-
Notifications
You must be signed in to change notification settings - Fork 22
/
dataprep2.py
1977 lines (1709 loc) · 84 KB
/
dataprep2.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
import mmh3
import logging
import numpy as np
import json
import os
import token_statistics
import re
import xml.etree.ElementTree as ET
import unicodedata
import stringmatch
import h5py
import collections
import gzip
import bz2
import typing
import sys
import html
import time
from enum import Enum
from queue import Queue
from threading import Thread
import settings
#
# Helpers 💁
#
def threaded_generator(g, maxsize=16):
q = Queue(maxsize=maxsize)
sentinel = object()
def fill_queue():
try:
for value in g:
q.put(value)
finally:
q.put(sentinel)
thread = Thread(name=repr(g), target=fill_queue, daemon=True)
thread.start()
yield from iter(q.get, sentinel)
def json_from_file(filename: str):
if filename.endswith(".bz2"):
open_fn = bz2.open
elif filename.endswith(".gz"):
open_fn = gzip.open
else:
open_fn = open
with open_fn(filename, "rt", encoding="UTF-8", errors="replace") as p:
for line in p:
try:
yield json.loads(line)
except ValueError as e:
logging.warning("Error while reading document (%s); skipping", e)
def json_from_files(filenames: typing.List[str]):
for filename in filenames:
yield from json_from_file(filename)
def normalize(s: str) -> str:
s = s.lower()
s = unicodedata.normalize("NFKC", s)
return s
def sanitize_for_json(s: typing.Optional[str]) -> typing.Optional[str]:
if s is not None:
return s.replace("\0", "\ufffd")
else:
return None
#
# Classes 🏫
#
def percentile_function_from_values_and_counts(values: np.ndarray, counts: np.ndarray):
if len(values) <= 0:
return lambda x: 0.5
assert (np.diff(values) >= 0.0).all() # make sure the values are sorted
cum_array = counts.cumsum()
if cum_array.dtype != np.float32:
cum_array = cum_array.astype(np.float32)
total = cum_array[-1]
cum_array /= total
cum_array = np.insert(cum_array, 0, 0.0)
cum_values = np.insert(values, 0, -np.inf)
def result(vs: np.ndarray) -> np.ndarray:
# Let's say we have one document with 2 tokens of size 5.0, and 200 tokens of size 8.0. Then
# we would want all tokens with size 8.0 to end up with a feature value around 0.5, because
# that's the "normal" font size. If we just take the percentile though, the value in this
# example will be 1.0. So instead, we take the percentile of 8.0 including (should be 1.00)
# and excluding (should be 0.10), and we average the two values.
assert cum_values.dtype == vs.dtype # If this is not true, numpy will cast one of them automatically, resulting in terrible performance.
indices = cum_values.searchsorted(vs).clip(1, len(cum_values) - 1)
indices_before = indices - 1
return (cum_array[indices] + cum_array[indices_before]) / 2.0
return result
def percentile_function_from_counts(counts: dict):
cum_array = np.fromiter(
counts.items(), dtype=[("item", np.float32), ("count", np.float32)]
)
cum_array.sort()
return percentile_function_from_values_and_counts(cum_array["item"], cum_array["count"])
def percentile_function_from_values(values: np.ndarray):
values, counts = np.unique(values, return_counts=True)
return percentile_function_from_values_and_counts(values, counts)
class TokenStatistics(object):
def __init__(self, filename):
self.filename = filename
self.tokens = None
self.token_count = None
self.cum_font_sizes = None
self.cum_space_widths = None
# We load all this stuff lazily.
def _ensure_loaded(self):
if self.tokens is not None:
return
# load the file
(texts, fonts, font_sizes, space_widths) = \
token_statistics.load_stats_file_no_coordinates(self.filename)
# prepare normalized tokens
self.tokens = {}
self.token_count = 0
for token, new_count in texts.items():
self.token_count += new_count
token = normalize(token)
old_count = self.tokens.get(token, 0)
self.tokens[token] = old_count + new_count
self.tokens = list(self.tokens.items())
self.tokens.sort(key=lambda x: (-x[1], x[0]))
# print ten least frequent tokens
logging.info("Ten least frequent tokens:")
for token, count in self.tokens[-10:]:
logging.info(" %s", token)
# prepare font sizes and token widths
self.percentile_function_for_font_size = percentile_function_from_counts(font_sizes)
self.percentile_function_for_space_width = percentile_function_from_counts(space_widths)
def get_font_size_percentile(self, font_size):
self._ensure_loaded()
