-
Notifications
You must be signed in to change notification settings - Fork 4
/
bilearn_hybrid.py
1416 lines (931 loc) · 50.6 KB
/
bilearn_hybrid.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
#
# version 2 of bilearner
#
#
# requires numpy, sklearn
#
#
import cPickle as pickle
import cPickle as pickle
from collections import defaultdict
import logging
import math
from pprint import pprint
import re
import biviewer
from indexnumbers import swap_num
import numpy as np
import pipeline
import progressbar
from scipy import stats
import scipy
from sklearn.feature_extraction import DictVectorizer
from sklearn.linear_model import LogisticRegression
from tokenizer import tag_words, MergedTaggedAbstractReader
# from journalreaders import LabeledAbstractReader
from sklearn.pipeline import Pipeline
from sklearn.svm import SVC
from sklearn.ensemble import RandomForestClassifier
from sklearn.feature_selection import SelectKBest
from sklearn.linear_model import RandomizedLogisticRegression
from sklearn.cross_validation import ShuffleSplit
from scipy.sparse import vstack
with open('data/brill_pos_tagger.pck', 'rb') as f:
pos_tagger = pickle.load(f)
logging.basicConfig(level=logging.INFO)
logging.info("Importing python modules")
def show_most_informative_features(vectorizer, clf, n=50):
c_f = sorted(zip(clf.coef_[0], vectorizer.get_feature_names()))
c_f = [(math.exp(w), i) for (w, i) in c_f if (w < 0) or (w > 0)]
if n == 0:
n = len(c_f)/2
top = zip(c_f[:n], c_f[:-(n+1):-1])
for (c1, f1), (c2, f2) in top:
print "\t%.4f\t%-15s\t\t%.4f\t%-15s" % (c1, f1, c2, f2)
class bilearnPipeline(pipeline.Pipeline):
def __init__(self, text, window_size):
self.text = re.sub('(?:[0-9]+)\,(?:[0-9]+)', '', text)
self.functions = [[{"w": word, "p": pos} for word, pos in pos_tagger.tag(self.word_tokenize(sent))] for sent in self.sent_tokenize(swap_num(text))]
self.load_templates()
self.w_pos_window = window_size
self.text = text
def load_templates(self):
self.templates = (
(("w_int", 0),),
# (("w", 1),),
# (("w", 2),),
# (("w", 3),),
# # (("wl", 4),),
# (("w", -1),),
# (("w", -2),),
# (("w", -3),),
# (("wl", -4),),
# (('w', -2), ('w', -1)),
# (('wl', -1), ('wl', -2), ('wl', -3)),
# (('stem', -1), ('stem', 0)),
# (('stem', 0), ('stem', 1)),
# (('w', 1), ('w', 2)),
# (('wl', 1), ('wl', 2), ('wl', 3)),
# (('p', 0), ('p', 1)),
# (('p', 1),),
# (('p', 2),),
# (('p', -1),),
# (('p', -2),),
# (('p', 1), ('p', 2)),
# (('p', -1), ('p', -2)),
# (('stem', -2), ('stem', -1), ('stem', 0)),
# (('stem', -1), ('stem', 0), ('stem', 1)),
# (('stem', 0), ('stem', 1), ('stem', 2)),
# (('p', -2), ),
# (('p', -1), ),
# (('p', 1), ),
# (('p', 2), ),
# (('num', -1), ),
# (('num', 1), ),
# (('cap', -1), ),
# (('cap', 1), ),
# (('sym', -1), ),
# (('sym', 1), ),
(('div10', 0), ),
(('>10', 0), ),
(('numrank', 0), ),
# (('p1', 1), ),
# (('p2', 1), ),
# (('p3', 1), ),
# (('p4', 1), ),
# (('s1', 1), ),
# (('s2', 1), ),
# (('s3', 0), ),
# (('s4', 0), ),
(('wi', 0), ),
(('si', 0), ),
# (('next_noun', 0), ),
# (('next_verb', 0), ),
# (('last_noun', 0), ),
# (('last_verb', 0), ),
)
self.answer_key = "w"
# self.w_pos_window = window_size # set 0 for no w_pos window features
def run_functions(self, show_progress=False):
# make dict to look up ranking of number in abstract
num_list_nest = [[int(word["w"]) for word in sent if word["w"].isdigit()] for sent in self.functions]
num_list = [item for sublist in num_list_nest for item in sublist] # flatten
num_list.sort(reverse=True)
num_dict = {num: len(num_list)+2-rank for rank, num in enumerate(num_list)}
num_index = 0
for i, sent_function in enumerate(self.functions):
last_noun_index = 0
