-
Notifications
You must be signed in to change notification settings - Fork 5
/
gpt3_relation.py
928 lines (806 loc) · 38.2 KB
/
gpt3_relation.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
import json
import statistics
import os
import pandas as pd
import argparse
import faiss
import sys
import math
from tqdm import tqdm
from transformers import pipeline
from gpt3_api import Demo
import random
import numpy as np
from testeval import compute_f1
from shared.const import semeval_reltoid
from shared.const import semeval_idtoprompt
from shared.const import ace05_reltoid
from shared.const import ace05_idtoprompt
from shared.const import tacred_reltoid
from shared.const import scierc_reltoid
from shared.const import wiki_reltoid
from shared.prompt import instance
from sklearn.metrics import classification_report
from knn_simcse import find_knn_example, find_lmknn_example
from simcse import SimCSE
from shared.prompt import generate_zero_prompt
from shared.prompt import generate_select_prompt
from shared.prompt import generate_select_auto_prompt
from shared.result import get_results_onebyone
from shared.result import get_results_select
#print(prompt_query)
def generate_relation_dict_label(dataset):
labels = []
with open(dataset, "r") as f:
relation_dict = {}
for line in f.read().splitlines():
tmp_dict = json.loads(line)
if tmp_dict["relations"]== [[]]:
rel = "None"
else:
rel = tmp_dict["relations"][0][0][4]
if rel not in relation_dict.keys():
relation_dict[rel] = len(relation_dict.keys())
labels.append(relation_dict[rel])
print(relation_dict)
return relation_dict, labels
def generate_label(dataset, relation_dict):
labels = []
with open(dataset, "r") as f:
for line in f.read().splitlines():
tmp_dict = json.loads(line)
if tmp_dict["relations"]== [[]]:
rel = "NONE"
else:
rel = tmp_dict["relations"][0][0][4]
if rel not in relation_dict.keys():
relation_dict[rel] = len(relation_dict.keys())
labels.append(relation_dict[rel])
print(relation_dict)
return labels
def generate_query(h_type, t_type, relation_list, query_dict):
query_list = []
#print(query_dict)
for rel in relation_list:
if rel == "None":
continue
else:
query = query_dict[str((h_type,rel,t_type))]
query_list.append(query)
return query_list
def build_query_dict(dataset):
with open("query_templates/ace2005.json", "r") as f:
whole_dict = json.load(f)
query_dict = whole_dict["qa_turn2"]
return query_dict
def get_train_example(example_path, reltoid, no_na):
example_dict = {k:list() for k in reltoid.values()}
with open(example_path, "r") as f:
for line in f.read().splitlines():
tmp_dict = json.loads(line)
if tmp_dict["relations"] == [[]]:
rel = "NONE"
example_dict[reltoid[rel]].append(tmp_dict)
else:
rel = tmp_dict["relations"][0][0][4]
example_dict[reltoid[rel]].append(tmp_dict)
return example_dict
def get_test_example(example_path, reltoid):
example_dict = {k:list() for k in reltoid.values()}
with open(example_path, "r") as f:
for line in f.read().splitlines():
tmp_dict = json.loads(line)
if tmp_dict["relations"] == [[]]:
rel = "NONE"
example_dict[reltoid[rel]].append(tmp_dict)
else:
rel = tmp_dict["relations"][0][0][4]
example_dict[reltoid[rel]].append(tmp_dict)
return example_dict
def auto_generate_example(example_dict, reltoid, idtoprompt, num_per_rel, num_na, random_label, reasoning, demo):
#ratio = 0.5
#num_per_rel = 4
num_example = num_per_rel * (len(example_dict.keys()) - 1) + num_na
#select_dict = {"0":0, "A":1,"B":2,"C":3,"D":4,"E":5,"F":6}
#reltoalpha = {0:"0", 1:"A", 2:"B", 3:"C", 4:"D", 5:"E", 6:"F"}
#reltoalpha = {0:"NONE", 1:"Physical", 2:"General and affiliation", 3:"Person and social", 4:"Organization and affiliation", 5:"Part and whole", 6:"Agent and artifact"}
#reltoalpha = {0:"NONE", 1:"PHYSICAL", 2:"GENERAL AND AFFILIATION", 3:"PERSON AND SOCIAL", 4:"ORGANIZATION AND AFFILIATION", 5:"PART AND WHOLE", 6:"AGENT AND ARTIFACT"}
#else:
# if random.random() > 0.9:
# example_list.append(tmp_dict)
# else:
# continue
#examples = [item for k,v in example_dict.items() for item in v]
examples = []
for relid in example_dict.keys():
if relid == 0:
examples.append(random.sample(example_dict[relid], num_na))
else:
examples.append(random.sample(example_dict[relid], num_per_rel))
flat_examples = [item for sublist in examples for item in sublist]
