File: | hmm.cc |
Warning: | line 693, column 10 Although the value stored to 'nw' is used in the enclosing expression, the value is never actually read from 'nw' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | |
2 | /* |
3 | * Copyright (C) 2005 Universitat d'Alacant / Universidad de Alicante |
4 | * |
5 | * This program is free software; you can redistribute it and/or |
6 | * modify it under the terms of the GNU General Public License as |
7 | * published by the Free Software Foundation; either version 2 of the |
8 | * License, or (at your option) any later version. |
9 | * |
10 | * This program is distributed in the hope that it will be useful, but |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 | * General Public License for more details. |
14 | * |
15 | * You should have received a copy of the GNU General Public License |
16 | * along with this program; if not, see <https://www.gnu.org/licenses/>. |
17 | */ |
18 | /* |
19 | * First order hidden Markov model (HMM) implementation (source) |
20 | * |
21 | * @author Felipe Sánchez-Martínez - fsanchez@dlsi.ua.es |
22 | */ |
23 | |
24 | #include <apertium/hmm.h> |
25 | #include "apertium_config.h" |
26 | #include <apertium/unlocked_cstdio.h> |
27 | #include <lttoolbox/compression.h> |
28 | |
29 | #include <stdio.h> |
30 | #include <unistd.h> |
31 | #include <vector> |
32 | #include <algorithm> |
33 | #include <lttoolbox/string_utils.h> |
34 | #include <apertium/file_morpho_stream.h> |
35 | |
36 | inline bool p_isnan(double v) { |
37 | #if __cplusplus202101L >= 201103L |
38 | return std::isnan(v); |
39 | #else |
40 | return ::isnan(v); |
41 | #endif |
42 | } |
43 | |
44 | inline bool p_isinf(double v) { |
45 | #if __cplusplus202101L >= 201103L |
46 | return std::isinf(v); |
47 | #else |
48 | return ::isinf(v); |
49 | #endif |
50 | } |
51 | |
52 | using namespace Apertium; |
53 | using namespace tagger_utils; |
54 | |
55 | TaggerData& HMM::get_tagger_data() { |
56 | return tdhmm; |
57 | } |
58 | |
59 | void HMM::deserialise(FILE *Serialised_FILE_Tagger) { |
60 | tdhmm.read(Serialised_FILE_Tagger); |
61 | eos = (tdhmm.getTagIndex())["TAG_SENT"_u]; |
62 | } |
63 | |
64 | std::vector<UString> &HMM::getArrayTags() { |
65 | return tdhmm.getArrayTags(); |
66 | } |
67 | |
68 | void HMM::serialise(FILE *Stream_) { tdhmm.write(Stream_); } |
69 | |
70 | void HMM::deserialise(const TaggerData &Deserialised_FILE_Tagger) { |
71 | tdhmm = TaggerDataHMM(Deserialised_FILE_Tagger); |
72 | eos = (tdhmm.getTagIndex())["TAG_SENT"_u]; |
73 | } |
74 | |
75 | void HMM::init_probabilities_from_tagged_text_(MorphoStream &stream_tagged, |
76 | MorphoStream &stream_untagged) { |
77 | init_probabilities_from_tagged_text(stream_tagged, stream_untagged); |
78 | apply_rules(); |
79 | } |
80 | |
81 | void HMM::init_probabilities_kupiec_(MorphoStream &lexmorfo) { |
82 | init_probabilities_kupiec(lexmorfo); |
83 | apply_rules(); |
84 | } |
85 | |
86 | void HMM::train(MorphoStream &morpho_stream, unsigned long count) { |
87 | for (; count > 0; --count) { |
88 | morpho_stream.rewind(); |
89 | train(morpho_stream); |
90 | } |
91 | |
92 | apply_rules(); |
93 | } |
94 | |
95 | HMM::HMM() {} |
96 | |
97 | HMM::HMM(TaggerFlags& Flags_) : FILE_Tagger(Flags_) {} |
98 | |
99 | HMM::HMM(TaggerDataHMM _tdhmm) |
100 | : tdhmm(_tdhmm) |
101 | { |
102 | eos = (tdhmm.getTagIndex())["TAG_SENT"_u]; |
103 | } |
104 | |
105 | HMM::HMM(TaggerDataHMM *tdhmm) : tdhmm(*tdhmm) {} |
106 | |
107 | HMM::~HMM() {} |
108 | |
109 | void |
110 | HMM::init() |
111 | { |
112 | } |
113 | |
114 | void |
115 | HMM::set_eos(TTag t) |
116 | { |
117 | eos = t; |
118 | } |
119 | |
120 | void |
121 | HMM::read_ambiguity_classes(FILE *in) |
122 | { |
123 | while(in) |
124 | { |
125 | int ntags = Compression::multibyte_read(in); |
126 | |
127 | if(feof(in)) |
128 | { |
129 | break; |
130 | } |
131 | set<TTag> ambiguity_class; |
132 | |
133 | for(; ntags != 0; ntags--) |
134 | { |
135 | ambiguity_class.insert(Compression::multibyte_read(in)); |
136 | } |
137 | |
138 | if(ambiguity_class.size() != 0) |
139 | { |
140 | tdhmm.getOutput().add(ambiguity_class); |
141 | } |
142 | } |
143 | |
144 | tdhmm.setProbabilities(tdhmm.getTagIndex().size(), tdhmm.getOutput().size()); |
145 | } |
146 | |
147 | void |
148 | HMM::write_ambiguity_classes(FILE *out) |
149 | { |
150 | for(int i=0, limit = tdhmm.getOutput().size(); i != limit; i++) |
151 | { |
152 | set<TTag> const &ac = (tdhmm.getOutput())[i]; |
153 | Compression::multibyte_write(ac.size(), out); |
154 | for(set<TTag>::const_iterator it = ac.begin(), limit2 = ac.end(); |
155 | it != limit2; it++) |
156 | { |
157 | Compression::multibyte_write(*it, out); |
158 | } |
159 | } |
160 | } |
161 | |
162 | void |
163 | HMM::read_probabilities(FILE *in) |
164 | { |
165 | tdhmm.read(in); |
166 | } |
167 | |
168 | void |
169 | HMM::write_probabilities(FILE *out) |
170 | { |
171 | tdhmm.write(out); |
172 | } |
173 | |
174 | void |
175 | HMM::init_probabilities_kupiec(MorphoStream &lexmorfo) |
176 | { |
177 | int N = tdhmm.getN(); |
178 | int M = tdhmm.getM(); |
179 | int i=0, j=0, k=0, k1=0, k2=0, nw=0; |
180 | vector <double> classes_ocurrences (M, 1); |
181 | vector <vector <double> > classes_pair_ocurrences(M, vector<double>(M, 1)); |
182 | vector <double> tags_estimate(N, 0); |
183 | vector <vector <double> > tags_pair_estimate(N, vector<double>(N, 0)); |
184 | |
185 | Collection &output = tdhmm.getOutput(); |
186 | |
187 | TaggerWord *word=NULL__null; |
188 | |
189 | set<TTag> tags; |
190 | tags.insert(eos); |
191 | k1=output[tags]; //The first tag (ambiguity class) seen is the end-of-sentence |
192 | |
193 | //We count for each ambiguity class the number of ocurrences |
194 | word = lexmorfo.get_next_word(); |
195 | while((word)) { |
196 | if (++nw%10000==0) cerr<<'.'<<flush; |
197 | |
198 | tags=word->get_tags(); |
199 | |
200 | if (tags.size()==0) { //This is an unknown word |
201 | tags = tdhmm.getOpenClass(); |
202 | } |
203 | else { |
204 | require_ambiguity_class(tdhmm, tags, *word, nw); |
205 | } |
206 | |
207 | k2=output[tags]; |
208 | |
209 | classes_ocurrences[k1]++; |
210 | classes_pair_ocurrences[k1][k2]++; //k1 followed by k2 |
211 | delete word; |
212 | word=lexmorfo.get_next_word(); |
213 | |
214 | k1=k2; |
215 | |
216 | } |
217 | |
218 | //Estimation of the number of time each tags occurs in the training text |
219 | for(i=0; i<N; i++) { |
220 | for(k=0; k<M; k++) { |
221 | |
222 | if(output[k].find(i) != output[k].end()) |
223 | tags_estimate[i] += classes_ocurrences[k]/output[k].size(); |