last_noun = "BEGINNING_OF_SENTENCE"
last_verb_index = 0
last_verb = "BEGINNING_OF_SENTENCE"
for j, function in enumerate(sent_function):
# print j
word = self.functions[i][j]["w"]
features = {"num": word.isdigit(),
"cap": word[0].isupper(),
"sym": not word.isalnum(),
"p1": word[0],
"p2": word[:2],
"p3": word[:3],
"p4": word[:4],
"s1": word[-1],
"s2": word[-2:],
"s3": word[-3:],
"s4": word[-4:],
# "stem": self.stem.stem(word),
"wi": j,
"si": i,
"wl": word.lower(),
"punct": not any(c.isalnum() for c in word) # all
}
if word.isdigit():
num = int(word)
num_index += 1
features[">10"] = num > 10
features["w_int"] = num
features["div10"] = ((num % 10) == 0)
features["numrank"] = num_dict[num]
features["numindex"] = num_index
self.functions[i][j].update(features)
# self.functions[i][j].update(words)
# if pos is a noun, back fill the previous words with 'next_noun'
# and the rest as 'last_noun'
pos = self.functions[i][j]["p"]
if re.match("NN.*", pos):
for k in range(last_noun_index, j):
self.functions[i][k]["next_noun"] = word
self.functions[i][k]["last_noun"] = last_noun
last_noun_index = j
last_noun = word
# and the same for verbs
elif re.match("VB.*", pos):
for k in range(last_verb_index, j):
self.functions[i][k]["next_verb"] = word
self.functions[i][k]["last_verb"] = last_verb
last_verb_index = j
last_verb = word
for k in range(last_noun_index, len(sent_function)):
self.functions[i][k]["next_noun"] = "END_OF_SENTENCE"
self.functions[i][k]["last_noun"] = last_noun
for k in range(last_verb_index, len(sent_function)):
self.functions[i][k]["next_verb"] = "END_OF_SENTENCE"
self.functions[i][k]["last_verb"] = last_verb
class bilearnPipelineCochrane(bilearnPipeline):
def __init__(self, text_dict, window_size):
self.functions = []
for key, value in text_dict.iteritems():
self.functions.extend([[{"w": word, "p": pos, "cochrane_part":key} for word, pos in pos_tagger.tag(self.word_tokenize(sent))] for sent in self.sent_tokenize(swap_num(value))])
self.load_templates()
self.w_pos_window = window_size
# self.text = text
def load_templates(self):
self.templates = (
(("w_int", 0),),
# (("w", 1),),
# (("w", 2),),
# (("w", 3),),
# # (("wl", 4),),
# (("w", -1),),
# (("w", -2),),
# (("w", -3),),
# (("wl", -4),),
# (('w', -2), ('w', -1)),
# (('wl', -1), ('wl', -2), ('wl', -3)),
# (('stem', -1), ('stem', 0)),
# (('stem', 0), ('stem', 1)),
# (('w', 1), ('w', 2)),
# (('wl', 1), ('wl', 2), ('wl', 3)),
# (('p', 0), ('p', 1)),
# (('p', 1),),
# (('p', 2),),
# (('p', -1),),
# (('p', -2),),
# (('p', 1), ('p', 2)),
# (('p', -1), ('p', -2)),
# (('stem', -2), ('stem', -1), ('stem', 0)),
# (('stem', -1), ('stem', 0), ('stem', 1)),
# (('stem', 0), ('stem', 1), ('stem', 2)),
# (('p', -2), ),
# (('p', -1), ),
# (('p', 1), ),
# (('p', 2), ),
# (('num', -1), ),
# (('num', 1), ),
# (('cap', -1), ),
# (('cap', 1), ),
# (('sym', -1), ),
# (('sym', 1), ),
(('div10', 0), ),
(('>10', 0), ),
(('numrank', 0), ),
# (('p1', 1), ),
# (('p2', 1), ),
# (('p3', 1), ),
# (('p4', 1), ),
# (('s1', 1), ),
# (('s2', 1), ),
# (('s3', 0), ),
# (('s4', 0), ),
(('wi', 0), ),
(('si', 0), ),
# (('cochrane_part', 0), ),
# (('next_noun', 0), ),
# (('next_verb', 0), ),
# (('last_noun', 0), ),
# (('last_verb', 0), ),
)
class BiLearner():
def __init__(self, test_mode=True, window_size=4):
logging.info("Initialising Bilearner")
self.data = {}
# load cochrane and pubmed parallel text viewer
self.biviewer = biviewer.BiViewer(in_memory=False, test_mode=test_mode)
self.window_size = window_size
# self.stem = PorterStemmer()
def initialise(self, seed="regex"):
self.data["y_lookup_init"] = {i: -1 for i in range(len(self.biviewer))}
