1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
|
// $Id$
/***********************************************************************
Moses - factored phrase-based language decoder
Copyright (C) 2006 University of Edinburgh
This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
This library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with this library; if not, write to the Free Software
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
***********************************************************************/
#include <cstring>
#include <iostream>
#include <memory>
#include <stdlib.h>
#include "lm/binary_format.hh"
#include "lm/enumerate_vocab.hh"
#include "lm/left.hh"
#include "lm/model.hh"
#include "LM/Ken.h"
#include "LM/Base.h"
#include "FFState.h"
#include "TypeDef.h"
#include "Util.h"
#include "FactorCollection.h"
#include "Phrase.h"
#include "InputFileStream.h"
#include "StaticData.h"
#include "ChartHypothesis.h"
#include <boost/shared_ptr.hpp>
using namespace std;
namespace Moses {
namespace {
struct KenLMState : public FFState {
lm::ngram::State state;
int Compare(const FFState &o) const {
const KenLMState &other = static_cast<const KenLMState &>(o);
if (state.length < other.state.length) return -1;
if (state.length > other.state.length) return 1;
return std::memcmp(state.words, other.state.words, sizeof(lm::WordIndex) * state.length);
}
};
/*
* An implementation of single factor LM using Ken's code.
*/
template <class Model> class LanguageModelKen : public LanguageModel {
public:
LanguageModelKen(const std::string &file, ScoreIndexManager &manager, FactorType factorType, bool lazy);
LanguageModel *Duplicate(ScoreIndexManager &scoreIndexManager) const;
bool Useable(const Phrase &phrase) const {
return (phrase.GetSize()>0 && phrase.GetFactor(0, m_factorType) != NULL);
}
std::string GetScoreProducerDescription(unsigned) const {
std::ostringstream oss;
oss << "LM_" << (unsigned)m_ngram->Order() << "gram";
return oss.str();
}
const FFState *EmptyHypothesisState(const InputType &/*input*/) const {
KenLMState *ret = new KenLMState();
ret->state = m_ngram->BeginSentenceState();
return ret;
}
void CalcScore(const Phrase &phrase, float &fullScore, float &ngramScore, size_t &oovCount) const;
FFState *Evaluate(const Hypothesis &hypo, const FFState *ps, ScoreComponentCollection *out) const;
FFState *EvaluateChart(const ChartHypothesis& cur_hypo, int featureID, ScoreComponentCollection *accumulator) const;
private:
LanguageModelKen(ScoreIndexManager &manager, const LanguageModelKen<Model> ©_from);
lm::WordIndex TranslateID(const Word &word) const {
std::size_t factor = word.GetFactor(m_factorType)->GetId();
return (factor >= m_lmIdLookup.size() ? 0 : m_lmIdLookup[factor]);
}
// Convert last words of hypothesis into vocab ids, returning an end pointer.
lm::WordIndex *LastIDs(const Hypothesis &hypo, lm::WordIndex *indices) const {
lm::WordIndex *index = indices;
lm::WordIndex *end = indices + m_ngram->Order() - 1;
int position = hypo.GetCurrTargetWordsRange().GetEndPos();
for (; ; ++index, --position) {
if (index == end) return index;
if (position == -1) {
*index = m_ngram->GetVocabulary().BeginSentence();
return index + 1;
}
*index = TranslateID(hypo.GetWord(position));
}
}
boost::shared_ptr<Model> m_ngram;
std::vector<lm::WordIndex> m_lmIdLookup;
FactorType m_factorType;
const Factor *m_beginSentenceFactor;
};
class MappingBuilder : public lm::EnumerateVocab {
public:
MappingBuilder(FactorCollection &factorCollection, std::vector<lm::WordIndex> &mapping)
: m_factorCollection(factorCollection), m_mapping(mapping) {}
void Add(lm::WordIndex index, const StringPiece &str) {
std::size_t factorId = m_factorCollection.AddFactor(str)->GetId();
if (m_mapping.size() <= factorId) {
// 0 is <unk> :-)
m_mapping.resize(factorId + 1);
}
m_mapping[factorId] = index;
}
private:
FactorCollection &m_factorCollection;
std::vector<lm::WordIndex> &m_mapping;
};
template <class Model> LanguageModelKen<Model>::LanguageModelKen(const std::string &file, ScoreIndexManager &manager, FactorType factorType, bool lazy) : m_factorType(factorType) {
lm::ngram::Config config;
IFVERBOSE(1) {
config.messages = &std::cerr;
} else {
config.messages = NULL;
}
FactorCollection &collection = FactorCollection::Instance();
MappingBuilder builder(collection, m_lmIdLookup);
config.enumerate_vocab = &builder;
config.load_method = lazy ? util::LAZY : util::POPULATE_OR_READ;
m_ngram.reset(new Model(file.c_str(), config));
m_beginSentenceFactor = collection.AddFactor(BOS_);
Init(manager);
}
template <class Model> LanguageModel *LanguageModelKen<Model>::Duplicate(ScoreIndexManager &manager) const {
return new LanguageModelKen<Model>(manager, *this);
}
template <class Model> LanguageModelKen<Model>::LanguageModelKen(ScoreIndexManager &manager, const LanguageModelKen<Model> ©_from) :
m_ngram(copy_from.m_ngram),
// TODO: don't copy this.