224 | } |
225 | } |
226 | |
227 | set<TTag> tags1, tags2; |
228 | set<TTag>::iterator itag1, itag2; |
229 | for(k1=0; k1<M; k1++) { |
230 | tags1=output[k1]; |
231 | for(k2=0; k2<M; k2++) { |
232 | tags2=output[k2]; |
233 | double nocurrences=classes_pair_ocurrences[k1][k2]/((double)(tags1.size()*tags2.size())); |
234 | for (itag1=tags1.begin(); itag1!=tags1.end(); itag1++) { |
235 | for (itag2=tags2.begin(); itag2!=tags2.end(); itag2++) |
236 | tags_pair_estimate[*itag1][*itag2]+=nocurrences; |
237 | } |
238 | } |
239 | } |
240 | |
241 | //a[i][j] estimation. |
242 | double sum; |
243 | for(i=0; i<N; i++) { |
244 | sum=0; |
245 | for(j=0; j<N; j++) |
246 | sum+=tags_pair_estimate[i][j]; |
247 | |
248 | for(j=0; j<N; j++) { |
249 | if (sum>0) |
250 | (tdhmm.getA())[i][j] = tags_pair_estimate[i][j]/sum; |
251 | else { |
252 | (tdhmm.getA())[i][j] = 0; |
253 | } |
254 | } |
255 | } |
256 | |
257 | //b[i][k] estimation |
258 | for(i=0; i<N; i++) { |
259 | for(k=0; k<M; k++) { |
260 | if (output[k].find(i)!=output[k].end()) { |
261 | if (tags_estimate[i]>0) |
262 | (tdhmm.getB())[i][k] = (classes_ocurrences[k]/output[k].size())/tags_estimate[i]; |
263 | else |
264 | (tdhmm.getB())[i][k] = 0; |
265 | } |
266 | } |
267 | } |
268 | cerr<<"\n"; |
269 | } |
270 | |
271 | |
272 | void |
273 | HMM::init_probabilities_from_tagged_text(MorphoStream &stream_tagged, |
274 | MorphoStream &stream_untagged) { |
275 | int i, j, k, nw=0; |
276 | int N = tdhmm.getN(); |
277 | int M = tdhmm.getM(); |
278 | vector <vector <double> > tags_pair(N, vector<double>(N, 0)); |
279 | vector <vector <double> > emission(N, vector<double>(M, 0)); |
280 | |
281 | |
282 | TaggerWord *word_tagged=NULL__null, *word_untagged=NULL__null; |
283 | Collection &output = tdhmm.getOutput(); |
284 | |
285 | |
286 | set<TTag> tags; |
287 | |
288 | TTag tag1, tag2; |
289 | tag1 = eos; // The first seen tag is the end-of-sentence tag |
290 | |
291 | word_tagged = stream_tagged.get_next_word(); |
292 | word_untagged = stream_untagged.get_next_word(); |
293 | while(word_tagged) { |
294 | cerr<<*word_tagged; |
295 | cerr<<" -- "<<*word_untagged<<"\n"; |
296 | |
297 | if (word_tagged->get_superficial_form()!=word_untagged->get_superficial_form()) { |
298 | cerr<<"\nTagged text (.tagged) and analyzed text (.untagged) streams are not aligned.\n"; |
299 | cerr<<"Take a look at tagged text (.tagged).\n"; |
300 | cerr<<"Perhaps this is caused by a multiword unit that is not a multiword unit in one of the two files.\n"; |
301 | cerr<<*word_tagged<<" -- "<<*word_untagged<<"\n"; |
302 | exit(1); |
303 | } |
304 | |
305 | if (++nw%100==0) cerr<<'.'<<flush; |
306 | |
307 | tag2 = tag1; |
308 | |
309 | if (word_untagged==NULL__null) { |
310 | cerr<<"word_untagged==NULL\n"; |
311 | exit(1); |
312 | } |
313 | |
314 | if (word_tagged->get_tags().size()==0) // Unknown word |
315 | tag1 = -1; |
316 | else if (word_tagged->get_tags().size()>1) // Ambiguous word |
317 | cerr<<"Error in tagged text. An ambiguous word was found: "<<word_tagged->get_superficial_form()<<"\n"; |
318 | else |
319 | tag1 = *(word_tagged->get_tags()).begin(); |
320 | |
321 | |
322 | if ((tag1>=0) && (tag2>=0)) |
323 | tags_pair[tag2][tag1]++; |
324 | |
325 | |
