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
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
|
// -*- mode: c++; tab-width: 2; indent-tabs-mode: nil -*-
#pragma once
#include <algorithm>
#include <boost/random.hpp>
#include <boost/thread.hpp>
#include <boost/thread/locks.hpp>
#include <boost/intrusive_ptr.hpp>
#include <boost/math/distributions/binomial.hpp>
#include "ug_bitext.h"
#include "ug_bitext_pstats.h"
#include "ug_sampling_bias.h"
#include "ug_tsa_array_entry.h"
#include "ug_bitext_phrase_extraction_record.h"
#include "moses/TranslationModel/UG/generic/threading/ug_ref_counter.h"
#include "moses/TranslationModel/UG/generic/threading/ug_thread_safe_counter.h"
#include "moses/TranslationModel/UG/generic/sorting/NBestList.h"
namespace Moses
{
namespace bitext
{
enum sampling_method { full_coverage, random_sampling, ranked_sampling };
typedef ugdiss::ttrack::Position TokenPosition;
class CandidateSorter
{
SamplingBias const& score;
public:
CandidateSorter(SamplingBias const& s) : score(s) {}
bool operator()(TokenPosition const& a, TokenPosition const& b) const
{ return score[a.sid] > score[b.sid]; }
};
template<typename Token>
class
BitextSampler : public reference_counter
{
typedef Bitext<Token> bitext;
typedef TSA<Token> tsa;
typedef SamplingBias bias;
typedef typename Bitext<Token>::iter tsa_iter;
mutable boost::condition_variable m_ready;
mutable boost::mutex m_lock;
// const members
// sptr<bitext const> const m_bitext; // keep bitext alive while I am
// should be an
iptr<bitext const> const m_bitext; // keep bitext alive as long as I am
size_t const m_plen; // length of lookup phrase
bool const m_fwd; // forward or backward direction?
sptr<tsa const> const m_root; // root of suffix array
char const* m_next; // current position
char const* m_stop; // end of search range
sampling_method const m_method; // look at all/random/ranked samples
sptr<bias const> const m_bias; // bias over candidates
size_t const m_samples; // how many samples at most
// non-const members
sptr<pstats> m_stats; // destination for phrase stats
size_t m_ctr; // number of samples considered
float m_total_bias; // for random sampling with bias
bool m_finished;
boost::taus88 m_rnd; // every job has its own pseudo random generator
double m_rnddenom; // denominator for scaling random sampling
double m_bias_total;
bool consider_sample(TokenPosition const& p);
size_t perform_ranked_sampling();
size_t perform_random_sampling();
int check_sample_distribution(uint64_t const& sid, uint64_t const& offset);
bool flip_coin(ugdiss::id_type & sid, ushort & offset);
public:
BitextSampler(BitextSampler const& other);
BitextSampler const& operator=(BitextSampler const& other);
BitextSampler(bitext const* const bitext, typename bitext::iter const& phrase,
sptr<SamplingBias const> const& bias, size_t const max_samples,
sampling_method const method);
~BitextSampler();
bool operator()(); // run sampling
sptr<pstats> stats();
bool done() const;
};
template<typename Token>
int
BitextSampler<Token>::
check_sample_distribution(uint64_t const& sid, uint64_t const& offset)
{ // ensure that the sampled distribution approximately matches the bias
// @return 0: SKIP this occurrence
// @return 1: consider this occurrence for sampling
// @return 2: include this occurrence in the sample by all means
typedef boost::math::binomial_distribution<> binomial;
// std::ostream* log = m_bias->loglevel > 1 ? m_bias->log : NULL;
std::ostream* log = NULL;
if (!m_bias) return 1;
float p = (*m_bias)[sid];
id_type docid = m_bias->GetClass(sid);
std::map<uint32_t,uint32_t>::const_iterator m = m_stats->indoc.find(docid);
uint32_t k = m != m_stats->indoc.end() ? m->second : 0 ;
// always consider candidates from dominating documents and
// from documents that have not been considered at all yet
bool ret = (p > .5 || k == 0);
if (ret && !log) return 1;
uint32_t N = m_stats->good; // number of trials
float d = cdf(complement(binomial(N, p), k));
// d: probability that samples contains k or more instances from doc #docid
ret = ret || d >= .05;
#if 0
if (log)
{
Token const* t = m_root->getCorpus()->sntStart(sid)+offset;
Token const* x = t - min(offset,uint64_t(3));
Token const* e = t + 4;
if (e > m_root->getCorpus()->sntEnd(sid))
e = m_root->getCorpus()->sntEnd(sid);
*log << docid << ":" << sid << " " << size_t(k) << "/" << N
<< " @" << p << " => " << d << " [";
std::map<uint32_t, uint32_t>::const_iterator m;
for (m = m_stats->indoc.begin(); m != m_stats->indoc.end(); ++m)
{
if (m != m_stats->indoc.begin()) *log << " ";
*log << m->first << ":" << m->second;
}
*log << "] ";
for (; x < e; ++x) *log << (*m_bitext->V1)[x->id()] << " ";
if (!ret) *log << "SKIP";
else if (p < .5 && d > .9) *log << "FORCE";
*log << std::endl;
}
#endif
return (ret ? (p < .5 && d > .9) ? 2 : 1 : 0);
}
template<typename Token>
bool
BitextSampler<Token>::
flip_coin(ugdiss::id_type & sid, ushort & offset)
{
int no_maybe_yes = m_bias ? check_sample_distribution(sid, offset) : 1;
if (no_maybe_yes == 0) return false; // no
if (no_maybe_yes > 1) return true; // yes
// ... maybe: flip a coin
size_t options_chosen = m_stats->good;
size_t options_total = max(m_stats->raw_cnt, m_ctr);
size_t options_left = (options_total - m_ctr);
size_t random_number = options_left * (m_rnd()/(m_rnd.max()+1.));
size_t threshold;
if (m_bias_total) // we have a bias and there are candidates with non-zero prob
threshold = ((*m_bias)[sid]/m_bias_total * options_total * m_samples);
else // no bias, or all have prob 0 (can happen with a very opinionated bias)
threshold = m_samples;
return random_number + options_chosen < threshold;
}
template<typename Token>
BitextSampler<Token>::
BitextSampler(Bitext<Token> const* const bitext,
typename bitext::iter const& phrase,
sptr<SamplingBias const> const& bias, size_t const max_samples,
sampling_method const method)
: m_bitext(bitext)
, m_plen(phrase.size())
, m_fwd(phrase.root == bitext->I1.get())
, m_root(m_fwd ? bitext->I1 : bitext->I2)
, m_next(phrase.lower_bound(-1))
, m_stop(phrase.upper_bound(-1))
, m_method(method)
, m_bias(bias)
, m_samples(max_samples)
, m_ctr(0)
, m_total_bias(0)
, m_finished(false)
{
m_stats.reset(new pstats);
m_stats->raw_cnt = phrase.ca();
m_stats->register_worker();
// cerr << phrase.str(bitext->V1.get()) << " [" << HERE << "]" << endl;
}
template<typename Token>
BitextSampler<Token>::
BitextSampler(BitextSampler const& other)
: m_bitext(other.m_bitext)
, m_plen(other.m_plen)
, m_fwd(other.m_fwd)
, m_root(other.m_root)
, m_next(other.m_next)
, m_stop(other.m_stop)
, m_method(other.m_method)
, m_bias(other.m_bias)
, m_samples(other.m_samples)
{
// lock both instances
boost::unique_lock<boost::mutex> mylock(m_lock);
boost::unique_lock<boost::mutex> yrlock(other.m_lock);
// actually, BitextSamplers should only copied on job submission
m_stats = other.m_stats;
m_stats->register_worker();
m_ctr = other.m_ctr;
m_total_bias = other.m_total_bias;
m_finished = other.m_finished;
}
// Ranked sampling sorts all samples by score and then considers the top-ranked
// candidates for phrase extraction.
template<typename Token>
size_t
BitextSampler<Token>::
perform_ranked_sampling()
{
if (m_next == m_stop) return m_ctr;
CandidateSorter sorter(*m_bias);
// below: nbest size = 4 * m_samples to allow for failed phrase extraction
NBestList<TokenPosition, CandidateSorter> nbest(4*m_samples,sorter);
ugdiss::tsa::ArrayEntry I(m_next);
while (I.next < m_stop)
{
++m_ctr;
nbest.add(m_root->readEntry(I.next,I));
}
for (size_t i = 0; m_stats->good < m_samples && i < nbest.size(); ++i)
consider_sample(nbest[i]);
return m_ctr;
}
// Ranked sampling sorts all samples by score and then considers the top-ranked
// candidates for phrase extraction.
template<typename Token>
size_t
BitextSampler<Token>::
perform_random_sampling()
{
if (m_next == m_stop) return m_ctr;
m_bias_total = 0;
if (m_bias)
{
m_stats->raw_cnt = 0;
for (ugdiss::tsa::ArrayEntry I(m_next); I.next < m_stop;)
{
m_root->readEntry(I.next,I);
++m_stats->raw_cnt;
m_bias_total += (*m_bias)[I.sid];
}
}
ugdiss::tsa::ArrayEntry I(m_next);
while (m_stats->good < m_samples && I.next < m_stop)
{
++m_ctr;
m_root->readEntry(I.next,I);
if (!flip_coin(I.sid, I.offset)) continue;
consider_sample(I);
}
return m_ctr;
}
template<typename Token>
bool
BitextSampler<Token>::
consider_sample(TokenPosition const& p)
{
std::vector<unsigned char> aln;
bitvector full_aln(100*100);
PhraseExtractionRecord
rec(p.sid, p.offset, p.offset + m_plen, !m_fwd, &aln, &full_aln);
int docid = m_bias ? m_bias->GetClass(p.sid) : -1;
if (!m_bitext->find_trg_phr_bounds(rec))
{ // no good, probably because phrase is not coherent
m_stats->count_sample(docid, 0, rec.po_fwd, rec.po_bwd);
return false;
}
// all good: register this sample as valid
size_t num_pairs = (rec.s2 - rec.s1 + 1) * (rec.e2 - rec.e1 + 1);
m_stats->count_sample(docid, num_pairs, rec.po_fwd, rec.po_bwd);
float sample_weight = 1./num_pairs;
Token const* o = (m_fwd ? m_bitext->T2 : m_bitext->T1)->sntStart(rec.sid);
// adjust offsets in phrase-internal aligment
for (size_t k = 1; k < aln.size(); k += 2)
aln[k] += rec.s2 - rec.s1;
vector<uint64_t> seen; seen.reserve(10);
// It is possible that the phrase extraction extracts the same
// phrase twice, e.g., when word a co-occurs with sequence b b b but
// is aligned only to the middle word. We can only count each phrase
// pair once per source phrase occurrence, or else run the risk of
// having more joint counts than marginal counts.
for (size_t s = rec.s1; s <= rec.s2; ++s)
{
TSA<Token> const& I = m_fwd ? *m_bitext->I2 : *m_bitext->I1;
sptr<tsa_iter> b = I.find(o + s, rec.e1 - s);
UTIL_THROW_IF2(!b || b->size() < rec.e1 - s, "target phrase not found");
for (size_t i = rec.e1; i <= rec.e2; ++i)
{
uint64_t tpid = b->getPid();
if (find(seen.begin(), seen.end(), tpid) != seen.end())
continue; // don't over-count
seen.push_back(tpid);
size_t raw2 = b->approxOccurrenceCount();
m_stats->add(tpid, sample_weight, m_bias ? (*m_bias)[p.sid] : 1,
aln, raw2, rec.po_fwd, rec.po_bwd, docid);
bool ok = (i == rec.e2) || b->extend(o[i].id());
UTIL_THROW_IF2(!ok, "Could not extend target phrase.");
}
if (s < rec.s2) // shift phrase-internal alignments
for (size_t k = 1; k < aln.size(); k += 2)
--aln[k];
}
return true;
}
template<typename Token>
bool
BitextSampler<Token>::
operator()()
{
if (m_finished) return true;
boost::unique_lock<boost::mutex> lock(m_lock);
if (m_method == ranked_sampling)
perform_ranked_sampling();
else if (m_method == random_sampling)
perform_random_sampling();
else UTIL_THROW2("Unsupported sampling method.");
m_finished = true;
m_ready.notify_all();
return true;
}
template<typename Token>
bool
BitextSampler<Token>::
done() const
{
return m_next == m_stop;
}
template<typename Token>
sptr<pstats>
BitextSampler<Token>::
stats()
{
// if (m_ctr == 0) (*this)();
// boost::unique_lock<boost::mutex> lock(m_lock);
// while (!m_finished)
// m_ready.wait(lock);
return m_stats;
}
template<typename Token>
BitextSampler<Token>::
~BitextSampler()
{
m_stats->release();
}
} // end of namespace bitext
} // end of namespace Moses
|