#print(len(examples))
example_list = random.sample(flat_examples, num_example)
#assert False
example_prompt = str()
for tmp_dict in example_list:
string = " ".join(tmp_dict["sentences"][0])
sub_head = tmp_dict["ner"][0][0][0]
sub_tail = tmp_dict["ner"][0][0][1] + 1
obj_head = tmp_dict["ner"][0][1][0]
obj_tail = tmp_dict["ner"][0][1][1] + 1
entity1 = " ".join(tmp_dict["sentences"][0][sub_head:sub_tail])
entity1_type = tmp_dict["ner"][0][0][2]
entity2 = " ".join(tmp_dict["sentences"][0][obj_head:obj_tail])
entity2_type = tmp_dict["ner"][0][1][2]
if random_label:
rel = random.choice([x for x in reltoid.keys()])
elif tmp_dict["relations"] == [[]]:
rel = 'NONE'
else:
rel = tmp_dict["relations"][0][0][4]
if not reasoning:
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ".\n"
else:
#tmp_query = "\nGiven the sentence: \"" + string + "\", What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + "?"
tmp_query = "What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + " in the sentence \"" + string + "\"?"
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ". It is because:"
#print(prompt_query)
#assert False
results, probs = demo.get_multiple_sample(tmp_query)
prompt_query = prompt_query + results[0] +"\n"
#print(prompt_query)
#assert False
example_prompt += prompt_query
return example_prompt
def find_prob(target, result, probs):
if False:
print(result)
print("targettarget\n")
print(target)
print("tokentoken\n")
print(probs["tokens"])
print("===============\n")
try:
index = [x.strip() for x in probs["tokens"]].index(str(target))
#print(probs["token_logprobs"][index])
return math.exp(probs["token_logprobs"][index])
except:
len_target = len(target)
for i in range(2, len_target+1):
for j in range(len(probs["tokens"])):
if i + j > len(probs["tokens"]):
continue
#print(j+i)
#print(len(probs["tokens"]))
tmp_word = "".join([probs["tokens"][x] for x in range(j, j+i)])
if tmp_word.strip() != target:
#print(tmp_word.strip())
continue
else:
#print(tmp_word.strip())
start = j
end = j + i
sum_prob = 0
for k in range(start, end):
sum_prob += math.exp(probs["token_logprobs"][k])
#print(sum_prob)
#if sum_prob == None:
#print(target)
#print(result)
return sum_prob / i
return 0.0
def smooth(x):
if True:
return np.exp(x)/sum(np.exp(x))
else:
return x
def compute_variance(knn_distribution):
count_dis = [0 for x in range(len(knn_distribution))]
for i in knn_distribution:
count_dis[i] += 1
tmp_distribution = 1.0 * np.array(count_dis)
var = statistics.variance(tmp_distribution)
print(var)
if np.argmax(tmp_distribution) == 0 and var < 5:
return 1
else:
return 0
def generate_ft_example(tmp_dict, ft_dict, reltoid, idtoprompt, demo, args):
tmp_example = instance(tmp_dict)
example_list = ft_dict[tmp_example.id]
if args.reverse:
example_list.reverse()
label_other = 0
tmp_knn = []
example_prompt = str()
if args.var:
knn_distribution = []
for tmp_dict in example_list:
if tmp_dict["relations"] == [[]]:
rel = 'NONE'
else:
rel = tmp_dict["relations"][0][0][4]
knn_distribution.append(reltoid[rel])
label_other = compute_variance(knn_distribution)
for tmp_dict in example_list:
string = " ".join(tmp_dict["sentences"][0])
sub_head = tmp_dict["ner"][0][0][0]
sub_tail = tmp_dict["ner"][0][0][1] + 1
obj_head = tmp_dict["ner"][0][1][0]
obj_tail = tmp_dict["ner"][0][1][1] + 1
entity1 = " ".join(tmp_dict["sentences"][0][sub_head:sub_tail])
entity1_type = tmp_dict["ner"][0][0][2]
entity2 = " ".join(tmp_dict["sentences"][0][obj_head:obj_tail])
entity2_type = tmp_dict["ner"][0][1][2]
if args.random_label:
rel = random.choice([x for x in reltoid.keys()])
elif tmp_dict["relations"] == [[]]:
rel = 'NONE'
else:
rel = tmp_dict["relations"][0][0][4]
tmp_knn.append(reltoid[rel])
tmp_example = instance(tmp_dict)
if not args.reasoning or label_other == 1:
if args.structure:
prompt_query = tmp_example.prompt + tmp_example.pred + idtoprompt[reltoid[rel]] + "\n"
else:
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ".\n"
#prompt_query = instance(tmp_dict).reference + " is " + idtoprompt[reltoid[rel]] + ".\n\n"
elif args.self_error:
prompt_query = tmp_example.get_self_error(tmp_dict, demo, reltoid, idtoprompt, args)
else:
#tmp_query = "\nGiven the sentence: \"" + string + "\", What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + "?"