326 | if (word_untagged->get_tags().size()==0) { // Unknown word |
327 | tags = tdhmm.getOpenClass(); |
328 | } |
329 | else { |
330 | require_ambiguity_class(tdhmm, word_untagged->get_tags(), *word_untagged, nw); |
331 | tags = word_untagged->get_tags(); |
332 | } |
333 | |
334 | k=output[tags]; |
335 | if(tag1>=0) |
336 | emission[tag1][k]++; |
337 | |
338 | delete word_tagged; |
339 | word_tagged=stream_tagged.get_next_word(); |
340 | delete word_untagged; |
341 | word_untagged=stream_untagged.get_next_word(); |
342 | } |
343 | |
344 | |
345 | //Estimate of a[i][j] |
346 | for(i=0; i<N; i++) { |
347 | double sum=0; |
348 | for(j=0; j<N; j++) |
349 | sum += tags_pair[i][j]+1.0; |
350 | for(j=0; j<N; j++) |
351 | (tdhmm.getA())[i][j] = (tags_pair[i][j]+1.0)/sum; |
352 | } |
353 | |
354 | |
355 | //Estimate of b[i][k] |
356 | for(i=0; i<N; i++) { |
357 | int nclasses_appear=0; |
358 | double times_appear=0.0; |
359 | for(k=0; k<M; k++) { |
360 | if (output[k].find(i)!=output[k].end()) { |
361 | nclasses_appear++; |
362 | times_appear+=emission[i][k]; |
363 | } |
364 | } |
365 | for(k=0; k<M; k++) { |
366 | if (output[k].find(i)!=output[k].end()) |
367 | (tdhmm.getB())[i][k] = (emission[i][k]+(((double)1.0)/((double)nclasses_appear)))/(times_appear+((double)1.0)); |
368 | } |
369 | } |
370 | |
371 | cerr<<"\n"; |
372 | } |
373 | |
374 | void |
375 | HMM::apply_rules() |
376 | { |
377 | vector<TForbidRule> &forbid_rules = tdhmm.getForbidRules(); |
378 | vector<TEnforceAfterRule> &enforce_rules = tdhmm.getEnforceRules(); |
379 | int N = tdhmm.getN(); |
380 | int i, j, j2; |
381 | bool found; |
382 | |
383 | for(i=0; i<(int) forbid_rules.size(); i++) { |
384 | (tdhmm.getA())[forbid_rules[i].tagi][forbid_rules[i].tagj] = ZERO1e-10; |
385 | } |
386 | |
387 | for(i=0; i<(int) enforce_rules.size(); i++) { |
388 | for(j=0; j<N; j++) { |
389 | found = false; |
390 | for (j2=0; j2<(int) enforce_rules[i].tagsj.size(); j2++) { |
391 | if (enforce_rules[i].tagsj[j2]==j) { |
392 | found = true; |
393 | break; |
394 | } |
395 | } |
396 | if (!found) |
397 | (tdhmm.getA())[enforce_rules[i].tagi][j] = ZERO1e-10; |
398 | } |
399 | } |
400 | |
401 | // Normalize probabilities |
402 | for(i=0; i<N; i++) { |
403 | double sum=0; |
404 | for(j=0; j<N; j++) |
405 | sum += (tdhmm.getA())[i][j]; |
406 | for(j=0; j<N; j++) { |
407 | if (sum>0) |
408 | (tdhmm.getA())[i][j] = (tdhmm.getA())[i][j]/sum; |
409 | else |
410 | (tdhmm.getA())[i][j] = 0; |
411 | } |
412 | } |
413 | } |
414 | |
415 | void |
416 | HMM::post_ambg_class_scan() { |
417 | int N = (tdhmm.getTagIndex()).size(); |
418 | int M = (tdhmm.getOutput()).size(); |
419 | cerr << N << " states and " << M <<" ambiguity classes\n"; |
420 | |
421 | tdhmm.setProbabilities(N, M); |
422 | } |
423 | |
424 | void |
425 | HMM::filter_ambiguity_classes(const char* input_file, UFILE* out) { |
426 | set<set<TTag> > ambiguity_classes; |
427 | FileMorphoStream morpho_stream(input_file, true, &tdhmm); |
428 | |
429 | TaggerWord *word = morpho_stream.get_next_word(); |
430 | |
431 | while(word) { |
432 | set<TTag> tags = word->get_tags(); |
433 | if(tags.size() > 0) { |
434 | if(ambiguity_classes.find(tags) == ambiguity_classes.end()) { |
435 | ambiguity_classes.insert(tags); |
436 | word->outputOriginal(out); |
