File size: 39,256 Bytes
b5beb60 |
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 |
import math
import os
import re
import string
import unicodedata
from dataclasses import dataclass
from decimal import Decimal, InvalidOperation
from typing import Optional
import regex
from vlmeval.smp.log import get_logger
logger = get_logger('MMDocBench Evaluation')
def digit_fix(text):
manualMap = {
'zero': '0',
'one': '1',
'two': '2',
'three': '3',
'four': '4',
'five': '5',
'six': '6',
'seven': '7',
'eight': '8',
'nine': '9',
'ten': '10',
}
ori_text = text
for string_number, digit_number in manualMap.items():
text = re.sub(rf'\b{string_number}\b', digit_number, text, flags=re.I)
if ori_text != text:
logger.debug(f'[Digit Fix] {repr(ori_text)} -----> {repr(text)}')
return text
def month_fix(text):
month_abbreviations = {
'jan': 'January',
'feb': 'February',
'mar': 'March',
'apr': 'April',
'may': 'May', # May is the same abbreviation and full form
'jun': 'June',
'jul': 'July',
'aug': 'August',
'sep': 'September',
'oct': 'October',
'nov': 'November',
'dec': 'December',
}
# Create a regex pattern that matches any of the abbreviations.
# The r'\b' ensures that we match whole words (word boundaries).
pattern = r'\b(' + '|'.join(month_abbreviations.keys()) + r')(?:\b|(?=\d))'
def replace_match(match):
abbr = match.group(0)
return month_abbreviations[abbr.lower()]
return re.sub(pattern, replace_match, text, flags=re.IGNORECASE)
def normalize_text(s, keep_newline=False):
"""Removing articles and punctuation, and standardizing whitespace are all typical text processing steps."""
def remove_articles(text):
_regex = regex.compile(r'\b(a|an|the)\b', regex.UNICODE)
return regex.sub(_regex, ' ', text)
def white_space_fix(text):
if keep_newline:
text = text.replace('\n', '⚽')
text = ' '.join(text.split())
text = text.replace('⚽', '\n')
return text
return ' '.join(text.split())
def remove_punc(text):
exclude = set(string.punctuation)
return ''.join(ch for ch in text if ch not in exclude)
def lower(text):
return text.lower()
return white_space_fix(digit_fix(remove_articles(remove_punc(lower(s)))))
def normalize_narrative_text(text: str, raw_question: str, sub_task_name: str) -> str:
if re.search(r'\bdate\b', raw_question, re.I):
# replace month abbreviations with full value, e.g., Dec -> December
text = month_fix(text)
# Remove diacritics
text = ''.join(c for c in unicodedata.normalize('NFKD', text) if unicodedata.category(c) != 'Mn')
# Normalize quotes and dashes
text = re.sub(r'[‘’´`]', "'", text)
text = re.sub(r'[“”]', '"', text)
text = re.sub(r'[‐‑‒–—−]', '-', text)
while sub_task_name == 'WTQ':
old_x = text
# Remove citations
text = re.sub(r'((?<!^)\[[^\]]*\]|\[\d+\]|[•♦†‡*#+])*$', '', text.strip())
# Remove details in parentheses
text = re.sub(r'(?<!^)( \([^)]*\))*$', '', text.strip())
# Remove outermost quotation mark
text = re.sub(r'^"([^"]*)"$', r'\1', text.strip())
if text == old_x:
break
return text
def levenshtein_distance(s1, s2):
if len(s1) > len(s2):
s1, s2 = s2, s1
distances = range(len(s1) + 1)
for i2, c2 in enumerate(s2):
distances_ = [i2 + 1]
for i1, c1 in enumerate(s1):
if c1 == c2:
distances_.append(distances[i1])
else:
distances_.append(1 + min((distances[i1], distances[i1 + 1], distances_[-1])))
distances = distances_
return distances[-1]
def anls_compute(groundtruth, prediction):
gt_answer = ' '.join(groundtruth.strip().lower().split())
det_answer = ' '.join(prediction.strip().lower().split())
dist = levenshtein_distance(gt_answer, det_answer)
length = max(len(groundtruth.upper()), len(prediction.upper()))
values = 0.0 if length == 0 else float(dist) / float(length)
return values
def is_reasoning_task(task_name):