# We have to search for the same data type as we have in the array. Otherwise this is super
# slow.
font_size = np.asarray(font_size, np.float32)
return self.get_font_size_percentiles(font_size)
def get_font_size_percentiles(self, font_sizes: np.array):
assert font_sizes.dtype == np.dtype(np.float32)
self._ensure_loaded()
return self.percentile_function_for_font_size(font_sizes)
def get_space_width_percentile(self, space_width):
self._ensure_loaded()
# We have to search for the same data type as we have in the array. Otherwise this is super
# slow.
space_width = np.asarray(space_width, np.float32)
return self.get_space_width_percentiles(space_width)
def get_space_width_percentiles(self, space_widths: np.array):
assert space_widths.dtype == np.dtype(np.float32)
self._ensure_loaded()
return self.percentile_function_for_space_width(space_widths)
def get_tokens_with_minimum_frequency(self, min_freq: int) -> typing.Generator[str, None, None]:
self._ensure_loaded()
# We can do this because self.tokens is sorted.
for token, count in self.tokens:
if count < min_freq:
break
yield token
def get_tokens_up_to_fraction(self, fraction: float) -> typing.Generator[str, None, None]:
self._ensure_loaded()
# We can do this because self.tokens is sorted.
count_yielded = 0
for token, count in self.tokens:
yield token
count_yielded += count
if count_yielded / self.token_count >= fraction:
break
class VisionOutput(object):
BoundingBox = collections.namedtuple("BoundingBox", [
"label",
"left",
"right",
"top",
"bottom",
"confidence"
])
def __init__(self, file_path):
self.boxes = {}
if file_path is not None:
with open(file_path) as file:
for line in file:
line = json.loads(line)
sha = line["docSha"]
bounding_boxes_for_sha = []
for json_page in line["pages"]:
bounding_boxes_for_page = []
for label, left, top, right, bottom, confidence in json_page:
bounding_boxes_for_page.append(
self.BoundingBox(label, left, right, top, bottom, confidence))
bounding_boxes_for_sha.append(bounding_boxes_for_page)
if sha in self.boxes:
logging.warning("Duplicate sha %s in %s", sha, file_path)
self.boxes[sha] = bounding_boxes_for_sha
def boxes_for_sha_and_page(self, sha: str, page: int) -> typing.List[BoundingBox]:
try:
return self.boxes[sha][page]
except KeyError:
# We don't have boxes for that document.
#logging.warning("Missing vision output for %s", sha) # This is probably never coming back.
return list()
except IndexError:
# We have boxes for that document, but not for that page.
return list()
def pages_for_sha(self, sha: str):
return len(self.boxes[sha])
class GloveVectors(object):
def __init__(self, filename: str):
# Open the file and get the dimensions in it. Vectors themselves are loaded lazily.
self.filename = filename
with gzip.open(filename, "rt", encoding="UTF-8") as lines:
for line in lines:
line = line.split()
self.dimensions = len(line) - 1
break
self.vectors = None
self.vectors_stddev = None
self.word2index = None
def _ensure_vectors(self):
if self.vectors is not None:
return
self.word2index = {}
self.vectors = []
with gzip.open(self.filename, "rt", encoding="UTF-8") as lines:
for line_number, line in enumerate(lines):
line = line.split(" ")
word = normalize(line[0])
try:
self.word2index[word] = len(self.vectors)
self.vectors.append(np.asarray(line[1:], dtype='float32'))
except:
logging.error(
"Error while loading line for '%s' at %s:%d",
word,
self.filename,
line_number)
raise
self.vectors = np.stack(self.vectors)
self.vectors_stddev = np.std(self.vectors)
def get_dimensions(self) -> int:
return self.dimensions
def get_vocab(self):
self._ensure_vectors()
return self.word2index.keys()
def get_vocab_size(self) -> int:
self._ensure_vectors()
return len(self.vectors)
def get_vector(self, word: str):
self._ensure_vectors()
index = self.word2index.get(normalize(word))
if index is None:
return None
else:
return self.vectors[index]
def get_dimensions_with_random(self):
return self.get_dimensions() + 1 # 1 for whether we found a vector or not
def get_vector_or_random(self, word: str):
vector = self.get_vector(word)
if vector is not None:
return np.insert(vector, 0, 0.5)
else:
seed = mmh3.hash(normalize(word)) % (2**31 - 1)
r = np.random.RandomState(seed)
vector = r.normal(
loc=0.0,
scale=self.vectors_stddev,
size=self.get_dimensions()+1
)
vector[0] = -0.5
return vector
class CombinedEmbeddings(object):
"""Combines token statistics and glove vectors to produce embeddings to start training with."""