tmp_query = "What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + " in the sentence \"" + string + "\"?"
#print(prompt_query)
#assert False
while(True):
try:
results, probs = demo.get_multiple_sample(tmp_query)
break
except:
continue
#prompt_query = prompt_query + results[0] +"\n"
if args.structure:
prompt_query = tmp_example.prompt + tmp_example.clue + results[0] + tmp_example.pred + idtoprompt[reltoid[rel]] + "\n"
else:
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ". It is because:\n" + results[0] + "\n"
#print(prompt_query)
#assert False
example_prompt += prompt_query
return example_prompt, tmp_knn, label_other, example_list
def generate_lm_example(gpu_index_flat, tmp_dict, train_dict, train_sentences, k, reltoid, idtoprompt, num_per_rel, num_na, random_label, reasoning, demo, var, args):
#train_list = [x for y in train_dict.values() for x in y]
#print(tmp_dict)
#assert False
#print(len(train_list))
example_list = find_lmknn_example(gpu_index_flat, tmp_dict,train_dict,train_sentences, k)
if args.reverse:
example_list.reverse()
label_other = 0
tmp_knn = []
example_prompt = str()
if var:
knn_distribution = []
for tmp_dict in example_list:
if tmp_dict["relations"] == [[]]:
rel = 'NONE'
else:
rel = tmp_dict["relations"][0][0][4]
knn_distribution.append(reltoid[rel])
label_other = compute_variance(knn_distribution)
for tmp_dict in example_list:
string = " ".join(tmp_dict["sentences"][0])
sub_head = tmp_dict["ner"][0][0][0]
sub_tail = tmp_dict["ner"][0][0][1] + 1
obj_head = tmp_dict["ner"][0][1][0]
obj_tail = tmp_dict["ner"][0][1][1] + 1
entity1 = " ".join(tmp_dict["sentences"][0][sub_head:sub_tail])
entity1_type = tmp_dict["ner"][0][0][2]
entity2 = " ".join(tmp_dict["sentences"][0][obj_head:obj_tail])
entity2_type = tmp_dict["ner"][0][1][2]
if random_label:
rel = random.choice([x for x in reltoid.keys()])
elif tmp_dict["relations"] == [[]]:
rel = 'NONE'
else:
rel = tmp_dict["relations"][0][0][4]
tmp_knn.append(reltoid[rel])
tmp_example = instance(tmp_dict)
if not reasoning or label_other == 1:
if args.structure:
prompt_query = tmp_example.prompt + tmp_example.pred + idtoprompt[reltoid[rel]] + "\n"
else:
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ".\n"
#prompt_query = instance(tmp_dict).reference + " is " + idtoprompt[reltoid[rel]] + ".\n\n"
else:
#tmp_query = "\nGiven the sentence: \"" + string + "\", What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + "?"
tmp_query = "What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + " in the sentence \"" + string + "\"?"