return task_name in [
'Arithmetic Reasoning',
'Logical Reasoning',
'Spatial Reasoning',
'Comparison',
'Sorting',
'Counting',
]
def is_narrative_kie(task_name: str, raw_question: str):
narrative_tags = [
'item name',
'add-on name',
'date',
'address',
'company',
'Store_addr_value',
'Store_name_value',
'Tel_value',
'Date_value',
'Time_value'
]
if task_name == 'Key Information Extraction':
return re.search(fr'"(?:{"|".join(narrative_tags)})"', raw_question, re.I)
return False
@dataclass
class NumberValue:
parsed_value: str
status: bool
from_percentage: bool = False
decimal_num: Optional[int] = None
def __post_init__(self):
self.decimal_num = self.decimal_num or 0
def restore_to_value_before_percentage(self):
self.decimal_num = max(0, self.decimal_num - 2)
if self.decimal_num == 0:
self.parsed_value = str(int(Decimal(self.parsed_value) * 100))
else:
self.parsed_value = str(float(Decimal(self.parsed_value) * 100))
self.from_percentage = False
def get_rounded_str(self, round_num: int = 2):
# make sure ('1.23', 3) -> '1.23'
try:
return str(float(round(Decimal(self.parsed_value), round_num)))
except InvalidOperation:
logger.warning(f'Invalid value: {self.parsed_value}')
return self.parsed_value
def get_int_str(self):
return str(int(float(self.parsed_value)))
def maybe_percentage_error(self, other_num: 'NumberValue'):
if self.status and other_num.status and Decimal(self.parsed_value) * 100 == Decimal(other_num.parsed_value):
return True
return False
def maybe_rounding_error(self, other_num: 'NumberValue'):
min_decimal_num = min(self.decimal_num, other_num.decimal_num)
diff_from_other = abs(Decimal(self.parsed_value) - Decimal(other_num.parsed_value))
tolerance = (Decimal('0.1') ** min_decimal_num) * Decimal('0.1')
if diff_from_other <= tolerance:
return True
return False
def maybe_sign_error(self, other_num: 'NumberValue'):
ori_unequal = self.parsed_value != other_num.parsed_value
if other_num.status and ori_unequal and abs(Decimal(self.parsed_value)) == abs(Decimal(other_num.parsed_value)):
return True
return False
def maybe_thousands_error(self, other_num: 'NumberValue'):
if self.status and other_num.status and Decimal(self.parsed_value) * 1000 == Decimal(other_num.parsed_value):
return True
class TextParser:
@staticmethod
def parse_number(value):
if re.search(r',\s?\d{2}\D*$', value):
idx = value.rfind(r',')
if idx != -1:
value = value[:idx] + '.' + value[idx + 1:]
percentage_flag = False
percentage_exists = re.search(r'(\s[-+]?[\d.]*\d+ ?%)|(^[-+]?[\d.]*\d+ ?%)', value)
not_alternative = re.search(
r'(?:\d{1,3}(?:,\d{3})*|\d+\.\d+ million,)\s+or\s+\d{1,2}(?:\.\d+)? ?%', value
) is None
if percentage_exists and not_alternative:
percentage_flag = True
# remove disturbing '-'
value = re.sub(r'(?<=[a-z] )-(?= [$€£¥]?\d+)', '', str(value), flags=re.I)
# remove disturbing '.'
value = re.sub(r'\([a-z,. ]+\)', '', str(value), flags=re.I)
# remove Rp.
value = re.sub(r'\brp\.', '', str(value), flags=re.I)
s = ''.join([ch for ch in str(value) if ch not in '\'"\\$€£¥%(),[]* '])
r_num = r'([+-]?\d+(\.\d+)?)|([+-]?\.\d+)'
groups = re.findall(r_num, s)
if len(groups) == 0:
return NumberValue(value, False, False, None)
num = groups[0][0] or groups[0][2]
if num == '':
return NumberValue(value, False, False, None)
if '.' in num:
if percentage_flag:
float_value = str(float(Decimal(num) / 100))
else:
float_value = str(float(num))
decimal_num = len(float_value) - float_value.rfind('.') - 1
return NumberValue(float_value, status=True, from_percentage=percentage_flag, decimal_num=decimal_num)
elif percentage_flag:
float_value = str(float(Decimal(num) / 100))
decimal_num = len(float_value) - float_value.rfind('.') - 1
return NumberValue(float_value, status=True, from_percentage=percentage_flag, decimal_num=decimal_num)
return NumberValue(str(int(num)), status=True, from_percentage=percentage_flag, decimal_num=None)
@staticmethod
def parse_ordinal_number(value):
pattern = r'(?:^|(?<=\s))(?:first|second|third|fourth|fifth|sixth|seventh|eighth|ninth|tenth)(?:$|(?=\s))'
m = re.search(pattern, value, re.I)
if m:
return m.group()
if re.search(r'(?:^|\s)1st(?:$|\s)', value, re.I):
return 'first'
if re.search(r'(?:^|\s)2nd(?:$|\s)', value, re.I):
return 'second'
if re.search(r'(?:^|\s)3rd(?:$|\s)', value, re.I):
return 'third'
if re.search(r'(?:^|\s)\dth(?:$|\s)', value, re.I):
idx = int(re.search(r'(?:^|\s)(\d)th(?:$|\s)', value, re.I).group(1)) - 4
return ['fourth', 'fifth', 'sixth', 'seventh', 'eighth', 'ninth', 'tenth'][idx]
return None
@staticmethod
def parse_date(value):
value = month_fix(value)
month_map = {
'january': 1,
'february': 2,
'march': 3,
'april': 4,
'may': 5,
'june': 6,
'july': 7,
'august': 8,
'september': 9,
'october': 10,
'november': 11,
'december': 12
}
month_pattern = r'january|february|march|april|may|june|july|august|september|october|november|december'
m = re.search(rf"({month_pattern})(\d{{1,2}})'(\d{{2}})", value, re.I)
if m:
year = int(m.group(3))
month = month_map[m.group(1).lower()]
day = int(m.group(2))
return year, month, day
m = re.search(rf"\b(\d{{2}})\s*({month_pattern})\s*(\d{{4}})\b", value, re.I)
if m:
month = month_map[m.group(2).lower()]
day = int(m.group(1))
year = int(m.group(3))
return year, month, day
m = re.search(rf"\b(\d{{2}})\s*({month_pattern})\s*(\d{{2}})\b", value, re.I)
if m:
# guess
month = month_map[m.group(2).lower()]
day = int(m.group(1))
year = int(m.group(3))
return year, month, day
m = re.search(rf'({month_pattern})\s+(\d{{1,2}}),\s*(\d{{4}})\b', value, re.I)
if m:
month = month_map[m.group(1).lower()]
day = int(m.group(2))
year = int(m.group(3))
return year, month, day
m = re.search(r'\b(\d{1,4})[-/](\d{1,4})[-/](\d{1,4})\b', value, re.I)
if m:
# guess
if len(m.group(1)) == 4 and int(m.group(2)) <= 12:
result_tuple = (int(m.group(1)), int(m.group(2)), int(m.group(3)))
elif len(m.group(1)) == 1 and int(m.group(2)) > 12:
result_tuple = (int(m.group(3)), int(m.group(1)), int(m.group(2)))
elif len(m.group(1)) == 1:
result_tuple = (int(m.group(3)), int(m.group(2)), int(m.group(1)))
else:
result_tuple = (int(m.group(1)), int(m.group(2)), int(m.group(3)))
return result_tuple
@classmethod
def number_in(cls, pred_text, gold_text):
"""
Determines if the `gold_text` is mentioned within `pred_text`.
"""
if len(gold_text) >= 3 or re.search(r'^\d+(?:, \d+){4,}$', pred_text):
if re.search(fr'(?:^| |\b){re.escape(gold_text)}\.?(?:\b(?![,.]\d)| |$)', pred_text, re.I):
return True
number_parts = re.findall(r'\b\d[\d,. ]*\b', pred_text)
number_parts = [pred_text] + number_parts
for one_part in number_parts:
pred_num = cls.parse_number(one_part)
gold_num = cls.parse_number(gold_text)
if pred_num.status and gold_num.status:
return pred_num.parsed_value == gold_num.parsed_value
return False
@staticmethod
def extract_from_equation(pred_text, gold_text, raw_question, is_evidence):
equation_m = re.search(
(
r'-?[$]?\d[\d,.]*(?:%| million)?(?:(?:\s*[+\-*/]\s*-?[$]?\d[\d,.]*(?:%| million)?)'
r'+\s*=\s*(-?[$]?\d[\d,.]*(?:%| million)?))+'
),
pred_text,
)
if equation_m:
return equation_m.group(1).strip('.')
return None
@staticmethod
def extract_number_from_by(pred_text, gold_text, raw_question, is_evidence):
if re.search('(?:by how (?:much|many))|(?:how much .+ change by)', raw_question, re.I):
targets = [t for t in re.split(r'\bby\b', pred_text, re.I)[1:] if re.search(r'\d', t)]
if targets:
return targets[-1].strip('.')
return None
@classmethod
def extract_from_predicate(cls, pred_text, gold_text, raw_question, is_evidence):
predicates = re.findall(r'\b(?:is|be|was|are)\b', pred_text)
if len(predicates) <= 1:
targets = [t for t in re.split(r'\b(?:is|be|was|are)\b', pred_text)[1:] if re.search(r'\d', t)]
if targets:
return targets[-1] # cls.parse_number can be used safely later
else:
# return gold text if it is in
targets = [t for t in re.split(r'\b(?:is|be|was|are)\b', pred_text)[1:] if re.search(r'\d', t)]
for one_target in targets:
if cls.number_in(one_target, gold_text):
return gold_text
# return the part where the prediction number is likely in
indicator_m = re.search(r'\b(?:therefore|which(?= (?:is|would be|was|are)))\b', pred_text, re.I)
if indicator_m:
interested_pred_text = pred_text[indicator_m.end():]
predicates = re.findall(r'\b(?:is|be|was|are)\b', interested_pred_text)
if len(predicates) <= 1:
interested_pred_text = re.split(r'\b(?:is|be|was|are)\b', interested_pred_text)[-1]
if re.search(r'\d', interested_pred_text):
return interested_pred_text # parse_number can be used safely
# remove noises
pred_text = re.sub(r'bounding box[^.]*?is', '', pred_text, re.I)
# split into candidates
targets = re.split(r'\b(?:is|be|was|are)\b', pred_text)
targets = [t for t in targets if re.search(r'^\D{,20}[$()]*\d[\d,. ]*[$()]*', t)]
if len(targets):
return targets[-1]
return None
@staticmethod
def extract_from_date_value_pair(pred_text, gold_text, raw_question, is_evidence):
m = re.search(r'(\d{4}:) (\$)?\d', pred_text)
if m:
return pred_text[m.end(1):]
if is_evidence and re.search(r'\byears?\b', raw_question, re.I):
m = re.search(r'\d{4}\s+(\d[\d ,.]*)', pred_text)
if m:
return m.group(1)
m = re.search(r'^\d{4} [a-z ]+: ?(\$?\d+)$', pred_text)
if m:
return m.group(1)
return None
@staticmethod
def extract_from_year(pred_text, gold_text, raw_question, is_evidence):
if 'which year' in raw_question and not is_evidence:
number_parts = re.findall(r'\b\d{4}\b', pred_text)
if len(number_parts) == 1:
return number_parts[0]
return None
@staticmethod
def extract_from_synonym(pred_text, gold_text, raw_question, is_evidence):
if re.search(r'what was the change', raw_question, re.I) and gold_text == '0':
if re.search(r'no change', pred_text, re.I):
return gold_text
elif re.search(r'^\d+ million$', gold_text, re.I) and re.search(r'^\d+,000,000$', pred_text, re.I):
gold_num = int(re.search(r'^(\d+) million$', gold_text, re.I).group(1)) * 1000000
pred_num = int(pred_text.replace(',', ''))
if gold_num == pred_num:
return gold_text
return None
@classmethod
def extract_number_part(cls, pred_text, gold_text, raw_question='', is_evidence=False):
pred_text = digit_fix(pred_text)
trials = [
cls.extract_from_synonym,
cls.extract_number_from_by,
cls.extract_from_equation,
cls.extract_from_predicate,
cls.extract_from_date_value_pair,
cls.extract_from_year,
]
for extract_fn in trials:
pred_number = extract_fn(pred_text, gold_text, raw_question, is_evidence)
if pred_number is not None:
return pred_number
number_parts = re.findall(r'\d[\d,. ]*', pred_text)
if len(pred_text) and len(number_parts) != 1:
if 'tax amount' in raw_question:
valid_number_parts = [
np for np in pred_text.split() if re.search(r'\d', np) and not re.search(r'[a-zA-Z%]', np)
]
if len(valid_number_parts) == 1:
return valid_number_parts[0]
# if multiple numbers, choose the one with overlap
if is_evidence:
if cls.number_in(pred_text, gold_text):
return gold_text
logger.debug(f'Fail to extract number from {repr(pred_text)}. Returning ori value')
elif len(number_parts) == 1:
# pred: 1,422,313 1,437,708 304,413 3,164,434
# gold: 3,164,434
if len(number_parts[0]) > len(gold_text):
if cls.number_in(pred_text, gold_text):
return gold_text
return pred_text
@staticmethod
def is_gold_number(gold_text):
m = re.search(
(
r'^(-|rp|\$|£|€|eur)?[()\d,. %]+(pts|million|billion|bil\.|bn|articles?|ps|earth years|years?'
r'|months?|weeks?|days?|earth days?|cm|mm|m|km|deaths?|experts?|'
r'seconds?|minutes?|hours?|hrs|hours ahead|events?|bps|kg|g|gallons?'
r'|percent|inches|votes?|k|b\.c\.?|a\.c\.?)?$'
),
gold_text,
re.I,
)
return m is not None
@staticmethod
def is_gold_yes(gold_text):
return re.search(r'^yes\b', gold_text, re.I) is not None
@staticmethod
def is_pred_yes(pred_text):
with_yes = re.search(r'^yes\b', pred_text, re.I) is not None
without_no = re.search(r'^no\b|\bnot\b', pred_text, re.I) is None
no_not = len(pred_text) > 0 and without_no
return no_not or with_yes
@staticmethod
def is_no_absolutely(gold_text):
return gold_text.lower() == 'no'
@staticmethod
def is_no_guessing(pred_text):
return re.search(r'^no\b|\bnot\b', pred_text, re.I) is not None
@staticmethod
def to_string(pred_text):
if pred_text is None:
return ''
return str(pred_text)
@classmethod
def is_same_judgement(cls, pred_text, gold_text):
if cls.is_gold_yes(gold_text) and cls.is_pred_yes(pred_text):
return True
elif cls.is_no_absolutely(gold_text) and cls.is_no_guessing(pred_text):
return True
elif cls.is_no_absolutely(pred_text) and cls.is_no_guessing(gold_text):
return True
return False
@classmethod
def is_same_ordinal_number(cls, pred_text, gold_text, raw_question):
if re.search('which (?:one|amendment)', raw_question, re.I):
pred_ordinal_number = cls.parse_ordinal_number(pred_text)
gold_ordinal_number = cls.parse_ordinal_number(gold_text)
if pred_ordinal_number is not None and gold_ordinal_number is not None:
return pred_ordinal_number == gold_ordinal_number
return False
@classmethod
def is_same_comparative_adjective(cls, pred_text, gold_text, raw_question):
if re.search('more or less|greater than or less than', raw_question, re.I):
pos_pred_adj = re.search(r'\b(?:more|greater)\b', pred_text, re.I)
neg_pred_adj = re.search(r'\b(?:less|lower)\b', pred_text, re.I)
pos_gold_adj = re.search(r'\b(?:more|greater)\b', gold_text, re.I)
neg_gold_adj = re.search(r'\b(?:less|lower)\b', gold_text, re.I)
if pos_pred_adj is not None and pos_gold_adj is not None:
return True
elif neg_pred_adj is not None and neg_gold_adj is not None:
return True
return False
@classmethod
def is_same_date(cls, pred_text, gold_text, raw_question):
date_question = re.search('"(?:date_value|date)"', raw_question, re.I)
if date_question is not None:
# check if string sub-match
pattern = r'date ?|mon ?|tue ?|wed ?|thu ?|fri ?|sat ?|sun ?'
cleaned_gold_text = re.sub(pattern, '', gold_text, flags=re.I).replace(' ', '').lower()
cleaned_pred_text = re.sub(pattern, '', pred_text, flags=re.I).replace(' ', '').lower()
if cleaned_gold_text in cleaned_pred_text:
return True
# check if parsed match
gold_tuple = cls.parse_date(gold_text)
pred_tuple = cls.parse_date(pred_text)
if gold_tuple is not None and pred_tuple is not None:
if gold_tuple[1:] == pred_tuple[1:] and abs(gold_tuple[0] - pred_tuple[0]) in [0, 2000]:
return True
elif gold_tuple == pred_tuple[::-1]:
return True
return False
@staticmethod
def is_same_trend(pred_text, gold_text, raw_question, task_name, sub_task_name):
if sub_task_name == 'CharXiv' and task_name == 'Logical Reasoning':
if re.search(r'\b(?:relationship|vary with|change as)\b', raw_question, re.I):
if re.search('^increase', gold_text, re.I):
straight_answer_pattern = r'(?:^increase)'
# as xxx increases, the xxx increases.
front_pattern = r'(?:as .+(?:increase|decrease).+increase[s.]*$)'
# the xxx increase, as xxx increases.
back_pattern = r'(?:increase.+as .+(?:increase|decrease))'
pattern = f'{straight_answer_pattern}|{front_pattern}|{back_pattern}'
if re.search(pattern, pred_text, re.I):
return True
elif re.search(r'^decrease', gold_text, re.I):
straight_answer_pattern = r'(?:^decrease)'
# as xxx increases, the xxx increases.
front_pattern = r'(?:as .+(?:increase|decrease).+decrease[s.]*$)'
# the xxx increase, as xxx increases.
back_pattern = r'(?:decrease.+as .+(?:increase|decrease))'
pattern = f'{straight_answer_pattern}|{front_pattern}|{back_pattern}'
if re.search(pattern, pred_text, re.I):
return True
elif re.search('^inverse', gold_text, re.I):
inverse_word_pattern = r'(?:\b(?:inversely|while|inverse)\b)'
inverse_sentence_pattern = r'(?:.+decrease.+as .+increase)|(?:.+increase.+as .+decrease)'
pattern = f'{inverse_word_pattern}|{inverse_sentence_pattern}'
if re.search(pattern, pred_text, re.I):
return True
elif re.search(r'change as .+higher', raw_question, re.I):
if re.search('^higher', gold_text, re.I):
front_pattern = r'(?:(?:higher|sharper|increase).+as .+(?:higher|increase))'
back_pattern = r'(?:as .+(?:higher|increase).+(?:higher|sharper|increase))'
pattern = f'{front_pattern}|{back_pattern}'
if re.search(pattern, pred_text, re.I):
return True
return False
@staticmethod
def string_in(pred_text, gold_text, raw_question, sub_task_name):
# in most cases, we check gold if is included in prediction, as prediction tends to be longer
if re.search(fr'(?:^| |\b){re.escape(gold_text)}\.?(?:\b| |$)', pred_text, re.I):
# different from kie's narrative string,
# we just check if gold is in prediction without length limit as prediction can be very long sometimes
# fix false inclusion
if sub_task_name == 'DUDE' and re.search(fr'\b{re.escape(gold_text)} or ', raw_question, re.I):
if re.search(r'\bboth\b', pred_text, re.I):
return False
return True
# in some cases, we check if prediction in included in gold, as predictions from lazy models tends to be shorter
selection_question = re.search(r'\b(?:which|where)\b', raw_question, re.I)
shorter_prediction = len(pred_text) < len(gold_text)
not_too_long_gold = gold_text.count(' ') <= 3
if selection_question and shorter_prediction and not_too_long_gold:
if re.search(fr'(?:^| |\b){re.escape(pred_text)}\.?(?:\b| |$)', gold_text, re.I):
return True
return False
@staticmethod
def remove_repetitions(s):
# meaningless point listing
points = [int(point) for point in re.findall(r'\b(\d+)\. (?=[a-z])', s, re.I)]
if points and len(points) > 10 and points == list(range(points[0], points[-1] + 1)):
logger.debug(f'[Repetition] remove invalid prediction completely: {repr(s[:100] + "...")}')
return ''
# general repetition removal
r = re.compile(r'(.+?)\1+')
detections = []
for match in r.finditer(s):
if match.end() - match.end(1) > 30:
if match.group(1) == '.':
# remove the '.' for correct number parsing later
if re.search(r'[a-z]$', s[:match.start()], re.I) and re.search(r'^ ?\d', s[match.end():], re.I):
detections.append((match.group(1), match.end(1) - 1, match.end()))
else:
detections.append((match.group(1), match.end(1), match.end()))
for r in detections:
logger.debug(f'[Repetition] remove {r[0]} from string')
s = s[:r[1]] + s[r[2]:]
return s
@staticmethod
def preprocess_for_amounts_in_special_format(text):
text_parts = text.split(',')
text_parts[0] = text_parts[0].replace('.', ',')
text = '.'.join(text_parts)
return text
@classmethod
def rule_parse_numeric_value(
cls, pred_text, gold_text, raw_question='', task_name='', sub_task_name='', is_evidence=False
):
"""
Extract and parse numbers from string text. Carefully handle the following cases:
+ amounts where '.' and ',' are used reversely
+ missing or redundant percentage errors
+ extra 0 error
+ calculation error
+ sign error
+ unit error
+ etc
"""
# cases checking
possible_negative = 'decrease' in pred_text
possible_thousands = ('$' in pred_text or re.search(r',\d00|[1-9]000$', pred_text) is not None) and re.search(
r'\bstars\b', raw_question, re.I
) is None
interested_pred_text = cls.extract_number_part(pred_text, gold_text, raw_question, is_evidence)
# handles amounts where '.' and ',' are used reversely for CORD dataset
if sub_task_name == 'CORD':
if re.search(r'^\D*\d{1,3}(?:\.\d{3})*(?:,\d{2})?\D*$', interested_pred_text):
interested_pred_text = cls.preprocess_for_amounts_in_special_format(interested_pred_text)
if re.search(r'^\D*\d{1,3}(?:\.\d{3})*(?:,\d{2})?\D*$', gold_text):
gold_text = cls.preprocess_for_amounts_in_special_format(gold_text)
pred_num = cls.parse_number(interested_pred_text)
gold_num = cls.parse_number(gold_text)
# handles percentage errors
question_for_percentage = re.search(r'[a-z ](?:%|\(%\))[a-z ]|\b(?:percentage|percent)\b', raw_question, re.I)
percent_sign_diff = (
gold_text == interested_pred_text.replace('%', '')
or f'{gold_text}%' in interested_pred_text
or interested_pred_text == gold_text.replace('%', '')
or f'{interested_pred_text}%' in gold_text
)
check_percentage = question_for_percentage or percent_sign_diff
if pred_num.from_percentage and not gold_num.from_percentage and check_percentage:
pred_num.restore_to_value_before_percentage()
elif gold_num.from_percentage and not pred_num.from_percentage and check_percentage:
if gold_num.maybe_percentage_error(pred_num):
gold_num.restore_to_value_before_percentage()
pred_text = pred_num.parsed_value
gold_text = gold_num.parsed_value
if '.' not in gold_text and re.search(r'\.0+$', pred_text):
# extra 00 error
pred_text = pred_num.get_int_str()
elif pred_num.status and pred_num.decimal_num != gold_num.decimal_num:
# extra decimal error and calculation error
if task_name == 'Key Information Extraction':
# decimal error in extraction is not acceptable
round_num = max(pred_num.decimal_num, gold_num.decimal_num)
else:
# slight decimal error in calculation is acceptable
min_decimal_num = min(pred_num.decimal_num, gold_num.decimal_num)
not_all_percentage = pred_num.from_percentage != gold_num.from_percentage
slight_error = gold_num.maybe_rounding_error(pred_num)
if slight_error or not_all_percentage:
round_num = min_decimal_num
else:
round_num = max(pred_num.decimal_num, gold_num.decimal_num)
pred_text = pred_num.get_rounded_str(round_num)
gold_text = gold_num.get_rounded_str(round_num)
# ignore sign error in decrease setting
if gold_text.startswith('-') and gold_text[1:] == pred_text and possible_negative:
pred_text = f'-{pred_text}'
# ignore sign error in subtraction questions
elif (
not is_evidence
and 'difference' in raw_question
and 'absolute' not in raw_question
and gold_num.maybe_sign_error(pred_num)
):
pred_text = gold_text
# ignore thousands error
elif possible_thousands and gold_num.maybe_thousands_error(pred_num):
pred_text = gold_text
return pred_text, gold_text
@classmethod
def rule_parse_extraction_value(
cls, pred_text, gold_text, raw_question='', task_name='', sub_task_name='', is_evidence=False
):
time_question = re.search('"(?:time_value)"', raw_question, re.I)
phone_question = re.search('tel_value', raw_question, re.I)
if gold_text.lower() in pred_text.lower() and len(pred_text) <= 2 * len(gold_text):
# e.g., item name with irrelevant quantity
# add restrictions to avoid false positive
pred_text = gold_text
elif re.sub(r"[ ']", '', gold_text).lower() == re.sub(r"[ ']", '', pred_text).lower():
# extra ' and spaces
pred_text = gold_text
elif phone_question and re.sub(r'\D', '', pred_text) == re.sub(r'\D', '', gold_text):
# 614-766-4494 is the same as (614) 766-4494
pred_text = gold_text
elif time_question:
# remove noisy word and do not care the length restrictions
pattern = r'time[: ]*'
cleaned_gold_text = re.sub(
r':(?=pm)', ' ', re.sub(pattern, '', gold_text, flags=re.I), flags=re.I
).replace(' ', '').lower()
cleaned_pred_text = re.sub(
r':(?=pm)', ' ', re.sub(pattern, '', pred_text, flags=re.I), flags=re.I
).replace(' ', '').lower()
if cleaned_gold_text in cleaned_pred_text:
pred_text = gold_text
elif cls.is_same_date(pred_text, gold_text, raw_question):
pred_text = gold_text
return pred_text, gold_text
@classmethod
def rule_parse_narrative_value(
cls, pred_text, gold_text, raw_question='', task_name='', sub_task_name='', is_evidence=False
):
pred_text = normalize_narrative_text(pred_text, raw_question, sub_task_name)
gold_text = normalize_narrative_text(gold_text, raw_question, sub_task_name)
if cls.is_same_ordinal_number(pred_text, gold_text, raw_question):
pred_text = gold_text
elif cls.is_same_comparative_adjective(pred_text, gold_text, raw_question):
pred_text = gold_text
elif cls.is_same_trend(pred_text, gold_text, raw_question, task_name, sub_task_name):
pred_text = gold_text
elif cls.string_in(pred_text, gold_text, raw_question, sub_task_name):
pred_text = gold_text
elif gold_text.replace(' ', '').lower() == pred_text.replace(' ', '').lower():
# space errors
pred_text = gold_text
elif gold_text.count(' ') >= 3 or (' ' in gold_text and len([i for i in gold_text if i.isalpha()]) >= 15):
# long answer with disturbing context information
# tolerant on spelling errors
# the original prediction
candidates = [pred_text]
# add another candidate((seperated by dot or 'including', 'is' keywords))
seps = ['. ', ', including', 'is']
for seperator in seps:
prediction_parts = pred_text.split(seperator)
if len(prediction_parts) > 1:
candidates.append(prediction_parts[0])
# check if any
for pred_candidate in candidates:
if anls_compute(gold_text, pred_candidate) < 0.5:
pred_text = gold_text
break
return pred_text, gold_text
@classmethod
def rule_parse(cls, pred_text, gold_text, raw_question='', task_name='', sub_task_name='', is_evidence=False):
"""
Check if pred_text expresses the same meaning as gold_text. If so, update pred_text to gold_text.
1. number comparison
1.1 float/int only
1.2 percentage
1.3 number + units
1.4 currency + number
2. yes or no
3. long answer with disturbing context information
"""
if cls.is_gold_number(gold_text):
pred_text, gold_text = cls.rule_parse_numeric_value(
pred_text=pred_text,
gold_text=gold_text,
raw_question=raw_question,
task_name=task_name,
sub_task_name=sub_task_name,
is_evidence=is_evidence
)
elif cls.is_same_judgement(pred_text, gold_text):
pred_text = gold_text
elif is_narrative_kie(task_name, raw_question):
pred_text, gold_text = cls.rule_parse_extraction_value(
pred_text=pred_text,
gold_text=gold_text,
raw_question=raw_question,
task_name=task_name,
sub_task_name=sub_task_name,
is_evidence=is_evidence
)
elif 'Question Answering' in task_name or (is_reasoning_task(task_name)):
pred_text, gold_text = cls.rule_parse_narrative_value(
pred_text=pred_text,
gold_text=gold_text,
raw_question=raw_question,
task_name=task_name,
sub_task_name=sub_task_name,
is_evidence=is_evidence
)
return pred_text, gold_text
@classmethod
def parse(cls, pred_text, gold_text, raw_question='', task_name='', sub_task_name='', is_evidence=False):
pred_text = cls.to_string(pred_text)
gold_text = cls.to_string(gold_text)
pred_text = cls.remove_repetitions(pred_text)
skip_parse = re.search(r'Recognition|Localization|Forgery', task_name) is not None
keep_newline = 'Recognition' in task_name
skip_normalize = False
if not skip_parse and pred_text is not None and gold_text is not None:
if cls.is_gold_number(gold_text):
skip_normalize = True
if pred_text != gold_text:
pred_text, gold_text = cls.rule_parse(
pred_text=pred_text,
gold_text=gold_text,
raw_question=raw_question,
task_name=task_name,
sub_task_name=sub_task_name,
is_evidence=is_evidence
)
if pred_text is not None and not skip_normalize:
pred_text = normalize_text(pred_text, keep_newline=keep_newline)
if gold_text is not None and not skip_normalize:
gold_text = normalize_text(gold_text, keep_newline=keep_newline)
return pred_text, gold_text
class BboxParser:
@classmethod
def parse(cls, bbox, model_name=''):
if bbox and isinstance(bbox, list):
if len(bbox) == 4 and (isinstance(bbox[0], int) or isinstance(bbox[0], float)):
if all(co < 2 for co in bbox):
x0 = math.floor(bbox[0] * 1000)
y0 = math.floor(bbox[1] * 1000)
x1 = math.ceil(bbox[2] * 1000)
y1 = math.ceil(bbox[3] * 1000)
return [x0, y0, x1, y1]
return bbox
elif isinstance(bbox[0], dict) and bbox[0].get('bbox'):
return cls.parse(bbox[0]['bbox'])
elif isinstance(bbox[0], list):
return cls.parse(bbox[0])
elif (
isinstance(bbox[0], dict)
and bbox[0].keys() == {'x0', 'y0'}
and isinstance(bbox[1], dict)
and bbox[1].keys() == {'x1', 'y1'}
):
return list(bbox[0].values()) + list(bbox[1].values())
logger.debug(f'Fail to parse bbox: {bbox}. Return empty')
return []
|