtual void AddLabelIndependentFeature(const StringPiece &name, float value);
virtual void AddLabelDependentFeature(const StringPiece &name, float value);
virtual FeatureType AddLabelIndependentFeature(const StringPiece &name, float value);
virtual FeatureType AddLabelDependentFeature(const StringPiece &name, float value);
virtual void AddLabelIndependentFeatureVector(const FeatureVector &features);
virtual void AddLabelDependentFeatureVector(const FeatureVector &features);
virtual void Train(const StringPiece &label, float loss);
virtual float Predict(const StringPiece &label);
@ -121,15 +135,17 @@ public:
VWPredictor(const std::string &modelFile, const std::string &vwOptions);
virtual ~VWPredictor();
virtual void AddLabelIndependentFeature(const StringPiece &name, float value);
virtual void AddLabelDependentFeature(const StringPiece &name, float value);
virtual FeatureType AddLabelIndependentFeature(const StringPiece &name, float value);
virtual FeatureType AddLabelDependentFeature(const StringPiece &name, float value);
virtual void AddLabelIndependentFeatureVector(const FeatureVector &features);
virtual void AddLabelDependentFeatureVector(const FeatureVector &features);
virtual void Train(const StringPiece &label, float loss);
virtual float Predict(const StringPiece &label);
friend class ClassifierFactory;
protected:
void AddFeature(const StringPiece &name, float values);
FeatureType AddFeature(const StringPiece &name, float values);
::vw *m_VWInstance, *m_VWParser;
::ezexample *m_ex;

14
vw/Normalizer.h
View File

@ -2,6 +2,7 @@
#define moses_Normalizer_h
#include <vector>
#include <algorithm>
#include "Util.h"
namespace Discriminative
@ -45,16 +46,25 @@ public:
virtual ~SquaredLossNormalizer() {}
};
// safe softmax
class LogisticLossNormalizer : public Normalizer
{
public:
virtual void operator()(std::vector<float> &losses) const {
float sum = 0;
std::vector<float>::iterator it;
float sum = 0;
float max = 0;
for (it = losses.begin(); it != losses.end(); it++) {
*it = exp(-*it);
*it = -*it;
max = std::max(max, *it);
}
for (it = losses.begin(); it != losses.end(); it++) {
*it = exp(*it - max);
sum += *it;
}
for (it = losses.begin(); it != losses.end(); it++) {
*it /= sum;
}

42
vw/VWPredictor.cpp
View File

@ -36,7 +36,7 @@ VWPredictor::~VWPredictor()
VW::finish(*m_VWInstance);
}
void VWPredictor::AddLabelIndependentFeature(const StringPiece &name, float value)
FeatureType VWPredictor::AddLabelIndependentFeature(const StringPiece &name, float value)
{
// label-independent features are kept in a different feature namespace ('s' = source)
@ -48,10 +48,10 @@ void VWPredictor::AddLabelIndependentFeature(const StringPiece &name, float valu
m_ex->addns('s');
if (DEBUG) std::cerr << "VW :: Setting source namespace\n";
}
AddFeature(name, value); // namespace 's' is set up, add the feature
return AddFeature(name, value); // namespace 's' is set up, add the feature
}
void VWPredictor::AddLabelDependentFeature(const StringPiece &name, float value)
FeatureType VWPredictor::AddLabelDependentFeature(const StringPiece &name, float value)
{
// VW does not use the label directly, instead, we do a Cartesian product between source and target feature
// namespaces, where the source namespace ('s') contains label-independent features and the target
@ -63,7 +63,37 @@ void VWPredictor::AddLabelDependentFeature(const StringPiece &name, float value)
m_ex->addns('t');
if (DEBUG) std::cerr << "VW :: Setting target namespace\n";
}
AddFeature(name, value);
return AddFeature(name, value);
}
void VWPredictor::AddLabelIndependentFeatureVector(const FeatureVector &features)
{
if (m_isFirstSource) {
// the first feature of a new example => create the source namespace for
// label-independent features to live in
m_isFirstSource = false;
m_ex->finish();
m_ex->addns('s');
if (DEBUG) std::cerr << "VW :: Setting source namespace\n";
}
// add each feature index using this "low level" call to VW
for (FeatureVector::const_iterator it = features.begin(); it != features.end(); it++)
m_ex->addf(it->first, it->second);
}
void VWPredictor::AddLabelDependentFeatureVector(const FeatureVector &features)
{
if (m_isFirstTarget) {
// the first target-side feature => create namespace 't'
m_isFirstTarget = false;
m_ex->addns('t');
if (DEBUG) std::cerr << "VW :: Setting target namespace\n";
}
// add each feature index using this "low level" call to VW
for (FeatureVector::const_iterator it = features.begin(); it != features.end(); it++)
m_ex->addf(it->first, it->second);
}
void VWPredictor::Train(const StringPiece &label, float loss)
@ -82,10 +112,10 @@ float VWPredictor::Predict(const StringPiece &label)
return loss;
}
void VWPredictor::AddFeature(const StringPiece &name, float value)
FeatureType VWPredictor::AddFeature(const StringPiece &name, float value)
{
if (DEBUG) std::cerr << "VW :: Adding feature: " << EscapeSpecialChars(name.as_string()) << ":" << value << "\n";
m_ex->addf(EscapeSpecialChars(name.as_string()), value);
return std::make_pair(m_ex->addf(EscapeSpecialChars(name.as_string()), value), value);
}
} // namespace Discriminative

18
vw/VWTrainer.cpp
View File

@ -25,7 +25,7 @@ VWTrainer::~VWTrainer()
close(m_bfos);
}
void VWTrainer::AddLabelIndependentFeature(const StringPiece &name, float value)
FeatureType VWTrainer::AddLabelIndependentFeature(const StringPiece &name, float value)
{
if (m_isFirstSource) {
if (m_isFirstExample) {
@ -43,9 +43,11 @@ void VWTrainer::AddLabelIndependentFeature(const StringPiece &name, float value)
}
AddFeature(name, value);
return std::make_pair(0, value); // we don't hash features
}
void VWTrainer::AddLabelDependentFeature(const StringPiece &name, float value)
FeatureType VWTrainer::AddLabelDependentFeature(const StringPiece &name, float value)
{
if (m_isFirstTarget) {
m_isFirstTarget = false;
@ -56,6 +58,18 @@ void VWTrainer::AddLabelDependentFeature(const StringPiece &name, float value)
}
AddFeature(name, value);
return std::make_pair(0, value); // we don't hash features
}
void VWTrainer::AddLabelIndependentFeatureVector(const FeatureVector &features)
{
throw logic_error("VW trainer does not support feature IDs.");
}
void VWTrainer::AddLabelDependentFeatureVector(const FeatureVector &features)
{
throw logic_error("VW trainer does not support feature IDs.");
}
void VWTrainer::Train(const StringPiece &label, float loss)