#print(prompt_query)
#assert False
while(True):
try:
results, probs = demo.get_multiple_sample(tmp_query)
break
except:
continue
#prompt_query = prompt_query + results[0] +"\n"
if args.structure:
prompt_query = tmp_example.prompt + tmp_example.clue + results[0] + tmp_example.pred + idtoprompt[reltoid[rel]] + "\n"
else:
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ". It is because:\n" + results[0] + "\n"
#print(prompt_query)
#assert False
example_prompt += prompt_query
return example_prompt, tmp_knn, label_other, example_list
def generate_knn_example(knn_model, tmp_dict, train_dict, k, reltoid, idtoprompt, num_per_rel, num_na, random_label, reasoning, demo, var, args):
#train_list = [x for y in train_dict.values() for x in y]
#print(tmp_dict)
#assert False
#print(len(train_list))
example_list = find_knn_example(knn_model, tmp_dict,train_dict,k, args.entity_info)
if args.reverse:
example_list.reverse()
label_other = 0
tmp_knn = []
example_prompt = str()
if var:
knn_distribution = []
for tmp_dict in example_list:
if tmp_dict["relations"] == [[]]:
rel = 'NONE'
else:
rel = tmp_dict["relations"][0][0][4]
knn_distribution.append(reltoid[rel])
label_other = compute_variance(knn_distribution)
for tmp_dict in example_list:
string = " ".join(tmp_dict["sentences"][0])
sub_head = tmp_dict["ner"][0][0][0]
sub_tail = tmp_dict["ner"][0][0][1] + 1
obj_head = tmp_dict["ner"][0][1][0]
obj_tail = tmp_dict["ner"][0][1][1] + 1
entity1 = " ".join(tmp_dict["sentences"][0][sub_head:sub_tail])
entity1_type = tmp_dict["ner"][0][0][2]
entity2 = " ".join(tmp_dict["sentences"][0][obj_head:obj_tail])
entity2_type = tmp_dict["ner"][0][1][2]
if random_label:
rel = random.choice([x for x in reltoid.keys()])
elif tmp_dict["relations"] == [[]]:
rel = 'NONE'
else:
rel = tmp_dict["relations"][0][0][4]
tmp_knn.append(reltoid[rel])
tmp_example = instance(tmp_dict)
if not reasoning or label_other == 1:
if args.structure:
prompt_query = tmp_example.prompt + tmp_example.pred + idtoprompt[reltoid[rel]] + "\n"
else:
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ".\n"
#prompt_query = instance(tmp_dict).reference + " is " + idtoprompt[reltoid[rel]] + ".\n\n"
else:
#tmp_query = "\nGiven the sentence: \"" + string + "\", What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + "?"
tmp_query = "What are the clues that lead the relation between \"" + entity1 + "\" and \"" + entity2 + "\" to be " + idtoprompt[reltoid[rel]] + " in the sentence \"" + string + "\"?"
#print(prompt_query)
#assert False
while(True):
try:
results, probs = demo.get_multiple_sample(tmp_query)
break
except:
continue
#prompt_query = prompt_query + results[0] +"\n"
if args.structure:
prompt_query = tmp_example.prompt + tmp_example.clue + results[0] + tmp_example.pred + idtoprompt[reltoid[rel]] + "\n"
else:
prompt_query = "\nContext: " + string + "\n" + "Given the context, the relation between " + entity1 + " and " + entity2 + " is " + idtoprompt[reltoid[rel]] + ". It is because:\n" + results[0] + "\n"
#print(prompt_query)
#assert False
example_prompt += prompt_query
return example_prompt, tmp_knn, label_other, example_list
def generate_ft_dict(args):
ft_dict = {}
knn_dict = {}
train_dict = {}
if args.use_dev and args.store_error_reason:
knn_path = "./knn_ids/knn_ids_{}_train_dev.txt".format(args.task)
elif args.use_dev:
knn_path = "./knn_ids/knn_ids_{}_dev.txt".format(args.task)
else:
knn_path = "./knn_ids/knn_ids_{}.txt".format(args.task)
with open(knn_path, "r") as f:
num_id = 0
for line in f.read().splitlines():
knn_num = line.split(" ")
ft_dict[num_id] = knn_num[:args.k]
num_id += 1
with open(args.test_dataset, "r") as f:
num_id = 0
for line in f.read().splitlines():
tmp_dict = json.loads(line)
knn_dict[tmp_dict["doc_key"]] = ft_dict[num_id]
num_id += 1
with open(args.example_dataset, "r") as f:
num_id = 0
for line in f.read().splitlines():
tmp_dict = json.loads(line)
train_dict[num_id] = tmp_dict
num_id += 1
knn_ft_dict = {}
for key in knn_dict.keys():
#print(knn_dict[key])
#print(train_dict)
knn_ft_dict[key] = [train_dict[int(x)] for x in knn_dict[key]]
return knn_ft_dict
def get_binary_select(pred, tmp_dict, demo, knn_list, reltoid, idtoprompt, args):
test_example = instance(tmp_dict)
prompt_list = str()
for example in knn_list:
knn_example = instance(example)
if pred == reltoid[knn_example.rel]:
prompt_list += knn_example.discriminator + idtoprompt[pred] + "?" + knn_example.answer + " yes.\n"
else:
prompt_list += knn_example.discriminator + idtoprompt[pred] + "?" + knn_example.answer + " no.\n"
prompt_list += test_example.discriminator + idtoprompt[pred] + "?" + test_example.answer
while True:
try:
results, probs = demo.get_multiple_sample(prompt_list)
break
except:
continue
#print(prompt_list)
print(results[0])
#assert False
if "no" in results[0]:
pred = 0
return pred, math.exp(probs[0]["token_logprobs"][0])
def run(reltoid, idtoprompt, store_path, args):
demo = Demo(
engine=args.model,
temperature=0,
max_tokens=256,
top_p=1,
frequency_penalty=0,
presence_penalty=0,
best_of=1,
logprobs=1,
)
#relation_dict = {'None': 0, 'PHYS': 1, 'GEN-AFF': 2, 'PER-SOC': 3, 'ORG-AFF': 4, 'PART-WHOLE': 5, 'ART': 6}relation_dict = {'Others': 0, 'PHYS': 1, 'GEN-AFF': 2, 'PER-SOC': 3, 'ORG-AFF': 4, 'PART-WHOLE': 5, 'ART': 6}
#reltoid = {'NONE': 0, 'PHYS': 1, 'GEN-AFF': 2, 'PER-SOC': 3, 'ORG-AFF': 4, 'PART-WHOLE': 5, 'ART': 6}
#idtoprompt = {0: "NONE", 1: "PHYSICAL", 2: "GENERAL AND AFFILIATION", 3: "PERSON AND SOCIAL", 4: "ORGANIZATION AND AFFILIATION", 5: "PART AND WHOLE", 6: "AGENT AND ARTIFACT"}
#relation_dict = {'OTHERS': 0, 'PHYS': 1, 'GEN-AFF': 2, 'PER-SOC': 3, 'ORG-AFF': 4, 'PART-WHOLE': 5, 'ART': 6}
#query_dict = build_query_dict(dataset)
#all_labels = generate_label(dataset, reltoid)
example_dict = get_train_example(args.example_dataset, reltoid, args.no_na)
test_dict = get_test_example(args.test_dataset, reltoid)
flat_examples = [item for sublist in test_dict.values() for item in sublist]
test_examples = random.sample(flat_examples, args.num_test)
if args.use_ft:
#ft_file = "./knn_ids/knn_ids_{}.txt".format(args.task)
ft_dict = generate_ft_dict(args)
elif args.use_knn:
#train_list = test_examples
train_list = [x for y in example_dict.values() for x in y]
if args.no_na:
if args.task == "semeval":
train_list = [x for x in train_list if reltoid[x["relations"][0][0][4]] != 0]
else:
train_list = [x for x in train_list if x["relations"] != [[]]]
#train_dict = {"The relation between" + "\"" + x["ner"][0][0][2] + "\" and \"" + x["ner"][0][1][2] + "\" in the sentence \"" + " ".join(x["sentences"][0]) + "\"":x for x in train_list}
if not args.lm_mask:
if args.entity_info:
train_dict = {instance(x).reference:x for x in train_list}
train_sentences = [instance(x).reference for x in train_list]
else:
train_dict = {instance(x).sentence:x for x in train_list}
train_sentences = [instance(x).sentence for x in train_list]
knn_model = SimCSE("princeton-nlp/sup-simcse-roberta-large")
#knn_model = SimCSE("princeton-nlp/sup-simcse-bert-base-uncased")
knn_model.build_index(train_sentences, device="cpu")
else:
train_dict = {instance(x).lm_mask:x for x in train_list}
train_sentences = [instance(x).lm_mask for x in train_list]