OOV = " ⚠ OOV ⚠ " # must be something that the tokenizer would destroy
OOV_INDEX = 1 # 0 is the keras masking value
def __init__(
self,
tokenstats: TokenStatistics,
glove: GloveVectors,
embedded_tokens_fraction: int
):
self.tokenstats = tokenstats
self.glove = glove
self.embedded_tokens_fraction = embedded_tokens_fraction
self.token2index = None
self.matrix = None
def _ensure_loaded(self):
if self.token2index is not None:
return
# build token2index
self.token2index = {
token: index + 2 # index 0 is the keras masking value, index 1 is the OOV token
for index, token
in enumerate(self.tokenstats.get_tokens_up_to_fraction(self.embedded_tokens_fraction))
}
self.token2index[self.OOV] = self.OOV_INDEX
# check whether there are duplicate indices
indices = set(self.token2index.values())
assert len(indices) == len(self.token2index)
# make sure that 0, the keras masking value, did not make it into the indices
assert 0 not in indices
# build the embedding matrix
self.matrix = np.zeros(
shape=(len(self.token2index)+1, self.glove.get_dimensions_with_random()), # +1 for the keras mask
dtype=np.float32)
for token, index in self.token2index.items():
self.matrix[index] = self.glove.get_vector_or_random(token)
# print out some stats
inv_count = np.sum(self.matrix[2:,0]) + (0.5 * (len(self.matrix) - 2)) # the first scalar in the word vector is -0.5 if it's OOV, or 0.5 otherwise
oov_count = len(self.matrix[2:]) - inv_count # 2: compensates for the keras mask and the OOV token
assert inv_count + oov_count == len(self.matrix) - 2
logging.info(
"%d words in vocab, %d of them from glove (%.2f%%)",
inv_count + oov_count,
inv_count,
(100 * inv_count) / (inv_count + oov_count))
# Print the 30 most frequent tokens that could not be found in glove
logging.info("Top tokens not embedded:")
tokens_printed = 0
for token in self.tokenstats.get_tokens_with_minimum_frequency(0):
if token not in self.token2index:
logging.info(" %s", token)
tokens_printed += 1
if tokens_printed >= 30:
break
def index_for_token(self, token: str) -> int:
self._ensure_loaded()
r = self.token2index.get(normalize(token), self.OOV_INDEX)
assert r != 0 # we must never return the keras masking value
return r
def dimensions(self):
self._ensure_loaded()
return self.matrix.shape[1]
def glove_vocab(self):
return self.glove.get_vocab()
def vocab_size(self):
self._ensure_loaded()
return self.matrix.shape[0] - 1 # -1 for the keras mask
def matrix_for_keras(self):
self._ensure_loaded()
return self.matrix
#
# Unlabeled Tokens 🗄
#
UNLABELED_TOKENS_VERSION = "tok6"
h5_unicode_type = h5py.special_dtype(vlen=np.unicode)
POTENTIAL_LABELS = [None, "title", "author", "bibtitle", "bibauthor", "bibvenue", "bibyear"]
NONE_LABEL = 0
TITLE_LABEL = POTENTIAL_LABELS.index("title")
AUTHOR_LABEL = POTENTIAL_LABELS.index("author")
BIBTITLE_LABEL = POTENTIAL_LABELS.index("bibtitle")
BIBAUTHOR_LABEL = POTENTIAL_LABELS.index("bibauthor")
BIBVENUE_LABEL = POTENTIAL_LABELS.index("bibvenue")
BIBYEAR_LABEL = POTENTIAL_LABELS.index("bibyear")