437 | //cerr<<word->get_string_tags()<<"\n"; |
438 | } |
439 | } |
440 | delete word; |
441 | word = morpho_stream.get_next_word(); |
442 | } |
443 | } |
444 | |
445 | void |
446 | HMM::train(MorphoStream &morpho_stream) { |
447 | int i, j, k, t, len, nw = 0; |
448 | TaggerWord *word=NULL__null; |
449 | TTag tag; |
450 | set<TTag> tags, pretags; |
451 | set<TTag>::iterator itag, jtag; |
452 | map <int, double> gamma; |
453 | map <int, double>::iterator jt, kt; |
454 | map < int, map <int, double> > alpha, beta, xsi, phi; |
455 | map < int, map <int, double> >::iterator it; |
456 | double prob, loli; |
457 | vector < set<TTag> > pending; |
458 | Collection &output = tdhmm.getOutput(); |
459 | |
460 | int ndesconocidas=0; |
461 | // alpha => forward probabilities |
462 | // beta => backward probabilities |
463 | |
464 | loli = 0; |
465 | tag = eos; |
466 | tags.clear(); |
467 | tags.insert(tag); |
468 | pending.push_back(tags); |
469 | |
470 | alpha[0].clear(); |
471 | alpha[0][tag] = 1; |
472 | |
473 | word = morpho_stream.get_next_word(); |
474 | |
475 | while (word) { |
476 | |
477 | //cerr<<"Enter para continuar\n"; |
478 | //getchar(); |
479 | |
480 | if (++nw%10000==0) cerr<<'.'<<flush; |
481 | |
482 | //cerr<<*word<<"\n"; |
483 | |
484 | pretags = pending.back(); |
485 | |
486 | tags = word->get_tags(); |
487 | |
488 | if (tags.size()==0) { // This is an unknown word |
489 | tags = tdhmm.getOpenClass(); |
490 | ndesconocidas++; |
491 | } |
492 | |
493 | require_ambiguity_class(tdhmm, tags, *word, nw); |
494 | |
495 | k = output[tags]; |
496 | len = pending.size(); |
497 | alpha[len].clear(); |
498 | |
499 | //Forward probabilities |
500 | for (itag=tags.begin(); itag!=tags.end(); itag++) { |
501 | i=*itag; |
502 | for (jtag=pretags.begin(); jtag!=pretags.end(); jtag++) { |
503 | j=*jtag; |
504 | //cerr<<"previous alpha["<<len<<"]["<<i<<"]="<<alpha[len][i]<<"\n"; |
505 | //cerr<<"alpha["<<len-1<<"]["<<j<<"]="<<alpha[len-1][j]<<"\n"; |
506 | //cerr<<"a["<<j<<"]["<<i<<"]="<<a[j][i]<<"\n"; |
507 | //cerr<<"b["<<i<<"]["<<k<<"]="<<b[i][k]<<"\n"; |
508 | alpha[len][i] += alpha[len-1][j]*(tdhmm.getA())[j][i]*(tdhmm.getB())[i][k]; |
509 | } |
510 | if (alpha[len][i]==0) |
511 | alpha[len][i]=DBL_MIN2.2250738585072014e-308; |
512 | //cerr<<"alpha["<<len<<"]["<<i<<"]="<<alpha[len][i]<<"\n--------\n"; |
513 | } |
514 | |
515 | if (tags.size()>1) { |
516 | pending.push_back(tags); |
517 | } else { // word is unambiguous |
518 | tag = *tags.begin(); |
519 | beta[0].clear(); |
520 | beta[0][tag] = 1; |
521 | |
522 | prob = alpha[len][tag]; |
523 | |
524 | //cerr<<"prob="<<prob<<"\n"; |
525 | //cerr<<"alpha["<<len<<"]["<<tag<<"]="<<alpha[len][tag]<<"\n"; |
526 | loli -= log(prob); |
527 | |
528 | for (t=0; t<len; t++) { // loop from T-1 to 0 |
529 | pretags = pending.back(); |
530 | pending.pop_back(); |
531 | k = output[tags]; |
532 | beta[1-t%2].clear(); |
533 | for (itag=tags.begin(); itag!=tags.end(); itag++) { |
534 | i=*itag; |
535 | for (jtag=pretags.begin(); jtag!=pretags.end(); jtag++) { |
536 | j = *jtag; |
537 | beta[1-t%2][j] += (tdhmm.getA())[j][i]*(tdhmm.getB())[i][k]*beta[t%2][i]; |
538 | xsi[j][i] += alpha[len-t-1][j]*(tdhmm.getA())[j][i]*(tdhmm.getB())[i][k]*beta[t%2][i]/prob; |
539 | } |
540 | double previous_value = gamma[i]; |
541 | |
542 | gamma[i] += alpha[len-t][i]*beta[t%2][i]/prob; |
543 | if (p_isnan(gamma[i])) { |
544 | cerr<<"NAN(3) gamma["<<i<<"] = "<<gamma[i]<<" alpha["<<len-t<<"]["<<i<<"]= "<<alpha[len-t][i] |
545 | <<" beta["<<t%2<<"]["<<i<<"] = "<<beta[t%2][i]<<" prob = "<<prob<<" previous gamma = "<<previous_value<<"\n"; |
546 | exit(1); |
547 | } |
548 | if (p_isinf(gamma[i])) { |
549 | cerr<<"INF(3) gamma["<<i<<"] = "<<gamma[i]<<" alpha["<<len-t<<"]["<<i<<"]= "<<alpha[len-t][i] |
550 | <<" beta["<<t%2<<"]["<<i<<"] = "<<beta[t%2][i]<<" prob = "<<prob<<" previous gamma = "<<previous_value<<"\n"; |
551 | exit(1); |
552 | } |
553 | if (gamma[i]==0) { |
554 | //cout<<"ZERO(3) gamma["<<i<<"] = "<<gamma[i]<<" alpha["<<len-t<<"]["<<i<<"]= "<<alpha[len-t][i] |
555 | // <<" beta["<<t%2<<"]["<<i<<"] = "<<beta[t%2][i]<<" prob = "<<prob<<" previous gamma = "<<previous_value<<"\n"; |
556 | gamma[i]=DBL_MIN2.2250738585072014e-308; |
557 | //exit(1); |
558 | } |
559 | phi[i][k] += alpha[len-t][i]*beta[t%2][i]/prob; |
560 | } |
561 | tags=pretags; |
562 | } |
563 | |
564 | tags.clear(); |
565 | tags.insert(tag); |
566 | pending.push_back(tags); |
567 | alpha[0].clear(); |
568 | alpha[0][tag] = 1; |
569 | } |
570 | |
571 | delete word; |
572 | word = morpho_stream.get_next_word(); |
573 | } |
574 | |
575 | if ((pending.size()>1) || ((tag!=eos)&&(tag != (tdhmm.getTagIndex())["TAG_kEOF"_u]))) { |
576 | cerr << "Warning: The last tag is not the end-of-sentence-tag " |
577 | << "but rather " << tdhmm.getArrayTags()[tag] << ". Line: " << nw |
578 | << ". Pending: " << pending.size() << ". Tags: "; |
579 | cerr << "\n"; |
580 | } |
581 | |
582 | int N = tdhmm.getN(); |
583 | int M = tdhmm.getM(); |
584 | |
585 | //Clean previous values |
586 | for(i=0; i<N; i++) { |
587 | for(j=0; j<N; j++) |
588 | (tdhmm.getA())[i][j]=ZERO1e-10; |
589 | for(k=0; k<M; k++) |
590 | (tdhmm.getB())[i][k]=ZERO1e-10; |
591 | } |
592 | |
593 | // new parameters |
594 | for (it=xsi.begin(); it!=xsi.end(); it++) { |
595 | i = it->first; |
596 | for (jt=xsi[i].begin(); jt!=xsi[i].end(); jt++) { |
597 | j = jt->first; |
598 | if (xsi[i][j]>0) { |
599 | if (gamma[i]==0) { |
600 | cerr<<"Warning: gamma["<<i<<"]=0\n"; |
601 | gamma[i]=DBL_MIN2.2250738585072014e-308; |
602 | } |
603 | |
604 | (tdhmm.getA())[i][j] = xsi[i][j]/gamma[i]; |
605 | |
606 | if (p_isnan((tdhmm.getA())[i][j])) { |
607 | cerr<<"NAN\n"; |
608 | cerr <<"Error: BW - NAN(1) a["<<i<<"]["<<j<<"]="<<(tdhmm.getA())[i][j]<<"\txsi["<<i<<"]["<<j<<"]="<<xsi[i][j]<<"\tgamma["<<i<<"]="<<gamma[i]<<"\n"; |
609 | exit(1); |
610 | } |
611 | if (p_isinf((tdhmm.getA())[i][j])) { |
612 | cerr<<"INF\n"; |
613 | cerr <<"Error: BW - INF(1) a["<<i<<"]["<<j<<"]="<<(tdhmm.getA())[i][j]<<"\txsi["<<i<<"]["<<j<<"]="<<xsi[i][j]<<"\tgamma["<<i<<"]="<<gamma[i]<<"\n"; |
614 | exit(1); |
615 | } |
616 | if ((tdhmm.getA())[i][j]==0) { |
617 | //cerr <<"Error: BW - ZERO(1) a["<<i<<"]["<<j<<"]="<<(tdhmm.getA())[i][j]<<"\txsi["<<i<<"]["<<j<<"]="<<xsi[i][j]<<"\tgamma["<<i<<"]="<<gamma[i]<<"\n"; |
618 | // exit(1); |
619 | } |
620 | } |
621 | } |
622 | } |
623 | |
624 | for (it=phi.begin(); it!=phi.end(); it++) { |
625 | i = it->first; |
626 | for (kt=phi[i].begin(); kt!=phi[i].end(); kt++) { |
627 | k = kt->first; |
628 | if (phi[i][k]>0) { |
629 | (tdhmm.getB())[i][k] = phi[i][k]/gamma[i]; |
630 | |
631 | if (p_isnan((tdhmm.getB())[i][k])) { |
632 | cerr<<"Error: BW - NAN(2) b["<<i<<"]["<<k<<"]="<<(tdhmm.getB())[i][k]<<"\tphi["<<i<<"]["<<k<<"]="<<phi[i][k]<<"\tgamma["<<i<<"]="<<gamma[i]<<"\n"; |
633 | exit(1); |
634 | } |
635 | if (p_isinf((tdhmm.getB())[i][k])) { |
636 | cerr<<"Error: BW - INF(2) b["<<i<<"]["<<k<<"]="<<(tdhmm.getB())[i][k]<<"\tphi["<<i<<"]["<<k<<"]="<<phi[i][k]<<"\tgamma["<<i<<"]="<<gamma[i]<<"\n"; |
637 | exit(1); |
638 | } |
639 | if ((tdhmm.getB())[i][k]==0) { |
640 | //cerr <<"Error: BW - ZERO(2) b["<<i<<"]["<<k<<"]="<<(tdhmm.getB())[i][k]<<"\tphi["<<i<<"]["<<k<<"]="<<phi[i][k]<<"\tgamma["<<i<<"]="<<gamma[i]<<"\n"; |
641 | // exit(1); |
642 | } |
643 | } |
644 | } |
645 | } |
646 | |
647 | //It can be possible that a probability is not updated |
648 | //We normalize the probabilitites |
649 | for(i=0; i<N; i++) { |
650 | double sum=0; |
651 | for(j=0; j<N; j++) |
652 | sum+=(tdhmm.getA())[i][j]; |
653 | for(j=0; j<N; j++) |
654 | (tdhmm.getA())[i][j]=(tdhmm.getA())[i][j]/sum; |
655 | } |
656 | |
657 | for(i=0; i<N; i++) { |
658 | double sum=0; |
659 | for(k=0; k<M; k++) { |
660 | if(output[k].find(i)!=output[k].end()) |
661 | sum+=(tdhmm.getB())[i][k]; |
662 | } |
663 | for(k=0; k<M; k++) { |
664 | if(output[k].find(i)!=output[k].end()) |
665 | (tdhmm.getB())[i][k]=(tdhmm.getB())[i][k]/sum; |
666 | } |
667 | } |
668 | |
669 | cerr<<"Log="<<loli<<"\n"; |
670 | } |
671 | |
672 | void |
673 | HMM::tagger(MorphoStream &morpho_stream, UFILE* Output) { |
674 | int i, j, k, nw; |
675 | TaggerWord *word = NULL__null; |
676 | TTag tag; |
677 | |
678 | set <TTag> ambg_class_tags, tags, pretags; |
679 | set <TTag>::iterator itag, jtag; |
680 | |
681 | double prob, loli, x; |
682 | int N = tdhmm.getN(); |
683 | vector <vector <double> > alpha(2, vector<double>(N)); |
684 | vector <vector <vector<TTag> > > best(2, vector <vector <TTag> >(N)); |
685 | |
686 | vector <TaggerWord> wpend; |
687 | int nwpend; |
688 | |
689 | morpho_stream.setNullFlush(TheFlags.getNullFlush()); |
690 | |
691 | Collection &output = tdhmm.getOutput(); |
692 | |
693 | loli = nw = 0; |
Although the value stored to 'nw' is used in the enclosing expression, the value is never actually read from 'nw' | |
694 | |
695 | //Initialization |
696 | tags.insert(eos); |
697 | alpha[0][eos] = 1; |
698 | |
699 | word = morpho_stream.get_next_word(); |
700 | |
701 | while (word) { |
702 | wpend.push_back(*word); |
703 | nwpend = wpend.size(); |
704 | |
705 | pretags = tags; // Tags from the previous word |
706 | |
707 | tags = word->get_tags(); |
708 | |
709 | if (tags.size()==0) // This is an unknown word |
710 | tags = tdhmm.getOpenClass(); |
711 | |
712 | ambg_class_tags = require_similar_ambiguity_class(tdhmm, tags, *word, TheFlags.getDebug()); |
713 | |
714 | k = output[ambg_class_tags]; //Ambiguity class the word belongs to |
715 | |
716 | clear_array_double(&alpha[nwpend%2][0], N); |
717 | clear_array_vector(&best[nwpend%2][0], N); |
718 | |
719 | //Induction |
720 | for (itag=tags.begin(); itag!=tags.end(); itag++) { //For all tag from the current word |
721 | i=*itag; |
722 | for (jtag=pretags.begin(); jtag!=pretags.end(); jtag++) { //For all tags from the previous word |
723 | j=*jtag; |
724 | x = alpha[1-nwpend%2][j]*(tdhmm.getA())[j][i]*(tdhmm.getB())[i][k]; |
725 | if (alpha[nwpend%2][i]<=x) { |
726 | if (nwpend>1) |
727 | best[nwpend%2][i] = best[1-nwpend%2][j]; |
728 | best[nwpend%2][i].push_back(i); |
729 | alpha[nwpend%2][i] = x; |
730 | } |
731 | } |
732 | } |
733 | |
734 | //Backtracking |
735 | if (tags.size() == 1) { |
736 | tag = *tags.begin(); |
737 | prob = alpha[nwpend%2][tag]; |
738 | |
739 | if (prob>0) |
740 | loli -= log(prob); |
741 | else { |
742 | if (TheFlags.getDebug()) |
743 | cerr<<"Problem with word '"<<word->get_superficial_form()<<"' "<<word->get_string_tags()<<"\n"; |
744 | } |
745 | for (unsigned t=0; t<best[nwpend%2][tag].size(); t++) { |
746 | if (TheFlags.getFirst()) { |
747 | UString const &micad = wpend[t].get_all_chosen_tag_first(best[nwpend%2][tag][t], (tdhmm.getTagIndex())["TAG_kEOF"_u]); |
748 | write(micad, Output); |
749 | } else { |
750 | // print Output |
751 | wpend[t].set_show_sf(TheFlags.getShowSuperficial()); |
752 | UString const &micad = wpend[t].get_lexical_form(best[nwpend%2][tag][t], (tdhmm.getTagIndex())["TAG_kEOF"_u]); |
753 | write(micad, Output); |
754 | } |
755 | } |
756 | |
757 | //Return to the initial state |
758 | wpend.clear(); |
759 | alpha[0][tag] = 1; |
760 | } |
761 | |
762 | delete word; |
763 | |
764 | if(morpho_stream.getEndOfFile()) |
765 | { |
766 | if(TheFlags.getNullFlush()) |
767 | { |
768 | u_fputcu_fputc_72('\0', Output); |
769 | tags.clear(); |
770 | tags.insert(eos); |
771 | alpha[0][eos] = 1; |
772 | } |
773 | |
774 | u_fflushu_fflush_72(Output); |
775 | morpho_stream.setEndOfFile(false); |
776 | } |
777 | word = morpho_stream.get_next_word(); |
778 | } |
779 | |
780 | if ((tags.size()>1)&&(TheFlags.getDebug())) { |
781 | cerr << "Error: The text to disambiguate has finished, but there are ambiguous words that has not been disambiguated.\n"; |
782 | cerr << "This message should never appears. If you are reading this ..... these are very bad news.\n"; |
783 | } |
784 | } |
785 | |
786 | |
787 | void |
788 | HMM::print_A() { |
789 | int i,j; |
790 | |
791 | cout<<"TRANSITION MATRIX (A)\n------------------------------\n"; |
792 | for(i=0; i != tdhmm.getN(); i++) |
793 | for(j=0; j != tdhmm.getN(); j++) { |
794 | cout<<"A["<<i<<"]["<<j<<"] = "<<(tdhmm.getA())[i][j]<<"\n"; |
795 | } |
796 | } |
797 | |
798 | void |
799 | HMM::print_B() { |
800 | int i,k; |
801 | |
802 | cout<<"EMISSION MATRIX (B)\n-------------------------------\n"; |
803 | for(i=0; i != tdhmm.getN(); i++) |
804 | for(k=0; k != tdhmm.getM(); k++) { |
805 | Collection &output = tdhmm.getOutput(); |
806 | if(output[k].find(i)!=output[k].end()) |
807 | cout<<"B["<<i<<"]["<<k<<"] = "<<(tdhmm.getB())[i][k]<<"\n"; |
808 | } |
809 | } |
810 | |
811 | void HMM::print_ambiguity_classes() { |
812 | set<TTag> ambiguity_class; |
813 | set<TTag>::iterator itag; |
814 | cout<<"AMBIGUITY CLASSES\n-------------------------------\n"; |
815 | for(int i=0; i != tdhmm.getM(); i++) { |
816 | ambiguity_class = (tdhmm.getOutput())[i]; |
817 | cout <<i<<": "; |
818 | for (itag=ambiguity_class.begin(); itag!=ambiguity_class.end(); itag++) { |
819 | cout << *itag <<" "; |
820 | } |
821 | cout << "\n"; |
822 | } |
823 | } |