res = faiss.StandardGpuResources()
index_flat = faiss.IndexFlatL2(1024)
gpu_index_flat = faiss.index_cpu_to_gpu(res, 0, index_flat)
extractor = pipeline(model="roberta-large", task="feature-extraction")
embed_array = []
for item in tqdm(train_sentences):
result = extractor(item, return_tensors=True)
embeds = result[0].detach().numpy().copy()
embed_array.append(embeds[-3,:])
embed_list = np.array(embed_array)
gpu_index_flat.add(embed_list)
print(len(test_examples))
micro_f1 = 0.0
#example_prompt = auto_generate_example(example_dataset, relation_dict, 18, True)
for run in range(args.num_run):
if args.fixed_example:
example_prompt = auto_generate_example(example_dict, reltoid, idtoprompt, args.num_per_rel, args.num_na, args.random_label, args.reasoning, demo)
print(example_prompt)
if not args.fixed_test:
test_examples = random.sample(flat_examples, args.num_test)
labels = []
preds = []
num = 0
whole_knn = []
whole_prob = []
whole_prob_on_rel = []
store_error_reason = {}
azure_error = []
for tmp_dict in test_examples:
tmp_knn = []
#tmp_dict = json.loads(line)
#na_filter = random.random()
#rel_filter = random.random()
if tmp_dict["relations"] == [[]] and args.no_na:
num += 1
continue
#elif tmp_dict["relations"] == [[]] and na_filter < 0.95:
# num += 1
# continue
if tmp_dict["relations"] != [[]] and tmp_dict["relations"][0][0][4] == "Other" and args.no_na:
num += 1
continue
#if rel_filter < 0.5:
# lineid += 1
# continue
#elif tmp_dict["relations"] == [[]] and na_filter < 0.95:
# num += 1
# continue
if tmp_dict["relations"] != [[]] and tmp_dict["relations"][0][0][4] != "Other" and args.null:
num += 1
continue
#example_dict = get_train_example(example_dataset, reltoid)
label_other = 0
if not args.fixed_example and not args.use_knn:
example_prompt = auto_generate_example(example_dict, reltoid, idtoprompt, args.num_per_rel, args.num_na, args.random_label, args.reasoning, demo)
if args.use_knn:
if args.use_ft:
example_prompt, tmp_knn, label_other, knn_list = generate_ft_example(tmp_dict, ft_dict, reltoid, idtoprompt, demo, args)
elif args.lm_mask:
example_prompt, tmp_knn, label_other, knn_list = generate_lm_example(gpu_index_flat, tmp_dict, train_dict, train_sentences, args.k, reltoid, idtoprompt, args.num_per_rel, args.num_na, args.random_label, args.reasoning, demo, args.var, args)
else:
example_prompt, tmp_knn, label_other, knn_list = generate_knn_example(knn_model, tmp_dict, train_dict, args.k, reltoid, idtoprompt, args.num_per_rel, args.num_na, args.random_label, args.reasoning, demo, args.var, args)
whole_knn.append(tmp_knn)
num += 1
if tmp_dict["relations"] == [[]]:
labels.append(0)
else:
labels.append(reltoid[tmp_dict["relations"][0][0][4]])
sentence = " ".join(tmp_dict["sentences"][0])
#prompt_list, subject, target = generate_zero_prompt(tmp_dict, query_dict, relation_dict.keys())
prompt_list, subject, target = generate_select_auto_prompt(tmp_dict, example_prompt, reltoid, args.no_na, args.reasoning, args)
#results, probs = demo.get_multiple_sample(prompt_list)
#pred, prob_on_rel = get_results_onebyone(demo, prompt_list, target)
#print(prompt_list)
#assert False
if args.var and label_other == 1:
pred = 0
prob_on_rel = 0
prob = {"NONE": 1}
else:
pred, prob_on_rel, prob, error = get_results_select(demo, prompt_list, reltoid, idtoprompt, args.verbalize, args)
if error:
azure_error.append(tmp_dict["doc_key"])
if args.discriminator and pred != 0:
ori_pred = pred
pred, prob = get_binary_select(pred, tmp_dict, demo, knn_list, reltoid, idtoprompt, args)
if pred != ori_pred:
print("work!")
if args.task == "wiki80" and pred == 0:
pred = labels[-1]
#print(prob_on_rel)
#assert False
whole_prob.append(prob)
whole_prob_on_rel.append(prob_on_rel)
preds.append(pred)
f1_result = compute_f1(preds, labels)
print(f1_result, end="\n")
if preds[-1] != labels[-1]:
if args.store_error_reason:
error_reason = instance(tmp_dict).get_error_reason(preds[-1], tmp_dict, example_prompt, demo, idtoprompt, reltoid, args)
store_error_reason[instance(tmp_dict).id] = error_reason
with open("{}/negtive.txt".format(store_path), "a") as negf:
#negf.write(args)
#negf.write("\n")
negf.write(prompt_list + "\n")
negf.write(str(reltoid) + "\n")
negf.write(str(prob_on_rel) + "\n")
negf.write("Prediction: " + str(preds[-1]) + "\n")
#negf.write(preds[num])
negf.write("Gold: " + str(labels[-1]) + "\n")
negf.write(tmp_dict["doc_key"])
negf.write("\n-----------------\n")
else:
if args.store_error_reason:
correct_reason = instance(tmp_dict).get_correct_reason(demo, idtoprompt, reltoid, args)
store_error_reason[instance(tmp_dict).id] = correct_reason
with open("{}/results.txt".format(store_path),"a") as negf:
#negf.write(args)
#negf.write("\n")
negf.write(prompt_list + "\n")
negf.write(str(reltoid) + "\n")
negf.write(str(prob_on_rel) + "\n")
negf.write("Prediction: " + str(preds[-1]) + "\n")
#negf.write(preds[num])
negf.write("Gold: " + str(labels[-1]) + "\n")
#negf.write(str(classification_report(labels[:num], preds, digits=4)))
negf.write(str(f1_result))
negf.write("\n")
#negf.write(labels[num])
negf.write(tmp_dict["doc_key"])
negf.write("\n-----------------\n")
#print(results[0])
#print(probs[0])
#if num > 100:
# assert False
print("processing:", 100*num/len(test_examples), "%", end="\n")
print(classification_report(labels, preds, digits=4))
report = classification_report(labels, preds, digits=4,output_dict=True)
if args.store_error_reason:
with open("stored_reason/{}_dev.txt".format(args.task), "w") as f:
json.dump(store_error_reason, f)
with open("{}/labels.csv".format(store_path), "w") as f:
f.write('\n'.join([str(labels)]))
with open("{}/preds.csv".format(store_path), "w") as f:
f.write('\n'.join([str(preds)]))
with open("{}/probs.csv".format(store_path), "w") as f:
for prob in whole_prob:
json.dump(prob, f)
f.write("\n")
with open("{}/prob_on_rel.csv".format(store_path), "w") as f:
f.write('\n'.join([str(x) for x in whole_prob_on_rel]))
micro_f1 += f1_result["f1"]
with open("{}/azure_error.csv".format(store_path), "w") as f:
f.write('\n'.join([str(azure_error)]))
with open("{}/knn.csv".format(store_path), "w") as f:
for line in whole_knn:
f.write('\n'.join([str(line)]))
f.write("\n")
df = pd.DataFrame(report).transpose()
df.to_csv("{}/result_per_rel.csv".format(store_path))
#print(report)
print(azure_error)
#assert False
avg_f1 = micro_f1 / args.num_run
print("AVG f1:", avg_f1)
print(args)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--task", type=str, default=None, required=True, choices=["ace05","semeval","tacred","scierc","wiki80"])
parser.add_argument("--model", default=None, type=str, required=True)
parser.add_argument("--num_test", type=int, default=100)
parser.add_argument("--example_dataset", type=str, default=None, required=True)
parser.add_argument("--test_dataset", type=str, default=None, required=True)
parser.add_argument("--fixed_example", type=int, default=1)
parser.add_argument("--fixed_test", type=int,default=1)
parser.add_argument("--num_per_rel", type=int, default=2)
parser.add_argument("--num_na", type=int, default=0)
parser.add_argument("--no_na", type=int, default=0)
parser.add_argument("--num_run", type=int, default=3)
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--random_label", type=int, default=0)
parser.add_argument("--reasoning", type=int, default=0)
parser.add_argument("--use_knn", type=int, default=0)
parser.add_argument("--lm_mask", type=int, default=0)
parser.add_argument("--k", type=int, default=0)
parser.add_argument("--bert_sim", type=int, default=1)
parser.add_argument("--var", type=int, default=0)
parser.add_argument("--reverse", type=int, default=0)
parser.add_argument("--verbalize", type=int, default=0)
parser.add_argument("--entity_info", type=int, default=0)
parser.add_argument("--structure", type=int, default=0)
parser.add_argument("--use_ft", type=int, default=0)
parser.add_argument("--self_error", type=int, default=0)
parser.add_argument("--use_dev", type=int, default=0)
parser.add_argument("--store_error_reason", type=int, default=0)
parser.add_argument("--discriminator", type=int, default=0)
parser.add_argument("--name", type=str, default=0)
parser.add_argument("--null", type=str, default=1)
tacred_idtoprompt = {tacred_reltoid[k]:k.upper() for k in tacred_reltoid.keys()}
scierc_idtoprompt = {scierc_reltoid[k]:k.upper() for k in scierc_reltoid.keys()}
wiki_idtoprompt = {wiki_reltoid[k]:k.upper() for k in wiki_reltoid.keys()}
args = parser.parse_args()
if args.null == 1:
args.null = True
else:
args.null = False
if args.lm_mask == 1:
args.lm_mask = True
else:
args.lm_mask = False
if args.verbalize == 1:
args.verbalize = True
else:
args.verbalize = False
if args.entity_info == 1:
args.entity_info = True
else:
args.entity_info = False
if args.reverse == 1:
args.reverse = True
else:
args.reverse = False
if args.var and args.no_na:
raise Exception("Sorry, if focus on no NA examples, please turn var into 0")
if args.var:
args.var = True
else:
args.var = False
if args.fixed_example and args.use_knn:
assert False
if args.fixed_example == 1:
args.fixed_example = True
else:
args.fixed_example = False
if args.fixed_test == 1:
args.fixed_test = True
else:
args.fixed_test = False
if args.reasoning == 1:
args.reasoning = True
else:
args.reasoning = False
if args.no_na == 1:
args.no_na = True
else:
args.no_na = False
if args.random_label == 1:
args.random_label = True
else:
args.random_label = False
print(args)
if args.no_na and args.num_na != 0:
print(args.no_na)
print(args.num_na)
assert False
store_path = "./results/knn_{}_results/test={}_knn={}_reverse={}_nona={}_var={}_{}_{}_seed={}_{}_randomlabel={}_fixedex={}_fixedtest={}_Reason={}_Verbalize={}_Entityinfo={}_structure={}_useft={}_selferror={}_usedev={}_discri={}_{}".format(args.task, args.num_test, args.k, args.reverse, args.no_na, args.var, args.num_per_rel,args.num_na,args.seed,args.model,str(args.random_label),str(args.fixed_example),str(args.fixed_test), str(args.reasoning), args.verbalize, args.entity_info,args.structure, args.use_ft, args.self_error, args.use_dev, args.discriminator, args.name)
if not os.path.exists(store_path):
os.mkdir(store_path)
#task = sys.argv[1]
#test_num = int(sys.argv[2])
#seed = sys.argv[3]
random.seed(args.seed)
if args.task == "semeval":
#example_dataset = "./dataset/semeval_gpt/train.json"
#dataset = "./dataset/semeval_gpt/test.json"
run(semeval_reltoid,semeval_idtoprompt, store_path, args)
elif args.task == "ace05":
#example_dataset = "./dataset/ace05/test.json"
#dataset = "./dataset/ace05/ace05_0.2/ace05_0.2_test.txt"
run(ace05_reltoid,ace05_idtoprompt, store_path, args)
elif args.task == "tacred":
run(tacred_reltoid, tacred_idtoprompt, store_path, args)
elif args.task == "scierc":
run(scierc_reltoid, scierc_idtoprompt, store_path, args)
elif args.task == "wiki80":
run(wiki_reltoid, wiki_idtoprompt, store_path, args)
#relation_list = ["\""+x+"\"" for x in tacred_relation.keys()]
#relation_set = ",".join(relation_list)
#nlp = stanza.Pipeline(lang='en', processors='tokenize,mwt,pos,lemma,depparse')
#results = [x.strip().strip("\"") for x in results[0]]
#print(results[0])
#print(rel)
#if False:
# if results[0].strip() in tacred_relation.keys():
# preds.append(tacred_relation[results[0].strip()])
# print(rel)
# if rel == results[0].strip():
# print("OK!")
# else:
# with open("./negative_results.txt","a") as negf:
# negf.write(neg_prompt)
# negf.write("\nPrediction:")
# negf.write(results[0])
# negf.write("\nGold:")
# negf.write(rel)
# negf.write("\n-----------------\n")
# else:
# print("None")
# preds.append(0)
#print(preds)
#print(labels)
#assert False
# result = compute_f1(preds, labels)
# print(result, end="\n")
#assert False
#dependency(nlp, string)