MAX_DOCS_PER_BUCKET = 6100
MAX_PAGE_COUNT = 40
# The effective page count used in training will be the minimum of this and the same setting in
# the model settings.
MAX_PAGES_PER_BUCKET = MAX_DOCS_PER_BUCKET * MAX_PAGE_COUNT
_sha1_re = re.compile(r'^[0-9a-f]{40}$')
_sha1DotPdf_re = re.compile(r'^[0-9a-f]{40}\.pdf$')
_sha1FromS2Url = re.compile(r'.*([0-9a-f]{4})/([0-9a-f]{36}).pdf$')
def make_unlabeled_tokens_file(
json_file_names: typing.Union[str, typing.List[str]],
output_file_name: str,
ignore_errors=False
):
if isinstance(json_file_names, str):
json_file_names = [json_file_names]
h5_file = h5py.File(output_file_name, "w-", libver="latest")
try:
h5_doc_metadata = h5_file.create_dataset(
"doc_metadata",
dtype=h5_unicode_type,
shape=(0,), # free-wheeling json structure
maxshape=(MAX_DOCS_PER_BUCKET,))
h5_token_text_features = h5_file.create_dataset(
"token_text_features",
dtype=h5_unicode_type,
shape=(0,2), # token, font name
maxshape=(None,2),
compression="gzip",
compression_opts=9)
h5_token_numeric_features = h5_file.create_dataset(
"token_numeric_features",
dtype=np.float32,
shape=(0, 6), # left, right, top, bottom, font_size, font_space_width
maxshape=(None, 6),
compression="gzip",
compression_opts=9)
for json_doc in json_from_files(json_file_names):
if "error" in json_doc:
if ignore_errors:
continue
else:
raise ValueError("Received error document when error was not expected")
if "doc" in json_doc:
json_doc = json_doc["doc"]
# find the proper doc id
doc_name = json_doc["docName"]
doc_sha = json_doc.get("docSha", None)
if doc_sha is None:
if _sha1DotPdf_re.match(doc_name) is not None:
doc_sha = doc_name[:40]
else:
doc_sha = _sha1FromS2Url.match(doc_name)
if doc_sha is not None:
doc_sha = doc_sha.group(1) + doc_sha.group(2)
else:
doc_name = doc_name.split("/")
for i, id_element in enumerate(doc_name):
if _sha1_re.match(id_element) is not None:
doc_name = doc_name[i:]
break
doc_sha = doc_name[0]
doc_name = "/".join(doc_name)
assert _sha1_re.match(doc_sha) is not None, doc_sha
doc_in_h5 = {} # the structure we are stuffing into doc_metadata
doc_in_h5["doc_id"] = doc_name
doc_in_h5["doc_sha"] = doc_sha
pages_in_h5 = []
try:
json_pages = json_doc["pages"]
except KeyError:
logging.warning("Document %s has no pages, skipping", doc_sha)
continue
effective_page_count = min(MAX_PAGE_COUNT, len(json_pages))
for json_page in json_pages[:effective_page_count]:
page_in_h5 = {}
width = float(json_page["width"])
height = float(json_page["height"])
page_in_h5["dimensions"] = (width, height)
# Get the tokens from the page
json_tokens = json_page.get("tokens", [])
# Filter out tokens that have NaN in them
numeric_fields = ["left", "right", "top", "bottom", "fontSize", "fontSpaceWidth"]
json_tokens = [token for token in json_tokens if
"NaN" not in [token[field_name] for field_name in numeric_fields]]
first_token_index = len(h5_token_text_features)
page_in_h5["first_token_index"] = first_token_index
page_in_h5["token_count"] = len(json_tokens)
h5_token_text_features.resize(first_token_index + len(json_tokens), axis=0)
def sanitize_string(s: str) -> str:
return s.replace("\0", "\ufffd")
h5_token_text_features[first_token_index:first_token_index+len(json_tokens)] = \
[(
sanitize_string(json_token["text"]).encode("utf-8"),
sanitize_string(json_token["font"]).encode("utf-8"),
) for json_token in json_tokens]
h5_token_numeric_features.resize(first_token_index + len(json_tokens), axis=0)
h5_token_numeric_features[first_token_index:first_token_index+len(json_tokens)] = \
[(
float(json_token["left"]),
float(json_token["right"]),
float(json_token["top"]),
float(json_token["bottom"]),
float(json_token["fontSize"]),
float(json_token["fontSpaceWidth"])
) for json_token in json_tokens]
# If we're missing font size or space width, fill in from the other value.
# This happens on a few PDFs. Especially some that Oren authored.
font_sizes = h5_token_numeric_features[first_token_index:first_token_index+len(json_tokens), numeric_fields.index("fontSize")]
space_widths = h5_token_numeric_features[first_token_index:first_token_index+len(json_tokens), numeric_fields.index("fontSpaceWidth")]
if np.all(font_sizes == 0):
h5_token_numeric_features[first_token_index:first_token_index+len(json_tokens), numeric_fields.index("fontSize")] = \
space_widths
if np.all(space_widths == 0):
h5_token_numeric_features[first_token_index:first_token_index+len(json_tokens), numeric_fields.index("fontSpaceWidth")] = \
font_sizes
pages_in_h5.append(page_in_h5)
doc_in_h5["pages"] = pages_in_h5
doc_index = len(h5_doc_metadata)
h5_doc_metadata.resize(doc_index + 1, axis=0)
h5_doc_metadata[doc_index] = json.dumps(doc_in_h5)
h5_file.close()
except:
# If something fails, try cleaning up after ourselves
try:
os.remove(output_file_name)
except FileNotFoundError:
pass
raise
def unlabeled_tokens_file(bucket_path: str):
"""Returns h5 file with unlabeled tokens"""
unlabeled_tokens_path = \
os.path.join(
bucket_path,
"unlabeled-tokens-%s.h5" % UNLABELED_TOKENS_VERSION)
if os.path.exists(unlabeled_tokens_path):
return h5py.File(unlabeled_tokens_path, "r")
logging.info("%s does not exist, will recreate", unlabeled_tokens_path)
temp_unlabeled_tokens_path = unlabeled_tokens_path + ".%d.temp" % os.getpid()
make_unlabeled_tokens_file(
os.path.join(bucket_path, "tokens6.json.bz2"),
temp_unlabeled_tokens_path)
os.rename(temp_unlabeled_tokens_path, unlabeled_tokens_path)
return h5py.File(unlabeled_tokens_path, "r")
#
# Labeling 🏷
#
LABELED_TOKENS_VERSION = "tok6"
_split_words_re = re.compile(r'(\W|\d+)')
_not_spaces_re = re.compile(r'\S+')
_word_characters_re = re.compile(r'[\w]+')
_leading_punctuation = re.compile(r'^[\.]+')
_trailing_punctuation = re.compile(r'[\.]+$')
def trim_punctuation(s: str) -> str:
s = _leading_punctuation.sub("", s)
s = _trailing_punctuation.sub("", s)
return s.strip()
def labeled_tokens_file(bucket_path: str):
"""Returns the h5 file with the labeled tokens"""
labeled_tokens_path = \
os.path.join(
bucket_path,
"labeled-tokens-%s.h5" % LABELED_TOKENS_VERSION)
if os.path.exists(labeled_tokens_path):
return h5py.File(labeled_tokens_path, "r")
total_matches = [0, 0, 0, 0]
logging.info("%s does not exist, will recreate", labeled_tokens_path)
with unlabeled_tokens_file(bucket_path) as unlabeled_tokens:
temp_labeled_tokens_path = labeled_tokens_path + ".%d.temp" % os.getpid()
labeled_file = h5py.File(temp_labeled_tokens_path, "w-", libver="latest")
try:
unlab_doc_metadata = unlabeled_tokens["doc_metadata"]
unlab_token_text_features = unlabeled_tokens["token_text_features"]
unlab_token_numeric_features = unlabeled_tokens["token_numeric_features"]
lab_doc_metadata = labeled_file.create_dataset(
"doc_metadata",
dtype=h5_unicode_type,
shape=(0,), # free-wheeling json structure
maxshape=(len(unlab_doc_metadata),)
)
lab_token_text_features = labeled_file.create_dataset(
"token_text_features",
dtype=h5_unicode_type,
shape=(0,2), # token, font name
maxshape=(len(unlab_token_text_features),2),
compression="gzip",
compression_opts=9)
lab_token_numeric_features = labeled_file.create_dataset(
"token_numeric_features",
dtype=np.float32,
shape=(0, 6), # left, right, top, bottom, font_size, font_space_width
maxshape=(len(unlab_token_numeric_features), 6),
compression="gzip",
compression_opts=9)
lab_token_labels = labeled_file.create_dataset(
"token_labels",
dtype=np.int8,
shape=(0,),
maxshape=(len(unlab_token_text_features),),
compression="gzip",
compression_opts=9)
for unlab_metadata in unlab_doc_metadata:
json_metadata = json.loads(unlab_metadata)
doc_sha = json_metadata["doc_sha"]
doc_id = json_metadata["doc_id"]
logging.info("Labeling %s", doc_id)
nxml_path = re.sub("\\.pdf$", ".nxml", doc_id)
nxml_path = os.path.join(bucket_path, "..", nxml_path)
try:
with open(nxml_path) as nxml_file:
nxml = ET.parse(nxml_file).getroot()
except FileNotFoundError:
logging.warning("Could not find %s; skipping doc", nxml_path)
continue
except UnicodeDecodeError:
logging.warning("Could not decode %s; skipping doc", nxml_path)
continue
def all_inner_text(node):
return "".join(node.itertext())
def textify_string_nodes(nodes):
return " ".join([all_inner_text(an) for an in nodes])
def tokenize(s: str):
"""Tokenizes strings exactly as dataprep does, for maximum matching potential."""