m_lmIdLookup(copy_from.m_lmIdLookup),
m_factorType(copy_from.m_factorType),
m_beginSentenceFactor(copy_from.m_beginSentenceFactor) {
Init(manager);
}
template <class Model> void LanguageModelKen<Model>::CalcScore(const Phrase &phrase, float &fullScore, float &ngramScore, size_t &oovCount) const {
fullScore = 0;
ngramScore = 0;
oovCount = 0;
if (!phrase.GetSize()) return;
typename Model::State state_backing[2];
typename Model::State *state0 = &state_backing[0], *state1 = &state_backing[1];
size_t position;
if (m_beginSentenceFactor == phrase.GetWord(0).GetFactor(m_factorType)) {
*state0 = m_ngram->BeginSentenceState();
position = 1;
} else {
*state0 = m_ngram->NullContextState();
position = 0;
}
size_t ngramBoundary = m_ngram->Order() - 1;
for (; position < phrase.GetSize(); ++position) {
const Word &word = phrase.GetWord(position);
if (word.IsNonTerminal()) {
// If there's a non-terminal at 1 and we have a 5-gram LM, then positions 2 3 4 and 5 will be incomplete while position 6 is complete.
ngramBoundary = m_ngram->Order() + position;
*state0 = m_ngram->NullContextState();
} else {
lm::WordIndex index = TranslateID(word);
if (index == m_ngram->GetVocabulary().BeginSentence()) {
std::cerr << "Either your data contains <s> in a position other than the first word or your language model is missing <s>. Did you build your ARPA using IRSTLM and forget to run add-start-end.sh?" << std::endl;
abort();
}
float score = TransformLMScore(m_ngram->Score(*state0, index, *state1));
std::swap(state0, state1);
if (position >= ngramBoundary) ngramScore += score;
fullScore += score;
if (!index) ++oovCount;
}
}
}
template <class Model> FFState *LanguageModelKen<Model>::Evaluate(const Hypothesis &hypo, const FFState *ps, ScoreComponentCollection *out) const {
const lm::ngram::State &in_state = static_cast<const KenLMState&>(*ps).state;
std::auto_ptr<KenLMState> ret(new KenLMState());
if (!hypo.GetCurrTargetLength()) {
ret->state = in_state;
return ret.release();
}
const std::size_t begin = hypo.GetCurrTargetWordsRange().GetStartPos();
//[begin, end) in STL-like fashion.
const std::size_t end = hypo.GetCurrTargetWordsRange().GetEndPos() + 1;
const std::size_t adjust_end = std::min(end, begin + m_ngram->Order() - 1);
std::size_t position = begin;
typename Model::State aux_state;
typename Model::State *state0 = &ret->state, *state1 = &aux_state;
float score = m_ngram->Score(in_state, TranslateID(hypo.GetWord(position)), *state0);
++position;
for (; position < adjust_end; ++position) {
score += m_ngram->Score(*state0, TranslateID(hypo.GetWord(position)), *state1);
std::swap(state0, state1);
}
if (hypo.IsSourceCompleted()) {
// Score end of sentence.
std::vector<lm::WordIndex> indices(m_ngram->Order() - 1);
const lm::WordIndex *last = LastIDs(hypo, &indices.front());
score += m_ngram->FullScoreForgotState(&indices.front(), last, m_ngram->GetVocabulary().EndSentence(), ret->state).prob;
} else if (adjust_end < end) {
// Get state after adding a long phrase.
std::vector<lm::WordIndex> indices(m_ngram->Order() - 1);
const lm::WordIndex *last = LastIDs(hypo, &indices.front());
m_ngram->GetState(&indices.front(), last, ret->state);
} else if (state0 != &ret->state) {
// Short enough phrase that we can just reuse the state.
ret->state = *state0;
}
score = TransformLMScore(score);
if (OOVFeatureEnabled()) {
std::vector<float> scores(2);
scores[0] = score;
scores[1] = 0.0;
out->PlusEquals(this, scores);
} else {
out->PlusEquals(this, score);
}
return ret.release();
}
class LanguageModelChartStateKenLM : public FFState {
public:
LanguageModelChartStateKenLM() {}
const lm::ngram::ChartState &GetChartState() const { return m_state; }
lm::ngram::ChartState &GetChartState() { return m_state; }
int Compare(const FFState& o) const
{
const LanguageModelChartStateKenLM &other = static_cast<const LanguageModelChartStateKenLM&>(o);
int ret = m_state.Compare(other.m_state);
return ret;
}
private:
lm::ngram::ChartState m_state;
};
template <class Model> FFState *LanguageModelKen<Model>::EvaluateChart(const ChartHypothesis& hypo, int featureID, ScoreComponentCollection *accumulator) const {
LanguageModelChartStateKenLM *newState = new LanguageModelChartStateKenLM();
lm::ngram::RuleScore<Model> ruleScore(*m_ngram, newState->GetChartState());
const AlignmentInfo::NonTermIndexMap &nonTermIndexMap = hypo.GetCurrTargetPhrase().GetAlignmentInfo().GetNonTermIndexMap();
const size_t size = hypo.GetCurrTargetPhrase().GetSize();
size_t phrasePos = 0;
// Special cases for first word.
if (size) {
const Word &word = hypo.GetCurrTargetPhrase().GetWord(0);
if (word.GetFactor(m_factorType) == m_beginSentenceFactor) {
// Begin of sentence
ruleScore.BeginSentence();
phrasePos++;
} else if (word.IsNonTerminal()) {
// Non-terminal is first so we can copy instead of rescoring.
const ChartHypothesis *prevHypo = hypo.GetPrevHypo(nonTermIndexMap[phrasePos]);
const lm::ngram::ChartState &prevState = static_cast<const LanguageModelChartStateKenLM*>(prevHypo->GetFFState(featureID))->GetChartState();
ruleScore.BeginNonTerminal(prevState, prevHypo->GetScoreBreakdown().GetScoresForProducer(this)[0]);
phrasePos++;
}
}
for (; phrasePos < size; phrasePos++) {
const Word &word = hypo.GetCurrTargetPhrase().GetWord(phrasePos);
if (word.IsNonTerminal()) {
const ChartHypothesis *prevHypo = hypo.GetPrevHypo(nonTermIndexMap[phrasePos]);
const lm::ngram::ChartState &prevState = static_cast<const LanguageModelChartStateKenLM*>(prevHypo->GetFFState(featureID))->GetChartState();
ruleScore.NonTerminal(prevState, prevHypo->GetScoreBreakdown().GetScoresForProducer(this)[0]);
} else {
ruleScore.Terminal(TranslateID(word));
}
}
accumulator->Assign(this, ruleScore.Finish());
return newState;
}
} // namespace
LanguageModel *ConstructKenLM(const std::string &file, ScoreIndexManager &manager, FactorType factorType, bool lazy) {
try {
lm::ngram::ModelType model_type;
if (lm::ngram::RecognizeBinary(file.c_str(), model_type)) {
switch(model_type) {
case lm::ngram::HASH_PROBING:
return new LanguageModelKen<lm::ngram::ProbingModel>(file, manager, factorType, lazy);
case lm::ngram::TRIE_SORTED:
return new LanguageModelKen<lm::ngram::TrieModel>(file, manager, factorType, lazy);
case lm::ngram::QUANT_TRIE_SORTED:
return new LanguageModelKen<lm::ngram::QuantTrieModel>(file, manager, factorType, lazy);
case lm::ngram::ARRAY_TRIE_SORTED:
return new LanguageModelKen<lm::ngram::ArrayTrieModel>(file, manager, factorType, lazy);
case lm::ngram::QUANT_ARRAY_TRIE_SORTED:
return new LanguageModelKen<lm::ngram::QuantArrayTrieModel>(file, manager, factorType, lazy);
default:
std::cerr << "Unrecognized kenlm model type " << model_type << std::endl;
abort();
}
} else {
return new LanguageModelKen<lm::ngram::ProbingModel>(file, manager, factorType, lazy);
}
} catch (std::exception &e) {
std::cerr << e.what() << std::endl;
abort();
}
}
}
|