# if seed == "annotations":
# self.seed_y_annotations() # generate y vector from regex
# else:
# self.seed_y_regex() # generate y vector from regex
print "Resetting joint predictions"
self.pred_joint = self.data["y_lookup_init"].copy()
# define empty arrays for the answers to both sides
self.y_cochrane_no = np.empty(shape=(len(self.data["words_cochrane"]),))
self.y_pubmed_no = np.empty(shape=(len(self.data["words_pubmed"]),))
# make a new metrics dict
self.metrics = defaultdict(list)
# with 0 = the seed rule iteration
# 1, 2, 3, n... = iterations
# defaultdict can accept non-lists fine, for non changing properties
def generate_features(self, test_mode=False, filter_uniques=True):
"""
generate the variables for the learning problem
each row represents a candidate answer
filter_uniques = exclude candidates where the number is not replicated in both Cochrane + Pubmed
"""
if test_mode:
logging.warning("In test mode: not processing all data")
logging.info("Loading biview data")
# feature variables
# each item is a candidate answer (here meaning an integer from the text)
# the X variables are static
X_cochrane_l = []
X_pubmed_l = []
X_pubmed_external_l = [] # so that predictions on 'unseen' abstracts can't use the internal cochrane features
# store the words (here=numbers) of interest in a separate list
# since these may not always be used as a feature, but still needed
words_cochrane_l = []
words_pubmed_l = []
# these find the corresponding article (biviewer id) which the
# candidate answer comes from, converted to numpy arrays later
study_id_lookup_cochrane_l = []
study_id_lookup_pubmed_l = []
# filter the training/test data to integers which appear in both datasets
# during training
# models for assessment run on the whole dataset
cochrane_distant_filter_l = []
pubmed_distant_filter_l = []
# answer variable PER STUDY (assumes one population per study)
# used to generate y which is used for training
# the y variables change at each iteration of the algorithm
# depending which answers are most probable
# key variable for the biviewer, 'correct' answers are passed
# between the two views with it
# called init since may be reused from stating position
# should make a copy
# self.data["y_lookup_init"] = {}
logging.info("Generating features")
p = progressbar.ProgressBar(len(self.biviewer), timer=True)
counter = 0 # number of studies initially found
for study_id, study in enumerate(self.biviewer):
p.tap()
cochrane_dict, pubmed_dict = study
cochrane_dict_subset = {k: cochrane_dict.get(k, "") for k in ('CHAR_PARTICIPANTS', 'CHAR_INTERVENTIONS', 'CHAR_OUTCOMES', 'CHAR_NOTES')}
pubmed_text = pubmed_dict.get("abstract", "")
# generate features for all studies
# the generate_features functions only generate features for integers
# words are stored separately since they are not necessarily used as features
# but are needed for the answers
# first generate features from the parallel texts
p_cochrane = bilearnPipelineCochrane(cochrane_dict_subset, self.window_size)
words_cochrane_study = p_cochrane.get_words(filter=lambda x: x["w"].isdigit(), flatten=True) # get the integers
p_pubmed = bilearnPipeline(pubmed_text, self.window_size)
words_pubmed_study = p_pubmed.get_words(filter=lambda x: x["w"].isdigit(), flatten=True) # get the integers
# generate filter vectors for integers which match between Cochrane + Pubmed
common_ints = set(words_cochrane_study) & set(words_pubmed_study)
# add presence in both texts as a common feature
p_cochrane.add_feature(feature_id="shared_num", feature_fn=lambda x: x["w"] in common_ints)
p_pubmed.add_feature(feature_id="shared_num", feature_fn=lambda x: x["w"] in common_ints)
# p_cochrane = bilearnPipeline(cochrane_dict_subset["CHAR_PARTICIPANTS"], self.window_size)
p_cochrane.generate_features()
p_pubmed.generate_features()
cochrane_filter_study = p_cochrane.get_answers(answer_key=lambda x: x["w"] in common_ints, filter=lambda x: x["num"], flatten=True)
pubmed_filter_study = p_pubmed.get_answers(answer_key=lambda x: x["w"] in common_ints, filter=lambda x: x["num"], flatten=True)
# get filtered + flattened feature dicts
X_cochrane_study = p_cochrane.get_features(filter=lambda x: x["num"], flatten=True)
X_pubmed_study = p_pubmed.get_features(filter=lambda x: x["num"], flatten=True)
# print X_pubmed_study
# these lists will be made into array to be used as lookup dicts
study_id_lookup_cochrane_l.extend([study_id] * len(X_cochrane_study))
study_id_lookup_pubmed_l.extend([study_id] * len(X_pubmed_study))
# Add features to the feature lists
X_cochrane_l.extend(X_cochrane_study)
X_pubmed_l.extend(X_pubmed_study)
# Add words to the word lists
words_cochrane_l.extend((int(word) for word in words_cochrane_study))
words_pubmed_l.extend((int(word) for word in words_pubmed_study))
# Add filters to the filter lists
cochrane_distant_filter_l.extend(cochrane_filter_study)
pubmed_distant_filter_l.extend(pubmed_filter_study)
logging.info("Creating NumPy arrays")
# create np arrays for fast lookup of corresponding answer
self.data["study_id_lookup_cochrane"] = np.array(study_id_lookup_cochrane_l)
self.data["study_id_lookup_pubmed"] = np.array(study_id_lookup_pubmed_l)
# create vectors for the 'words' which are the candidate answers
self.data["words_cochrane"] = np.array(words_cochrane_l)
self.data["words_pubmed"] = np.array(words_pubmed_l)
# create filter vectors for training
self.data["distant_filter_cochrane"] = np.array(cochrane_distant_filter_l)
self.data["distant_filter_pubmed"] = np.array(pubmed_distant_filter_l)
# set up vectorisers for cochrane and pubmed
self.data["vectoriser_cochrane"] = DictVectorizer(sparse=True)
self.data["vectoriser_pubmed"] = DictVectorizer(sparse=True)
# train vectorisers
self.data["X_cochrane"] = self.data["vectoriser_cochrane"].fit_transform(X_cochrane_l)
self.data["X_pubmed"] = self.data["vectoriser_pubmed"].fit_transform(X_pubmed_l)
self.data["X_pubmed_external"] = self.data["vectoriser_pubmed"].fit_transform(X_pubmed_external_l)
# self.reset()
def seed_y_regex(self, annotation_viewer):
"""
initialises the joint y vector with data from manually annotated abstracts
filter_ids = ids of the MergedTaggedAbstractReader to pay attention to
"""
self.initialise()
self.annotation_viewer_to_biviewer = {}
self.answers = {}
self.annotation_viewer = annotation_viewer
logging.info("Generating answers for test set")
p = progressbar.ProgressBar(len(self.annotation_viewer), timer=True)
for study in range(len(self.annotation_viewer)):
p.tap()
biview_id = annotation_viewer[study]["biview_id"]
self.annotation_viewer_to_biviewer[study] = biview_id
# set answers
parsed_tags = [item for sublist in annotation_viewer.get(study) for item in sublist] # flatten list
tagged_numbers = [w[0] for w in parsed_tags if 'n' in w[1]] # then get any tagged numbers
if tagged_numbers:
number = int(tagged_numbers[0])
else:
number = -2
self.answers[biview_id] = number
logging.info("Generating seed data from regular expression")
p = progressbar.ProgressBar(len(self.biviewer), timer=True)
counter = 0 # number of studies initially found
for study_id, (cochrane_dict, pubmed_dict) in enumerate(self.biviewer):
p.tap()
pubmed_text = pubmed_dict.get("abstract", "")
# use simple rule to identify population sizes (low sens/recall, high spec/precision)
pubmed_text = swap_num(pubmed_text)
matches = re.findall('([1-9][0-9]*) (?:\w+ )*(?:participants|men|women|patients|children|people) were (?:randomi[sz]ed)', pubmed_text)
# matches += re.findall('(?:[Ww]e randomi[sz]ed )([1-9][0-9]*) (?:\w+ )*(?:participants|men|women|patients)', pubmed_text)
# matches += re.findall('(?:[Aa] total of )([1-9][0-9]*) (?:\w+ )*(?:participants|men|women|patients)', pubmed_text)
if len(matches) == 1:
self.data["y_lookup_init"][study_id] = int(matches[0])
counter += 1
self.seed_abstracts = counter
logging.info("%d seed abstracts found", counter)
def seed_y_annotations(self, annotation_viewer, hybrid=False):
"""
initialises the joint y vector with data from manually annotated abstracts
filter_ids = ids of the MergedTaggedAbstractReader to pay attention to
hybrid: set to True if want to combine manually annotated with regex seed data
"""
self.initialise()
self.annotation_viewer_to_biviewer = {}
self.answers = {}
self.annotation_viewer = annotation_viewer
logging.info("Generating answers for test set")
p = progressbar.ProgressBar(len(self.annotation_viewer), timer=True)
for study in range(len(self.annotation_viewer)):
p.tap()
biview_id = annotation_viewer[study]["biview_id"]
self.annotation_viewer_to_biviewer[study] = biview_id
# set answers
parsed_tags = [item for sublist in annotation_viewer.get(study) for item in sublist] # flatten list
tagged_numbers = [w[0] for w in parsed_tags if 'n' in w[1]] # then get any tagged numbers
if tagged_numbers:
number = int(tagged_numbers[0])
else:
number = -2
self.answers[biview_id] = number
self.data["y_lookup_init"][biview_id] = number # we can label all in annotated mode since the 'hidden' results
# are hidden by the filter later, not at this stage
if hybrid:
logging.info("Generating seed data from regular expression for hybrid model")
p = progressbar.ProgressBar(len(self.biviewer), timer=True)
counter = 0 # number of studies initially found
for study_id, (cochrane_dict, pubmed_dict) in enumerate(self.biviewer):
p.tap()
pubmed_text = pubmed_dict.get("abstract", "")
# use simple rule to identify population sizes (low sens/recall, high spec/precision)
pubmed_text = swap_num(pubmed_text)
matches = re.findall('([1-9][0-9]*) (?:\w+ )*(?:participants|men|women|patients|children|people) were (?:randomi[sz]ed)', pubmed_text)
# matches += re.findall('(?:[Ww]e randomi[sz]ed )([1-9][0-9]*) (?:\w+ )*(?:participants|men|women|patients)', pubmed_text)
# matches += re.findall('(?:[Aa] total of )([1-9][0-9]*) (?:\w+ )*(?:participants|men|women|patients)', pubmed_text)
if len(matches) == 1 and self.data["y_lookup_init"][study_id] == -1:
self.data["y_lookup_init"][study_id] = int(matches[0])
counter += 1
self.seed_abstracts = counter
logging.info("%d seed abstracts found", counter)
def reset(self, hide_reader_ids, test_reader_ids):
self.test_reader_ids = test_reader_ids
self.hide_reader_ids = hide_reader_ids
hide_biviewer_ids = []
test_biviewer_ids = []
for study_id in test_reader_ids:
test_biviewer_ids.append(self.annotation_viewer_to_biviewer[study_id])
for study_id in hide_reader_ids:
hide_biviewer_ids.append(self.annotation_viewer_to_biviewer[study_id])
self.visible_biviewer_ids = np.array(list(set(range(len(self.biviewer))) - set(hide_biviewer_ids)))
self.test_biviewer_ids = np.array(test_biviewer_ids)
self.pred_joint = self.data["y_lookup_init"].copy()
# these weird bits of code find the indices of the pubmed and cochrane
# training data vectors which correspond to the biviewer ids found in
# the loop above
self.visible_cochrane_ids = np.arange(self.data["study_id_lookup_cochrane"].shape[0])[np.in1d(self.data["study_id_lookup_cochrane"], self.visible_biviewer_ids)]
self.visible_pubmed_ids = np.arange(self.data["study_id_lookup_pubmed"].shape[0])[np.in1d(self.data["study_id_lookup_pubmed"], self.visible_biviewer_ids)]
self.test_pubmed_ids = np.arange(self.data["study_id_lookup_pubmed"].shape[0])[np.in1d(self.data["study_id_lookup_pubmed"], self.test_biviewer_ids)]
def save_data(self, filename):
with open(filename, 'wb') as f:
pickle.dump(self.data, f)
def load_data(self, filename):
logging.info("loading %s", filename)
with open(filename, 'rb') as f:
self.data = pickle.load(f)
logging.info("%s loaded successfully", filename)
def learn(self, iterations=1, C=2, aperture=0.95, sample_weight=1, aperture_type="probability", test_abstract_file="data/test_abstracts.pck"):
# save the current settings
(self.metrics["iterations"], self.metrics["C"], self.metrics["aperture"],
self.metrics["aperture_type"]) = (iterations, C, aperture, aperture_type)
# open the test data
with open(test_abstract_file, 'rb') as f:
raw_data = pickle.load(f)
test_data = []
counter = 0 # number of matches using regular expression rule
# convert to feature vector
for entry in raw_data:
# matches = re.findall('([1-9][0-9]*) (?:participants|men|women|patients) were (?:randomi[sz]ed)', self.numberswapper.swap(entry["text"]))
# if len(matches) == 1:
# counter += 1
features, words = self.generate_prediction_features(entry["text"])
test_data.append({"features":features, "words": words, "answer": entry["answer"], "text": entry['text']})
# logging.info("%d/%d accuracy with seed rule", counter, len(raw_data))
# self.metrics["study_accuracy"] = [counter]
# .get_shape()[1] of the sparse matrix returns the number of columns
# i.e. the total number of features (get_shape() = (nrows, ncolumns))
self.metrics["samples_cochrane"], self.metrics["features_cochrane"] = self.data["X_cochrane"].get_shape()
self.metrics["samples_pubmed"], self.metrics["features_pubmed"] = self.data["X_pubmed"].get_shape()
p = progressbar.ProgressBar(iterations)
p_filter, c_filter = self.data["distant_filter_pubmed"], self.data["distant_filter_cochrane"]
for i in xrange(iterations):
p.tap()
confusion = []
# self.learn_cochrane(C=C, aperture=aperture, aperture_type=aperture_type)
if i > 0:
# # if i % 2 == 0:
# cochrane_model_f = self.learn_view(self.data["X_cochrane"][c_filter], self.data["words_cochrane"][c_filter], self.data["study_id_lookup_cochrane"][c_filter],
# C=C, aperture=aperture, aperture_type=aperture_type, update_joint=True)
# # else:
# pubmed_model_f = self.learn_view(self.data["X_pubmed"][p_filter], self.data["words_pubmed"][p_filter], self.data["study_id_lookup_pubmed"][p_filter],
# C=C, aperture=aperture, aperture_type=aperture_type, update_joint=True)
# print p_filter
# print c_filter
# print len(p_filter), len(c_filter)
cochrane_model_f = self.learn_view(self.data["X_cochrane"], self.data["words_cochrane"], self.data["study_id_lookup_cochrane"],
C=C, aperture=aperture, aperture_type=aperture_type, update_joint=True, sample_weight=sample_weight)
# print show_most_informative_features(self.data["vectoriser_pubmed"], cochrane_model_f)
pubmed_model = self.learn_view(self.data["X_pubmed"], self.data["words_pubmed"], self.data["study_id_lookup_pubmed"],
C=C, aperture=aperture, aperture_type=aperture_type, update_joint=True, sample_weight=sample_weight)
else:
pubmed_model = self.learn_view(self.data["X_pubmed"], self.data["words_pubmed"], self.data["study_id_lookup_pubmed"],
C=0.8, aperture=aperture, aperture_type=aperture_type, update_joint=False, sample_weight=sample_weight)
score = 0
score2 = 0 # score positive and negative results
denominator2 = 0
for entry in test_data:
if entry["features"] is not None:
prediction = self.predict_population_features(entry["features"], entry["words"], pubmed_model)
denominator2 += len(entry["words"])
if int(prediction[0])==entry["answer"]:
score += 1
score2 += len(entry["words"])
else:
score2 += len(entry["words"]) - 2
confusion.append(entry["text"])
logging.info("%d/%d accuracy after %d iterations", score, len(test_data), i)
self.metrics["study_accuracy"].append(score)
self.metrics["study_accuracy_2"].append(score2)
self.metrics["study_denominator_2"].append(denominator2)
with open('confused.txt', 'wb') as f:
f.write("\n\n".join(confusion))
def learn_pubmed(self, C=1.0, update=False, aperture_type="absolute", aperture=20, sample_weight=1):
pubmed_model = self.learn_view(self.data["X_pubmed"][self.visible_pubmed_ids], self.data["words_pubmed"][self.visible_pubmed_ids], self.data["study_id_lookup_pubmed"][self.visible_pubmed_ids],
C=C, aperture=aperture, aperture_type=aperture_type, update_joint=update, sample_weight=sample_weight)
return pubmed_model
def learn_cochrane(self, C=1.0, update=False, aperture_type="absolute", aperture=20, sample_weight=1):
cochrane_model = self.learn_view(self.data["X_cochrane"][self.visible_cochrane_ids], self.data["words_cochrane"][self.visible_cochrane_ids], self.data["study_id_lookup_cochrane"][self.visible_cochrane_ids],
C=C, aperture=aperture, aperture_type=aperture_type, update_joint=update, sample_weight=sample_weight)
return cochrane_model
def learn_view(self, X_view, words_view, joint_from_view_index,
C=1.0, aperture=0.90, aperture_type='probability', update_joint=True,
sample_weight=1):
initial_test_filter = np.empty(shape=(len(words_view),), dtype=bool)
for word_id in xrange(len(words_view)):
y = self.data["y_lookup_init"][joint_from_view_index[word_id]]
initial_test_filter[word_id] = (y != -1)
# print "self.pred_joint info"
# no_known = len([i for i in self.pred_joint.values() if i != -1])
# print "no known = %d/%d" % (no_known, len(self.pred_joint.values()))
# print
# extend all the variables appropriately
X_view_w = X_view
words_view_w = words_view
joint_from_view_index_w = joint_from_view_index
# print "X VIEW"
# print X_view.get_shape()
# print X_view
# print "X VIEW WINDOW - pre"
# print X_view_w.get_shape()
# print X_view_w
initial_test_filter = initial_test_filter.nonzero()[0]
# print "initial test filter"
# print initial_test_filter
# print len(initial_test_filter)
for i in range(sample_weight-1):
X_view_w = vstack((X_view_w, X_view[initial_test_filter]), format="csr")
words_view_w = np.concatenate((words_view_w, words_view[initial_test_filter]))
joint_from_view_index_w = np.concatenate((joint_from_view_index_w, joint_from_view_index[initial_test_filter]))
# print "X VIEW WINDOW - post"
# print X_view_w.get_shape()
# print X_view_w
pred_view_w = np.empty(shape=(len(words_view_w),), dtype=int) # make a new empty vector for predicted values
# (pred_view is predicted population sizes; not true/false)
# print self.pred_joint
# create answer vectors with the seed answers
for word_id in xrange(len(pred_view_w)):
pred_view_w[word_id] = self.pred_joint[joint_from_view_index_w[word_id]]
y_view_w = (pred_view_w == words_view_w) #* 2 - 1 # set Trues to 1 and Falses to -1
# set filter vectors (-1 = unknown)
filter_train = (pred_view_w != -1).nonzero()[0]
filter_test = (pred_view_w == -1).nonzero()[0]
# print len(filter_train), len(filter_test)
# print filter_train, len(filter_train)
# print filter_train, len(filter_test)
# self.metrics["cochrane_training_examples"].append(len(filter_train))
# self.metrics["cochrane_test_examples"].append(len(filter_test))
if len(filter_test)==0:
print "leaving early - run out of data!"
raise IndexError("out of data")
# set training vectors
X_train = X_view_w[filter_train]
y_train = y_view_w[filter_train]
# and test vectors as the rest
X_test = X_view_w[filter_test]
y_test = y_view_w[filter_test]
# and the numbers to go with it for illustration purposes
words_test = words_view_w[filter_test]
joint_from_view_index_test = joint_from_view_index_w[filter_test]
# make and fit new LR model
# model = LogisticRegression(C=C, penalty='l1')
model = self.model(C=C)
logging.debug("fitting model to cochrane data...")
model.fit(X_train, y_train)
if update_joint:
preds = model.predict_proba(X_test)[:,1] # predict unknowns
# get top results (by aperture type selected)
if aperture_type == "percentile":
top_pc_score = stats.scoreatpercentile(preds, aperture)
top_result_indices = (preds > top_pc_score).nonzero()[0]
elif aperture_type == "absolute":
top_result_indices = np.argsort(preds)[-aperture:]
else:
top_pc_score = aperture
top_result_indices = (preds > top_pc_score).nonzero()[0]
# extend the joint predictions
for i in top_result_indices:
self.pred_joint[joint_from_view_index_test[i]] = words_test[i]
# pubmed = self.biviewer[joint_from_view_index_test[i]][1]['abstract']
# cochrane = self.biviewer[joint_from_view_index_test[i]][0]['CHAR_PARTICIPANTS']
# # print
# print pubmed
# print
# print cochrane
# print words_test
# print words_test[i]
return model
def test_pubmed(self, model, display_preds=False, default_stats=False):
# print "train on"
# print len(self.visible_pubmed_ids)
# print self.visible_pubmed_ids
# print "test on"
# print len(self.test_pubmed_ids)
# print self.test_pubmed_ids
assert len(set(self.visible_pubmed_ids) & set(self.test_pubmed_ids)) == 0 # visible and test must not overlap
return self.test(model, self.data["X_pubmed"][self.test_pubmed_ids], self.data["words_pubmed"][self.test_pubmed_ids], self.data["study_id_lookup_pubmed"][self.test_pubmed_ids], display_preds=display_preds, default_stats=default_stats)
def test(self, model, X_test, words_test, joint_from_view_index, display_preds, default_stats=False):
preds = model.predict_proba(X_test)[:,1] # predict unknowns
metrics = {}
accuracy_count = 0
total_abstracts = len(set(joint_from_view_index))
total_candidates = 0
per_integer_accuracy_count = 0
potential_correct_answers = 0
true_pos_preds = []
false_pos_preds = []
# enforce one pop size per abstract
for id in list(set(joint_from_view_index)):
max_index = preds[joint_from_view_index==id].argmax()
no_candidates = len(preds[joint_from_view_index==id])
pred_probs_neg = np.delete(preds[joint_from_view_index==id], max_index)
std_neg = np.std(pred_probs_neg)
mean_neg = np.mean(pred_probs_neg)
prob_pos = preds[joint_from_view_index==id][max_index]
# print preds[joint_from_view_index==id][~max_index]
# print np.std(preds[joint_from_view_index==id][~max_index])
std_from_mean = (prob_pos - mean_neg) / std_neg
prob_mass = prob_pos / np.sum(preds[joint_from_view_index==id])
expected_ratio = prob_mass / (float(1)/float(no_candidates))
if default_stats: # default tagger selects the first integer
y_pred = words_test[joint_from_view_index==id][0]
else:
y_pred = words_test[joint_from_view_index==id][max_index]
# print
# print self.biviewer[id][1]['abstract']
# print
y_actual = self.answers[id]
# print y_pred, y_actual
# print type(y_pred), type(y_actual)
if y_actual != -2:
potential_correct_answers += 1
if display_preds:
print "Abstract %d; answer=%d" % (id, y_actual)
print "Predicted answer %d\n%f SDs from negative examples\n%f percent of probability mass\n%f times expected probability mass" % (y_pred, std_from_mean, prob_mass, expected_ratio)
result = np.around(np.vstack((preds[joint_from_view_index==id], words_test[joint_from_view_index==id])).T, decimals=2)
np.set_printoptions(precision=3, suppress=True)
print result
total_candidates += no_candidates