return filter(_not_spaces_re.fullmatch, _split_words_re.split(s))
# read title from nxml
gold_title = nxml.findall("./front/article-meta/title-group/article-title")
if len(gold_title) != 1:
logging.warning("Found %d gold titles for %s; skipping doc", len(gold_title), doc_id)
continue
gold_title = " ".join(tokenize(all_inner_text(gold_title[0])))
gold_title = trim_punctuation(gold_title)
gold_title.replace("\u2026", ". . .") # replace ellipsis
if len(gold_title) <= 4:
logging.warning("Title '%s' is too short; skipping doc", gold_title)
continue
# read authors from nxml
author_nodes = \
nxml.findall("./front/article-meta/contrib-group/contrib[@contrib-type='author']/name")
gold_authors = []
for author_node in author_nodes:
given_names = \
" ".join(tokenize(textify_string_nodes(author_node.findall("./given-names"))))
surnames = \
" ".join(tokenize(textify_string_nodes(author_node.findall("./surname"))))
if len(surnames) <= 0:
logging.warning("No surnames for one of the authors; skipping author")
continue
gold_authors.append((given_names, surnames))
if len(gold_authors) == 0:
logging.warning("Found no gold authors for %s; skipping doc", doc_id)
continue
if len(gold_authors) != len(author_nodes):
logging.warning(
"Didn't find the expected %d authors in %s; skipping doc",
len(author_nodes),
doc_id)
continue
if not gold_title or not gold_authors:
logging.error(
"No title or no authors in %s. This should have been caught earlier.",
doc_id)
continue
# read bibtitles from nxml
gold_bib_nodes = nxml.findall("./back/ref-list/ref/mixed-citation")
gold_bib_nodes += nxml.findall("./back/ref-list/ref/element-citation")
gold_bib_nodes += nxml.findall("./back/ref-list/ref/citation")
if len(gold_bib_nodes) == 0:
logging.warning("Found no gold bib nodes for %s", doc_id)
gold_bib_titles = [None for x in gold_bib_nodes]
gold_bib_author_nodes = [None for x in gold_bib_nodes]
gold_bib_venues = [None for x in gold_bib_nodes]
gold_bib_years = [None for x in gold_bib_nodes]
gold_bib_pubids = [None for x in gold_bib_nodes]
for idx, gold_bib_node in enumerate(gold_bib_nodes):
title = gold_bib_node.findall("./article-title")
if len(title) == 0:
logging.warning("Found no gold bib title for %s entry %s", doc_id, idx)
else:
gold_bib_titles[idx] = title[0]
authors = gold_bib_node.findall("./person-group/name")
if len(authors) == 0:
authors = gold_bib_node.findall("./name")
if len(authors) == 0:
authors = gold_bib_node.findall("./collab")
if len(authors) == 0:
logging.warning("Found no gold bib authors for %s entry %s", doc_id, idx)
else:
gold_bib_author_nodes[idx] = authors
venue = gold_bib_node.findall("./source")
if len(venue) == 0:
logging.warning("Found no venue for %s entry %s", doc_id, idx)
else:
gold_bib_venues[idx] = venue[0]
year = gold_bib_node.findall("./year")
if len(year) == 0:
logging.warning("Found no year for %s entry %s", doc_id, idx)
else:
gold_bib_years[idx] = year[0]
pubid = gold_bib_node.findall("./pub-id")
if len(pubid) == 0:
logging.warning("Found no pubid for %s entry %s", doc_id, idx)
else:
gold_bib_pubids[idx] = pubid
def stringify_elements(e, punct=True):
out = [" ".join(tokenize(" ".join(x.itertext()))) if x is not None else "" for x in e]
if punct:
out = [trim_punctuation(x) for x in out]
out = [x.replace("\u2026", ". . .") for x in out]
return out
def stringify_lists_of_elements(le, delim=" "):
out = [stringify_elements(e, False) if e is not None else [] for e in le]
out = [delim.join(x) for x in out]
return out
gold_bib_alls = stringify_elements(gold_bib_nodes)
gold_bib_titles = stringify_elements(gold_bib_titles)
gold_bib_venues = stringify_elements(gold_bib_venues, False)
gold_bib_pubids = stringify_lists_of_elements(gold_bib_pubids)
#strip pubids, which don't occur in doc
for i, a in enumerate(gold_bib_alls):
if not gold_bib_pubids[i] is None:
gold_bib_alls[i] = gold_bib_alls[i].replace(gold_bib_pubids[i], "")
gold_bib_authors = [[] for x in gold_bib_alls]
for bib_idx, authors_node in enumerate(gold_bib_author_nodes):
if not authors_node is None:
for author_node in authors_node:
given_names = \
" ".join(tokenize(textify_string_nodes(author_node.findall("./given-names"))))
surnames = \
" ".join(tokenize(textify_string_nodes(author_node.findall("./surname"))))
if len(surnames) <= 0:
logging.warning("No surnames for one of the bib authors; skipping author")
continue
gold_bib_authors[bib_idx].append((given_names, surnames))
gold_bib_years = stringify_elements(gold_bib_years)
effective_page_count = min(
MAX_PAGE_COUNT,
len(json_metadata["pages"]))
# find titles, authors, bibs in the document
title_match = None
author_matches = []
for author_index in range(len(gold_authors)):
author_matches.append([])
bib_all_matches = [[] for x in gold_bib_titles]
bib_title_matches = [[] for x in gold_bib_titles]
bib_venue_matches = [[] for x in gold_bib_titles]
bib_author_matches = [[] for x in gold_bib_titles]
bib_year_matches = [[] for x in gold_bib_titles]
FuzzyMatch = collections.namedtuple("FuzzyMatch", [
"page_number",
"first_token_index",
"one_past_last_token_index",
"cost",
"matched_string",
"average_font_size"
])
for page_number in range(effective_page_count):
json_page = json_metadata["pages"][page_number]
page_first_token_index = int(json_page["first_token_index"])
token_count = int(json_page["token_count"])
tokens = unlab_token_text_features[page_first_token_index:page_first_token_index+token_count,0]
font_sizes = unlab_token_numeric_features[page_first_token_index:page_first_token_index+token_count,4]
# concatenate the document into one big string, but keep a way to refer back to
# the tokens
page_text = []
page_text_length = 0
start_pos_to_token_index = {}
token_index_to_start_pos = {}
for token_index, token in enumerate(tokens):
if len(page_text) > 0:
page_text.append(" ")
page_text_length += 1
start_pos_to_token_index[page_text_length] = token_index
token_index_to_start_pos[token_index] = page_text_length
normalized_token_text = normalize(token)
page_text.append(normalized_token_text)
page_text_length += len(normalized_token_text)
page_text = "".join(page_text)
assert page_text_length == len(page_text)
def find_string_in_page(string: str, begin = None, end = None) -> typing.Generator[FuzzyMatch, None, None]:
string = normalize(string)
if len(string) == 0:
return
if begin is None:
offset = 0
else:
offset = token_index_to_start_pos.get(begin, 0)
if end is None:
end = len(page_text)
else:
end = token_index_to_start_pos.get(end, len(page_text))
while offset < end:
fuzzy_match = stringmatch.match(string, page_text[offset:])
if fuzzy_match.cost > ((len(string) - string.count(" ")) // 5):
# stringmatch.match() returns results in increasing order of cost.
# Once the cost is too high, it'll never get better, so we can
# bail here.
return
start = fuzzy_match.start_pos + offset
first_token_index = None
while not first_token_index and start >= 0:
first_token_index = start_pos_to_token_index.get(start, None)
start -= 1
if not first_token_index:
first_token_index = 0
end = fuzzy_match.end_pos + offset
one_past_last_token_index = None
while one_past_last_token_index is None and end < len(page_text):
one_past_last_token_index = start_pos_to_token_index.get(end, None)
end += 1
if one_past_last_token_index is None:
one_past_last_token_index = token_count
assert first_token_index != one_past_last_token_index
matched_string = tokens[first_token_index:one_past_last_token_index]
matched_string = " ".join(matched_string)
yield FuzzyMatch(
page_number,
first_token_index,
one_past_last_token_index,
fuzzy_match.cost,
matched_string,
np.average(font_sizes[first_token_index:one_past_last_token_index])
)
offset += fuzzy_match.end_pos
#
# find title
#
def title_match_sort_key(match: FuzzyMatch):
return (
match.cost,
-match.average_font_size,
match.first_token_index
)
title_matches_on_this_page = list(find_string_in_page(gold_title))
if len(title_matches_on_this_page) > 0:
title_match_on_this_page = min(title_matches_on_this_page, key=title_match_sort_key)
if title_match is None or title_match_on_this_page.cost < title_match.cost:
title_match = title_match_on_this_page
#
# find authors
#
for author_index, author in enumerate(gold_authors):
def initials(names, space=" "):
return space.join(
(x[0] for x in filter(_word_characters_re.fullmatch, tokenize(names)))
)
given_names, surnames = author
if len(given_names) == 0:
author_variants = {surnames}
else:
author_variants = {
"%s %s" % (given_names, surnames),
"%s %s" % (initials(given_names, " "), surnames),
"%s . %s" % (initials(given_names, " . "), surnames),
"%s %s" % (initials(given_names, ""), surnames),
"%s , %s" % (surnames, given_names),
"%s %s" % (given_names[0], surnames),
"%s . %s" % (given_names[0], surnames)}
for author_variant in author_variants:
author_matches[author_index].extend(find_string_in_page(author_variant))
#
# find bibs
#
def bib_match_sort_key(match: FuzzyMatch):
return match.cost, match.first_token_index
# find all bib text first. Other fields will be found within these matches
bib_entries_this_page = [1E10, -1] # holds index range of bib entries appearing on this page
bib_all_match = None
for bib_all_index, gold_bib_all in enumerate(gold_bib_alls):
if len(gold_bib_all) == 0:
continue
bib_all_matches_on_this_page = list(find_string_in_page(gold_bib_all))
if len(bib_all_matches_on_this_page) > 0:
bib_all_match_on_this_page = \
min(bib_all_matches_on_this_page, key=bib_match_sort_key)
if bib_all_match is None or bib_all_match_on_this_page.cost < bib_all_match.cost:
bib_all_match = bib_all_match_on_this_page
if not bib_all_match:
continue
bib_all_matches[bib_all_index] = bib_all_match
bib_entries_this_page = [
min(bib_all_index, bib_entries_this_page[0]),
max(bib_all_index, bib_entries_this_page[1])
]
bib_all_match = None
#
# find bibtitles
#
for bib_title_index, gold_bib_title in enumerate(gold_bib_titles):
if len(gold_bib_title) == 0:
continue
bib_title_matches_on_this_page = list(find_string_in_page(gold_bib_title))
if len(bib_title_matches_on_this_page) > 0:
bib_title_matches[bib_title_index] = \
min(bib_title_matches_on_this_page, key=bib_match_sort_key)
def find_authors_in_bounds(out_matches, to_find):
for idx, ses in enumerate(to_find):
if ses is None or len(ses) == 0 or idx < bib_entries_this_page[0] or idx > bib_entries_this_page[1]:
continue
def check(author):
author_matches = []
given_names, surnames = author
if len(given_names) == 0:
author_variants = {surnames}
else:
author_variants = {
"%s %s" % (surnames, given_names),
"%s %s" % (given_names, surnames),
"%s , %s" % (surnames, given_names)
}
if len(bib_all_matches[idx]) == 0: # just find it anywhere:
for author_variant in author_variants: