-
Notifications
You must be signed in to change notification settings - Fork 0
/
scfgLearner.py
1306 lines (1134 loc) · 37.4 KB
/
scfgLearner.py
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
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from __future__ import absolute_import, division, print_function
'''""""""""""""""""""""""""""""""""""""""""""""""""""""""""
Learn Stochastic Context-Free Grammars from input with uncertainties.
Author:
Kyuhwa Lee
Imperial College of Science, Technology and Science
Notations:
NJP: Normalized Joint Probability ( JP(S)^(1/len(S)) )
V: Rule score (V= sigma(NJP(S)) * len(S))
Data structures:
input= {'symbols':[], 'values':[]}
Input stream with uncertainties.
G= {'NT':[rule1, rule2, ...], ...}
Grammar object.
DLT(global)= OrderedDict{string:{score, count, parent, terms}}
Description Length Table.
GNode= class{g, dlt, pri, lik, mdl, bestmdl, gid, worse}
Grammar node of a search tree, gList.
bestmdl: best MDL score observed so far in the current branch
worse: for beam search (worse += 1 if new_mdl > bestmdl)
self.t_stat= {string: {count, prob}}
Statistics of terminal symbols.
self.t_dic= {'a':'A', 'b':'B',...}
Global terminal list.
Basics of Merging & Substituting:
A. Stolcke, PhD Thesis, UCB, p.93-97
TODO:
- Support terminals of 2 or more characters (currently 25 terminal symbols (a-y) are supported)
- Get rid of global constants
""""""""""""""""""""""""""""""""""""""""""""""""""""""""'''
import math, os, sys, time
from collections import OrderedDict
from string import ascii_uppercase
from copy import deepcopy
from operator import itemgetter
from pathos.multiprocessing import Pool
from multiprocessing import Manager, cpu_count, current_process
import sartparser as sp
import q_common as qc
'''""""""""""""""""""""""""""""""""
Experiment settings
TODO: Split this part into a separate config file
""""""""""""""""""""""""""""""""'''
ALGORITHM= 'LEEQ' # 'LEEQ' or 'STOLCKE'
VERBOSE= 0; # min 0, max 2.
BEAMSIZE= 3 # beam search width (1= best-first search). Not recommended over 3.
PRUNE_P= 0.01 # prune a rule if low probability
TERM_P= 0.9 # sensor uncertainty to make robust grammar; prob of terminal 'x' being really 'x'
MAX_NGRAMS= 46 # maximum sub-patterns to consider (warning: too large value can lead to very slow performance)
EXPORT_INPUT= False # export input into *.seq each time for parsing
'''""""""""""""""""""""""""""""""""
Grammar node object
""""""""""""""""""""""""""""""""'''
class GNode:
def __init__(self, g,dlt,pri,lik,mdl,bestmdl,gid,worse=0):
self.g= g
self.dlt= dlt
self.pri= pri
self.lik= lik
self.mdl= mdl
self.bestmdl= bestmdl
self.gid= gid
self.worse= worse
'''""""""""""""""""""""""""""""""""
Helper Functions
""""""""""""""""""""""""""""""""'''
def uniquify(l):
"""
Remove duplicates: doesn't preserve orders
"""
return list(set(l))
'''""""""""""""""""""""""""""""""""
Main Functions
""""""""""""""""""""""""""""""""'''
class ScfgLearner:
def conv2NT(self, t):
"""
Convert input terminals into corresponding NT's
"""
str= ''
for x in t:
str += self.t_dic[x]
return str
def conv2T(self, nt):
"""
Convert preterminals to terminals
"""
str= ''
for x in nt:
str += self.t_dic_rev[x]
return str
def getPrior(self, g):
"""
Description length of prior probability P(G)
Stolcke PhD Thesis,"Bayesian learning of probabilistic language models",Sec 2.5.5
"""
# parameter prior (terminal symobls + non-terminal symbols)
num_symbols= len(self.t_dic.keys()) + len(g.keys())
dl_theta= -math.log(qc.dirichlet(num_symbols),2)
# structure prior (Poisson distribution of the grammar length)
dl_S= 0
mu= 3
for s in g:
for r in g[s]:
rlen= len(r)+1
# expected bits for (length prior + each rule length * num_symbols)
dl_S += -math.log(qc.poisson(mu,rlen),3) + rlen*math.log(num_symbols,2)
pri= dl_S+dl_theta
return pri
def grammar2sartgrammar(self, g, dlt, method=0):
"""
Convert to a SARTParser's CFGrammar object to get the likelihood.
"""
gs= sp.CFGrammar()
# Define axiom
gs.addAxiom('Z')
# Define non-terminals
for s in g.keys():
gs.addNonTerminal(s)
if method==2:
skip= ascii_uppercase[len(self.t_dic)] # add a SKIP terminal
gs.addNonTerminal(skip)
# Define terminals
for s in self.t_seq:
gs.addTerminal(s)
# Define rules
# Default method
if method==0:
for t in self.t_dic:
gs.addRule(self.t_dic[t], [t], 1.0)
# Robust method 1: A -> a|b|c|d, B -> a|b|c|d ...
elif method==1:
for nt in self.t_dic_rev:
for t in self.t_dic:
if self.t_dic[t]==nt:
rulescore= TERM_P
else:
rulescore= self.term_p_other
gs.addRule(nt, [t], rulescore)
# Robust method2: A -> a|SKIP, SKIP -> SKIP SKIP|a|b|c|d ...
elif method==2:
pskipself= 0.01
pskip= (1-pskipself) / len(self.t_dic)
gs.addRule(skip, [skip, skip], pskipself)
for t in self.t_dic:
gs.addRule(skip, t, pskip)
for nt in self.t_dic_rev:
gs.addRule(nt, [self.t_dic_rev[t]], TERM_P)
gs.addRule(nt, skip, 1-TERM_P)
# Actual grammar body
for nt in self.getNTlist(g):
sum= 0.0 # sum of all rule scores belonging to nt
for r in g[nt]: # for each rule of a non-terminal
rulescore= 0.0
for t in dlt[r]['terms']:
if t not in self.t_stat:
self.bug('%s key is not in self.t_stat'%t, g, dlt)
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
#rulescore *= len(r)
sum += rulescore
for r in g[nt]:
rulescore= 0.0 # the score of each rule
for t in dlt[r]['terms']:
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
#rulescore *= len(r)
if sum==0: # when input probabilities of all terminals are 0
gs.addRule(nt, list(r), 1.0/len(g[nt]))
else:
gs.addRule(nt, list(r), rulescore/sum)
return gs
def getLikelihood(self, g, dlt, verbose=VERBOSE):
"""
Compute likelihood using Viterbi parsing
"""
gs= self.grammar2sartgrammar(g, dlt)
parser= sp.SParser(gs)
for i in range(len(self.input_list)):
fs= open(self.testfile % i)
for s in fs:
sw= s.strip()
if len(sw)==0: continue
if sw[0] != '#':
parser.parseLine(sw)
psc= parser.getViterbiParse().probability.scaled
return min(self.max_mdl, psc)
def getMDL(self, g, dlt, verbose=False):
"""
MDL score
"""
return self.getPrior(g) + self.getLikelihood(g, dlt, verbose)
def getStringCount(self, g, strings):
"""
Count the number of appearances of given string in RHS of g
"""
c= 0
for s in g:
for r in g[s]:
c += r.count(strings)
return c
def getNTlist(self, g):
"""
Returns the list of NT's in g except terminal NT's. e.g. [Z,Y,X..]
"""
ntlist=[]
for x in g.keys():
if x not in self.t_dic.values():
ntlist.append(x)
return sorted(ntlist, reverse=True)
def ngrams(self, seq, maxw=MAX_NGRAMS):
"""
Build n-grams from seq up to n=maxw
"""
nglist = []
inlen = len(seq)
if maxw > inlen: maxw = inlen
# n-grams with 1 <= n <= maxw
for w in range(1, maxw + 1):
for x in range(inlen - w + 1):
nt = ''.join(seq[x:x + w])
if nt not in nglist:
nglist.append(nt)
return nglist
def printMsg(self, minVerbosity, *args):
if VERBOSE < minVerbosity: return
for msg in args:
print(msg, end=' ')
print()
def printInput(self, minVerbosity, inp):
"""
Print input strings
"""
if VERBOSE < minVerbosity: return
print('\n-- New Input Sequence --')
for x in range(len(inp['symbols'])):
print('%s %0.2f'% (inp['symbols'][x],inp['values'][x]))
print()
def printDLT(self, minVerbosity, dlt, msg='Description Length Table'):
"""
Print Description Length Table
"""
if VERBOSE < minVerbosity: return
print(' '*5 + '-'*15,msg,'-'*15,'<%s>'% current_process().name)
print(' SCORE COUNT PARENT STRING',' '*11,'TERMINALS')
for s in dlt:
print('%6d %6d %4s %-18s %-s' % (dlt[s]['score'],\
dlt[s]['count'], dlt[s]['parent'], s, dlt[s]['terms'][0]), end='')
for t in range(1,len(dlt[s]['terms'])):
print( ',%s' % dlt[s]['terms'][t], end='' )
print( ',%s' % dlt[s]['terms'][t], end='' )
print()
print()
def printTSTAT(self, minVerbosity, msg='Terminal Symbol Statistics'):
if VERBOSE < minVerbosity: return
print(' --',msg,'--','<%s>'% current_process().name)
print('%-5s %-8s %s' % ('COUNT','PROB','STRING') )
for s in self.t_stat:
print( '%-5d %0.6f %s' % (self.t_stat[s]['count'],self.t_stat[s]['prob'],s) )
print()
def printMDL(self, minVerbosity, g, dlt):
if VERBOSE < minVerbosity: return
pri= self.getPrior(g)
lik= self.getLikelihood(g, dlt)
print('DL_prior: %.3f'% pri )
print('DL_likelihood: %.3f'% lik )
print('MDL: %.3f'% (pri+lik) )
print()
def printGrammar(self, minVerbosity, g, dlt, msg=''):
"""
Print grammar with msg
"""
if VERBOSE < minVerbosity: return
print(msg)
for nt in self.getNTlist(g):
sum= 0.0 # sum of all rule scores belonging to nt
sortedRules= [] # grammar with each rule sorted by scores
for r in range(len(g[nt])): # for each RHS rule of a LHS symbol
rulescore= 0.0
for t in dlt[g[nt][r]]['terms']:
if t not in self.t_stat:
self.bug('%s key is not in self.t_stat'%t, g, dlt)
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
#rulescore *= len(g[nt][r])
sortedRules.append((rulescore,r))
sum += rulescore
sortedRules= sorted(sortedRules, key=itemgetter(0), reverse=True)
firstLine=True
for r in sortedRules:
rs= g[nt][r[1]]
if firstLine:
firstLine=False
print('%s -> '% nt, end='')
else:
print(' ', end='')
rulescore= 0.0 # the score of each rule
terms= ''
for t in dlt[rs]['terms']:
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
terms += t + ' '
#rulescore *= len(rs)
if sum== 0:
print('%-18s [%0.6f] (%0.2f) <%s> %s'%
(rs, 1.0, rulescore, dlt[rs]['parent'], terms))
else:
print('%-18s [%0.6f] (%0.2f) <%s> %s'%
(rs, rulescore/sum, rulescore, dlt[rs]['parent'], terms))
if dlt[rs]['parent']== None:
#self.bug('no parent', g, dlt)
pass
print()
def printGrammarSimple(self, minVerbosity, g, msg=''):
"""
Print grammar without rule probabilities
"""
if VERBOSE < minVerbosity: return
print(msg)
for nt in sorted(g.keys(), reverse=True):
firstLine=True
for rs in g[nt]:
if firstLine:
firstLine=False
print('%s '% nt, end='')
else:
print(' ', end='')
print('-> %-18s'% rs)
def printGrammarMDL(self, minVerbosity, g, dlt, msg=''):
if VERBOSE < minVerbosity: return
self.printGrammar(g, dlt, msg)
self.self.printMDL(minVerbosity, g, dlt)
def printGrammarList(self, minVerbosity, gList, msg=''):
if VERBOSE < minVerbosity: return
print('\n'+'='*60)
print(' ', msg, 'Showing the best grammars')
print('='*60)
for n in range(len(gList)):
self.printGrammar(minVerbosity, gList[n].g, gList[n].dlt, \
'#%d, MDL: %.6f, DL_pri: %.6f, DL_lik: %.6f'%\
(n+1, gList[n].mdl, gList[n].pri, gList[n].lik) )
#self.printDLT(gList[n].dlt)
print('-'*60)
def exportGrammarList(self, gList):
if not os.path.exists('results'):
os.mkdir('results')
for n in range(len(gList)):
self.exportGrammar(gList[n].g, gList[n].dlt, 'results/rank%d.grm'% (n+1), 2 )
def exportGrammar(self, g, dlt, filename, METHOD=0):
"""
Export grammar
"""
f=open(filename,'w')
f.write('Section Terminals\n')
for s in self.t_seq: f.write('%s '% s)
f.write('\n\nSection NonTerminals\n')
for s in g: f.write('%s '% s)
if METHOD==2:
skip= ascii_uppercase[len(self.t_dic)]
f.write(skip)
f.write('\n\nSection Axiom\n')
f.write('Z\n\nSection Productions\n')
# default
if METHOD==0:
for t in self.t_dic:
f.write('%s: %-20s [%0.6f]\n' % (self.t_dic[t], t, 1.0))
# Robust Input Method 1: A -> a|b|c|d, B -> a|b|c|d ...
elif METHOD==1:
for nt in self.t_dic_rev:
firstLine= True
for t in self.t_dic:
if firstLine:
firstLine= False
id='%s:'% nt
else:
id=' |'
if self.t_dic[t]==nt:
rulescore= TERM_P
else:
rulescore= self.term_p_other
f.write('%s %-20s [%0.6f]\n' % (id, t, rulescore))
# Robust Input Method2: A -> a|SKIP, SKIP -> SKIP SKIP|a|b|c|d ...
elif METHOD==2:
pskipself= 0.01
pskip= (1-pskipself) / len(self.t_dic)
f.write('%s: %s %-18s [%0.6f]\n' % (skip, skip, skip, pskipself) )
for t in self.t_dic:
f.write(' | %-20s [%0.6f]\n' % (t, pskip) )
for t in self.t_dic_rev:
f.write('%s: %-20s [%0.6f]\n' % (t, self.t_dic_rev[t], TERM_P))
f.write(' | %-20s [%0.6f]\n' % (skip, 1-TERM_P))
for nt in self.getNTlist(g):
sum= 0.0 # sum of all rule scores belonging to nt
for r in g[nt]: # for each rule of a non-terminal
rulescore= 0.0
for t in dlt[r]['terms']:
if t not in self.t_stat:
self.bug('%s key is not in self.t_stat'%t, g, dlt)
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
#rulescore *= len(r)
sum += rulescore
firstLine= True
for r in g[nt]:
if firstLine:
firstLine= False
id='%s:'% nt
else:
id=' |'
rulescore= 0.0 # the score of each rule
for t in dlt[r]['terms']:
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
#rulescore *= len(r)
if sum==0: # when input probabilities of all terminals are 0
f.write('%s %-20s [%0.6f]\n' % (id, ' '.join(list(r)), 1.0/len(g[nt])))
else:
f.write('%s %-20s [%0.6f]\n' % (id, ' '.join(list(r)), rulescore/sum))
f.close()
def exportInput(self):
"""
Export input data
"""
tlist= self.t_seq
for r in range(len(self.input_list)):
f=open(self.testfile%r, 'w')
f.write('#')
for t in tlist: f.write('%8s '% t)
f.write('\n#')
for t in tlist: f.write('%-9s'% ('-'*8) )
f.write('\n')
input= self.input_list[r]
for i in range(len(input['symbols'])):
p= input['values'][i]
li= (1-p) / (len(tlist)-1)
s= self.t_dic_rev[input['symbols'][i]]
for t in tlist:
if s==t: f.write(' %0.6f'% p)
else: f.write(' %0.6f'% li)
f.write(' # %s'% s)
f.write('\n')
f.close()
def bug(self, msg, g='', dlt=''):
sys.stderr.write('>> BUG FOUND: %s'% msg)
print('#'*80)
print('>> BUG FOUND: %s'% msg)
print('#'*80)
if g: self.printGrammarSimple(0, g, '## Raw Grammar Rules')
if dlt: self.printDLT(0, dlt)
print('#'*80)
raise RuntimeError
def buildInput(self, seq):
"""
Build input from raw data
"""
assert len(seq) % 2==0, 'You have wrong length of input sequence.'
input= {'symbols':[], 'values':[]}
for i in range(0,len(seq),2):
input['symbols'].append(self.conv2NT(seq[i]))
input['values'].append(float(seq[i+1]))
return input
def getNextNT(self, g):
"""
Return the next available NT symbol
"""
for s in self.upper:
if s not in g:
assert s not in self.t_dic.values(), 'self.getNextNT(): assertion error!'
return s
def splitTNT(self, st):
"""
Return the list of strings split by NT and chunk of T's. [NT,T,NT,NT,T,NT...]
"""
seq= []
wasNT= True
for x in st:
if x not in self.t_dic or wasNT:
seq += x
else:
seq[-1] += x
wasNT= x not in self.t_dic
return seq
def addRule(self, g, nt, string):
"""
Add a new rule that belongs to nt
"""
if nt not in g:
g[nt]= [string]
elif string not in g[nt]:
g[nt].append(string) # if not identical input
else:
#print('>> AddRule(): Ignored adding pre-existing strings',string)
pass
return g
def delRule(self, g, nt):
"""
Delete a rule that belongs to nt
"""
if nt not in g: self.bug('self.delRule(): NT(%s) is not in G!'% nt)
del g[nt]
return g
def sortRules(self, g):
"""
Sort RHS rules in dictionary order
"""
for s in self.getNTlist(g): g[s]= sorted(g[s])
return g
def sortSymbols(self, g, dlt):
"""
Sort LHS symbols in reversed alphabetical order(e.g. W, X -> Z, Y)
"""
ntlist= self.getNTlist(g)
for x in range(len(ntlist)):
if ntlist[x] != self.upper[x]:
nt_old= ntlist[x]
nt_new= self.upper[x]
self.printMsg(2, '>> Re-ordering %s to %s <%s>'% (nt_old, nt_new, current_process().name) )
for s in g:
for x in range(len(g[s])):
r= g[s][x]
if nt_old in r:
if r in dlt and dlt[r]['parent'] != None:
dlt[r]['parent']= dlt[r]['parent'].replace(nt_old, nt_new)
g[s][x]= r.replace(nt_old, nt_new)
if s==nt_old:
if nt_new in g: self.bug('self.sortSymbols(): nt_new(%s) already in G!'% nt_new)
g[nt_new]= g.pop(nt_old)
for s in dlt:
if nt_old in s:
new_s= s.replace(nt_old, nt_new)
if new_s in dlt: self.bug('self.sortSymbols(): new_s(%s) already in DLT!'% new_s)
dlt[new_s]= dlt.pop(s)
return g, dlt
def updateDLT(self, g, dlt, maxw=MAX_NGRAMS):
"""
Update DLT's count and score to reflect the current grammar
"""
# invalidate DLT
for x in dlt:
dlt[x]['count']= 0
# build n-grams
for s in self.getNTlist(g):
for r in g[s]:
nglist= self.ngrams(r, maxw)
for nt in nglist:
if nt not in dlt:
self.bug('self.updateDLT(): %s is not in DLT'% nt, g, dlt)
dlt[nt]['count'] += r.count(nt)
# sort DLT according to score
for nt in dlt:
n= dlt[nt]['count']
w= len(nt)
dlt[nt]['score']= (n-1)*(w-1)-2
##################################################################################
# THINK: THIS SHOULD BE RECONSIDERED, ALSO, WHY -1?
if dlt[nt]['parent'] != None: dlt[nt]['score'] -= 1
##################################################################################
dlt= OrderedDict(sorted(dlt.items(), key=lambda t: t[1]['score'], reverse=True))
return dlt
def updateTStat(self, input, maxw=MAX_NGRAMS):
"""
Update self.t_stat with new input (substrings up to length=maxw)
"""
inlen= len(input['symbols'])
if maxw > inlen: maxw= inlen
# n-grams with 1 <= n <= maxw
for w in range(1,maxw+1):
for x in range(inlen-w+1):
if self.algorithm=='STOLCKE':
njp= 0.0
else:
njp= 1.0
term= ''
for i in range(x,x+w):
njp *= input['values'][i]
term += self.t_dic_rev[input['symbols'][i]]
njp= math.pow(njp,1/w)
if term not in self.t_stat:
self.t_stat[term]={'count':1, 'prob':njp}
else:
t= self.t_stat[term]
self.t_stat[term]['prob']= (t['count'] * t['prob'] + njp ) / (t['count']+1)
self.t_stat[term]['count'] += 1
def inputDLT(self, input, dlt, maxw=MAX_NGRAMS):
"""
Update DLT with new input (substrings up to length=maxw)
"""
inlen= len(input['symbols'])
if maxw > inlen: maxw= inlen
# n-grams with 1 <= n <= maxw
for w in range(1,maxw+1):
for x in range(inlen-w+1):
nt= ''.join([z[0] for z in input['symbols'][x:x+w]])
term= ''
for i in range(x,x+w):
term += self.t_dic_rev[input['symbols'][i]]
if nt not in dlt:
dlt[nt]={'score':-1,'terms':[term],'count':0,'parent':None}
return dlt
def getFirstDLT(self, dlt, strings='', parent=False):
"""
Return the highest score string X in DLT with conditions:
1. parent: {False: X has no parent | True: X has parent}
2. strings: X is substring of strings; '' if don't care
"""
global PRUNE_P
lastScore= -1 # minimum score
maxItem= []
for s in dlt:
if dlt[s]['score'] < lastScore: break
#if parent and parent==(dlt[s]['parent']==None): continue
if strings != '' and s not in strings: continue
termprob= 0.0
for t in dlt[s]['terms']:
termprob += self.t_stat[t]['prob']
if termprob <= PRUNE_P: continue
lastScore= dlt[s]['score']
maxItem.append(s)
return maxItem
def getFirstDLThack(self, dlt, strings='', parent=False):
"""
Consider only limited-length words
"""
global PRUNE_P
maxItem= []
for s in dlt:
if len(s) < 2 or len(s) > 5: continue
termprob= 0.0
for t in dlt[s]['terms']:
termprob += self.t_stat[t]['prob']
if termprob <= PRUNE_P: continue
if dlt[s]['count'] > 1:
maxItem.append(s)
return maxItem
def substituteInput(self, g, r, dlt, strings, new_nt, nt='Z'):
"""
Substitute symbols in input and update DLT
"""
if VERBOSE: print( '>> SUBSTITUTE input (%s)=%s'% (strings,new_nt) )
assert nt in g
assert nt in dlt
assert new_nt in g
assert new_nt in dlt
oldRule= g[nt][r]
assert strings in oldRule
assert strings in dlt
ref= dlt[oldRule]
newRule= oldRule.replace(strings, new_nt)
g[nt][r]= oldRule.replace(strings, '_') # will be changed later; this is important
if newRule not in dlt:
dlt[newRule]= {'terms':deepcopy(ref['terms']),
'score':ref['score'], 'count':ref['count'], 'parent':ref['parent']}
else:
dlt[newRule]['parent']= ref['parent']
dlt[oldRule]['parent']= None # it's broken now
# build new n-grams including new NT
for ngram in self.ngrams(g[nt][r]): # for each sub-pattern
newgram= ngram.replace('_', new_nt)
if newgram in dlt: continue
ref= ngram.replace('_', strings)
#if len(ref) > MAX_NGRAMS: continue
try:
dlt[newgram]= deepcopy(dlt[ref])
except Exception:
self.bug('in self.substitute(): while trying to copy dlt[%s] to dlt[%s]'%(ref,ngram), g, dlt)
g[nt][r]= g[nt][r].replace('_', new_nt) # back to normal
dlt= self.updateDLT(g, dlt)
return g, dlt
def substitute(self, g, strings, new_nt, dlt, gid):
"""
Substitute symbols
"""
#print( '>> SUBSTITUTE(%s)=%s on Grammar #%d'% (strings, new_nt, gid) )
#assert(new_nt not in g)
# update existing rules
for nt in self.getNTlist(g):
for r in range(len(g[nt])):
#if nt==dlt[strings]['parent']: continue ############ DON't NEED IT
oldRule= g[nt][r]
ref= dlt[oldRule]
if strings in oldRule:
newRule= oldRule.replace(strings, new_nt)
g[nt][r]= oldRule.replace(strings, '_') # will be changed later; important
if newRule not in dlt:
dlt[newRule]= {'terms':deepcopy(ref['terms']),
'score':ref['score'], 'count':ref['count'], 'parent':ref['parent']}
dlt[oldRule]['parent']= None # it's "broken" now
g[nt]= uniquify(g[nt])
g= self.addRule(g, new_nt, strings)
# update child-parent relationship
if new_nt not in dlt:
dlt[new_nt]={'terms':deepcopy(dlt[strings]['terms']),'score':-1,'count':0,'parent':None}
dlt[strings]['parent']= new_nt
# build new n-grams including new NT
for s in self.getNTlist(g): # for each NT in g
for r in range(len(g[s])): # for each rule in NT
for ngram in self.ngrams(g[s][r]): # for each sub-pattern
newgram= ngram.replace('_', new_nt)
if newgram in dlt: continue
ref= ngram.replace('_', strings)
#if len(ref) > MAX_NGRAMS: continue
try:
dlt[newgram]= deepcopy(dlt[ref])
except Exception:
self.bug('in self.substitute(): while trying to copy dlt[%s] to dlt[%s]'%(ref,ngram), g, dlt)
g[s][r]= g[s][r].replace('_', new_nt) # back to normal
g[s]= uniquify(g[s])
dlt= self.updateDLT(g, dlt)
return g, dlt
def mergeSet(self, g):
"""
Return NT combinations to merge
"""
s= []
l= self.getNTlist(g)
for n1 in range(len(l)):
for n2 in range(n1+1,len(l)):
s.append( (l[n1],l[n2]) )
return s
def merge(self, nt1, nt2, new_nt, g, dlt, gid):
"""
Merge operation
Two special cases of Merge:
1. need to check self-recursion!
Y -> X
self.merge(X,Y):
Y -> Y should be deleted
2. need to check for uniqueness!
Z1 -> AYB (c1)
-> AXB (c2)
self.merge(X,Y):
Z1 -> AYB (c1)
-> AYB (c2)
should become:
Z1 -> AYB (c1+c2)
"""
#print( '>> MERGE(%s,%s)=%s on Grammar #%d'% (nt1, nt2, new_nt, gid) )
# replace nt1, nt2 to new_nt in both LHS & RHS
for s in self.getNTlist(g):
for r in range(len(g[s])):
if dlt[g[s][r]]['parent']:
if nt1 in dlt[g[s][r]]['parent']:
dlt[g[s][r]]['parent']= dlt[g[s][r]]['parent'].replace(nt1, new_nt)
if nt2 in dlt[g[s][r]]['parent']:
dlt[g[s][r]]['parent']= dlt[g[s][r]]['parent'].replace(nt2, new_nt)
g[s][r]= g[s][r].replace(nt1, new_nt)
g[s][r]= g[s][r].replace(nt2, new_nt)
g[s]= uniquify(g[s]) # case 2
# add merged rule
new_rules= g[nt1] + g[nt2]
delList= []
for r in range(len(new_rules)):
new_rules[r]= new_rules[r].replace(nt1, new_nt)
new_rules[r]= new_rules[r].replace(nt2, new_nt)
if new_rules[r]==new_nt:
delList.append(r) # case 1
for d in sorted(list(set(delList)), reverse=True):
del(new_rules[d])
new_rules= uniquify(new_rules) # case 2
if len(new_rules)==0:
self.bug('This actually happened. While merging, no rules are left on merged RHS !')
# update grammar
for r in new_rules:
g= self.addRule(g, new_nt, r)
g= self.delRule(g, nt1)
g= self.delRule(g, nt2)
# update DLT
for s in dlt:
if nt1 in s:
new_string= s.replace(nt1, new_nt)
if new_string not in dlt:
dlt[new_string]= dlt.pop(s)
else:
dlt[new_string]['terms'].extend(dlt[s]['terms'])
dlt[new_string]['terms']= uniquify(dlt[new_string]['terms'])
################ THIS IS HACK #####################################
# TO AVOID MULTIPLE PARENTS WHICH ACTUALLY DOESN'T AFFECT PROGRAM #
###################################################################
if dlt[s]['parent']: dlt[new_string]['parent']= dlt[s]['parent']
del(dlt[s])
for s in dlt:
if nt2 in s:
new_string= s.replace(nt2, new_nt)
if new_string not in dlt:
dlt[new_string]= dlt.pop(s)
else:
dlt[new_string]['terms'].extend(dlt[s]['terms'])
dlt[new_string]['terms']= uniquify(dlt[new_string]['terms'])
# TO AVOID MULTIPLE PARENTS WHICH ACTUALLY DOESN'T AFFECT THE PERFORMANCE
if dlt[s]['parent']: dlt[new_string]['parent']= dlt[s]['parent']
del(dlt[s])
# THINK: SHOULD DLT UPDATED HERE?
dlt= self.updateDLT(g, dlt)
############# THIS IS SLOWER METHOD ########################
# Instead of adding new_nt, just merge into higher-order NT
# e.g. MERGE(Z,X)=Z
############################################################
if self.upper.index(nt1) < self.upper.index(nt2):
nt_old= new_nt
nt_new= nt1
else:
nt_old= nt1
nt_new= new_nt
for s in self.getNTlist(g):
for x in range(len(g[s])):
r= g[s][x]
if r in dlt and dlt[r]['parent'] != None:
dlt[r]['parent']= dlt[r]['parent'].replace(nt_old, nt_new)
g[s][x]= r.replace(nt_old, nt_new)
if s==nt_old:
if nt_new in g: self.bug('self.merge(): nt_new(%s) is already in G!'% nt_new)
g[nt_new]= g.pop(nt_old)
for s in dlt:
if nt_old in s:
new_s= s.replace(nt_old, nt_new)
if new_s in dlt: self.bug('self.merge(): new_s(%s) is already in G!'% new_s)
dlt[new_s]= dlt.pop(s)
###########################################################
# DLT should be updated here!
dlt= self.updateDLT(g, dlt)
g, dlt= self.sortSymbols(g, dlt)
#if self.getLikelihood(g, dlt) < self.max_mdl: return g, dlt
#else: return None, None
return g, dlt
def pruneGrammar(self, g, dlt):
"""
Prune rules with low probability
"""
pruned= []
for nt in self.getNTlist(g):
sum= 0.0 # sum of all rule scores belonging to nt
for r in range(len(g[nt])): # for each RHS rule of a LHS symbol
rulescore= 0.0
for t in dlt[g[nt][r]]['terms']:
if t not in self.t_stat:
self.bug('%s key is not in self.t_stat'%t, g, dlt)
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
#rulescore *= len(g[nt][r])
sum += rulescore
if sum==0: # all rule probs are identical; cannot continue
return g, dlt, pruned
# compute rule probs and prune if needed
for r in g[nt]:
rulescore= 0.0 # the score of each rule
for t in dlt[r]['terms']:
rulescore += self.t_stat[t]['prob'] * self.t_stat[t]['count']
#rulescore *= len(rs)
ruleprob= rulescore / sum
if ruleprob < PRUNE_P:
if dlt[r]['parent']==nt:
dlt[r]['parent']= None
pruned.append( (nt, r, ruleprob) )
g[nt].remove(r)
# at least one rule must be remained
if len(g[nt])==0:
self.bug('NT %s has no rules!'% nt, g, dlt)
return g, dlt, pruned
def substituteMulti(self, argList):
"""
argList= [gn, ntlist.pop(), new_gid, P_LOCK]
"""
gn, nt, new_gid, P_LOCK= argList[0], argList[1], argList[2], argList[3]
timer= qc.Timer()
worse= gn.worse
g_new, dlt_new= self.substitute(deepcopy(gn.g), nt, self.getNextNT(gn.g), deepcopy(gn.dlt), gn.gid)
if g_new != None:
new_pri= self.getPrior(g_new)
new_lik= self.getLikelihood(g_new, dlt_new)
g_new, dlt_new, pruned= self.pruneGrammar(g_new, dlt_new)
if pruned:
new_pri= self.getPrior(g_new)
new_lik= self.getLikelihood(g_new, dlt_new)
new_mdl= new_pri + new_lik
'''
if new_mdl >= gn.bestmdl:
worse += 1
else:
worse= 0
'''
if VERBOSE > 1:
msg= '[#%d] After SUBSTITUTE(%s) on #%d\n'% (new_gid, nt, gn.gid)
else: