File size: 10,810 Bytes
c0d65de | 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 | #!/usr/bin/env python3
"""Screen every ACL/EMNLP Findings record independently of awards/presentation.
Raw official XML and one disposition per record are retained. Network retrieval
and PDF inspection are resumable; admission decisions are a separate input.
"""
import concurrent.futures
import csv
import gzip
import hashlib
import json
import re
import shutil
import subprocess
import sys
import xml.etree.ElementTree as ET
from collections import Counter
from pathlib import Path
import requests
ROOT = Path(__file__).resolve().parents[1]
CACHE = Path('/tmp/ai4finance_findings_review')
INSPECTION_VERSION = 2
FIN = re.compile(r'financ(?:e|ial)|\bstocks?\b|\btrading\b|\bportfolio\b|\basset pric|\bderivative pric|\bcredit\b|\bloan\b|\bbank(?:ing|s)?\b|\binsurance\b|\bfraud\b|\bmoney\b|\binvest(?:ment|ing|or)|\bearnings\b|\baccounting\b|\btax(?:ation)?\b|\bESG\b|\beconomic(?:s)?\b|\bmacroeconom|\binflat|\bGDP\b|\bcrypto|\bblockchain|\bAML\b|\bhedging\b|\binvoice|\breceipt|\bbanknote|\bFinQA\b|\bConvFinQA\b|\bTAT-QA\b|\bMultiHiertt\b|\bFiQA\b|\bFinTabNet\b|\bFinBERT\b|\bFinancialPhraseBank\b|\bFinanceBench\b|\bFinTextQA\b|\bXBRL\b|\bEDGAR\b|\bFOMC\b|\bBanking77\b|\bFinCausal\b|\bFinNum\b', re.I)
FOCUSED_FINANCE = re.compile(r'financ(?:e|ial)|\b(?:finqa|convfinqa|tat-qa|multiHiertt|fiqa|fintabnet|banking77|banking 77|financialphrasebank|fincausal|finnum|finbench|sroie|cord dataset)\b|credit[- ](?:risk|scor|default|card)|stock[- ](?:market|pric|movement|predict)|\bloan\b|\binsurance\b|\binvoice|\breceipt|\bESG\b|\bscam|\bfraud|\bcrypto|\bearnings\b|\bFOMC\b', re.I)
def body_length_evidence(text):
"""Require substantive page-six content before explicit back matter.
Uses reading-order extraction, not a PDF's total page count. A failure is
unverified, never an automatic short-paper label.
"""
pages = text.split('\f')
boundary = re.compile(r'(?im)^\s*(?:\d+[. ]+)?(?:References|Bibliography|Limitations|Ethical Considerations|Ethics Statement|Acknowledg(?:e)?ments|Broader Impact|Appendix|Appendices)\s*$')
first = next(((i,m.start()) for i,p in enumerate(pages) if (m:=boundary.search(p))), None)
candidates = []
if first and first[0] >= 5:
for index in range(5, min(first[0]+1, len(pages))):
content = pages[index][:first[1]] if index == first[0] else pages[index]
candidates.append((index,content))
index,sixth = max(candidates, key=lambda x:len(re.findall(r'\b[A-Za-z]{3,}\b',x[1])), default=(5,''))
prose = re.findall(r'\b[A-Za-z]{3,}\b', sixth)
evidence = dict(first_backmatter_page=first[0]+1 if first else None,
inspected_content_page=index+1,
sixth_content_page_words=len(prose),
sixth_content_page_excerpt=re.sub(r'\s+',' ',sixth)[:2400])
return bool(first and len(prose)>=160), evidence
def apply_length_override(result):
"""Use reviewed content boundaries only for the exact inspected publisher PDF."""
path = ROOT/'checks/track_screening/length_overrides.json'
if not path.exists():
return result
review = json.loads(path.read_text()).get(result.get('id'))
if not review:
return result
assert result.get('sha256') == review['pdf_sha256'], 'Reviewed PDF identity changed'
result['content_exceeds_short_limit'] = True
result['length_inspection'] = review['inspection']
result['manual_length_review'] = review
return result
def save(path, value):
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + '\n')
def prepare(source):
target = ROOT / 'findings/sources'
target.mkdir(parents=True, exist_ok=True)
records = []
manifest = []
for year in range(2020, 2027):
name = f'findings-{year}-anthology'
for suffix in ('.gz', '.meta.json'):
shutil.copy2(source / 'sources/raw' / (name + suffix), target / (name + suffix))
data = gzip.decompress((target / (name + '.gz')).read_bytes())
url = json.loads((target / (name + '.meta.json')).read_text())['url']
manifest.append(dict(file=f'findings/sources/{name}.gz', url=url,
uncompressed_sha256=hashlib.sha256(data).hexdigest()))
root = ET.fromstring(data)
for volume in root.findall('volume'):
venue = volume.get('id')
if venue not in ('acl', 'emnlp'):
continue
for node in volume.findall('paper'):
text = lambda tag: ' '.join(''.join(e.itertext()) for e in node.findall(tag)).strip()
ident = f'{year}.findings-{venue}.{node.get("id")}'
page = f'https://aclanthology.org/{ident}/'
record = dict(id=ident, venue=venue.upper(), year=year, track='findings',
title=text('title'), abstract=text('abstract'),
authors=[' '.join(''.join(a.find(t).itertext()) for t in ('first','last')
if a.find(t) is not None) for a in node.findall('author')],
paper_url=page, pdf_url=page[:-1]+'.pdf', doi=text('doi'),
pages=text('pages'), booktitle=''.join(volume.find('meta/booktitle').itertext()),
metadata_sources=[url], abstract_source=page,
awards=[], award_evidence=[], oral=False, oral_evidence=[],
presentation_original=[], code_urls=[])
record['metadata_hits'] = sorted(set(m.group().lower() for m in FIN.finditer(record['title']+' '+record['abstract'])))
# Every record without an abstract is queued for PDF screening.
record['fulltext_candidate'] = bool(record['metadata_hits'] or not record['abstract'])
records.append(record)
save(ROOT/'checks/track_screening/findings_records.json', records)
save(ROOT/'checks/track_screening/findings_sources.json', manifest)
print(json.dumps(dict(records=len(records), candidates=sum(r['fulltext_candidate'] for r in records),
missing_abstracts=sum(not r['abstract'] for r in records),
venue_year=dict(Counter(f"{r['venue']}/{r['year']}" for r in records))), indent=2))
def inspect(record):
CACHE.mkdir(exist_ok=True)
out = CACHE/(record['id']+'.json')
if out.exists():
cached=json.loads(out.read_text())
if cached.get('inspection_version')==INSPECTION_VERSION:
return apply_length_override(cached)
pdf = CACHE/(record['id']+'.pdf')
result = dict(id=record['id'], source_url=record['pdf_url'], status='unavailable',inspection_version=INSPECTION_VERSION)
try:
if not pdf.exists():
response = requests.get(record['pdf_url'], timeout=70,
headers={'User-Agent':'OpenFinMind-LiteratureCollection/1.0'})
response.raise_for_status()
assert response.content.startswith(b'%PDF-'), 'Not a PDF'
pdf.write_bytes(response.content)
text = subprocess.check_output(['pdftotext','-layout',str(pdf),'-'], timeout=45,stderr=subprocess.DEVNULL).decode(errors='replace')
(CACHE/(record['id']+'.txt')).write_text(text)
# Financial evidence must precede References, not be a reference-only hit.
match = re.search(r'(?m)^\s*(?:\d+[. ]+)?References\s*$', text)
body = text[:match.start()] if match else text
hits = []
for found in FIN.finditer(body):
snippet = re.sub(r'\s+', ' ', body[max(0,found.start()-180):found.end()+400]).strip()
if snippet not in hits:
hits.append(snippet)
result.update(status='inspected', bytes=pdf.stat().st_size,
sha256=hashlib.sha256(pdf.read_bytes()).hexdigest(),
pages=len([p for p in text.split('\f') if p.strip()]),
fulltext_hits=hits, references_boundary_found=bool(match))
# Conservative content-length evidence, distinct from total PDF pages:
# substantive prose on page 6 before excluded back matter exceeds the
# maximum 5 content pages for a final ACL short paper.
reading = subprocess.check_output(['pdftotext',str(pdf),'-'], timeout=45,stderr=subprocess.DEVNULL).decode(errors='replace')
result['content_exceeds_short_limit'], result['length_inspection'] = body_length_evidence(reading)
first=reading.split('\f')[0]
normalize=lambda s:re.sub('[^a-z0-9]','',s.lower())
title_words=set(re.findall(r'[a-z]{3,}',record['title'].lower()))
result['title_in_first_page']=normalize(record['title']) in normalize(first)
result['title_token_overlap']=len(title_words & set(re.findall(r'[a-z]{3,}',first.lower())))/max(1,len(title_words))
reference=re.search(r'(?im)^\s*References\s*$',reading)
review_body=reading[:reference.start()] if reference else reading
result['references_boundary_found']=bool(reference)
result['focused_finance_hits']=[re.sub(r'\s+',' ',review_body[max(0,m.start()-130):m.end()+230])
for m in FOCUSED_FINANCE.finditer(review_body)]
abstract=re.search(r'(?is)\bAbstract\b(.*?)(?:\n\s*1\s*\n|\n\s*1\s+Introduction|\n\s*Introduction)',first)
if abstract:
result['recovered_abstract']=re.sub(r'\s+',' ',abstract.group(1)).strip()
except Exception as exc:
result['error'] = str(exc)[:400]
result = apply_length_override(result)
save(out, result)
return result
def download():
records=json.loads((ROOT/'checks/track_screening/findings_records.json').read_text())
selected=[r for r in records if r['fulltext_candidate']]
with concurrent.futures.ThreadPoolExecutor(max_workers=6) as pool:
for n,result in enumerate(pool.map(inspect,selected),1):
if n%40==0:
print(f'Inspected {n}/{len(selected)}',flush=True)
results=[json.loads((CACHE/(r['id']+'.json')).read_text()) for r in selected]
save(ROOT/'checks/track_screening/findings_fulltext_results.json',results)
by_id={r['id']:r for r in results}
for record in records:
recovered=by_id.get(record['id'],{}).get('recovered_abstract')
if not record['abstract'] and recovered:
record.update(abstract=recovered,abstract_source=record['pdf_url'])
save(ROOT/'checks/track_screening/findings_records.json',records)
print(dict(Counter(r['status'] for r in results)), flush=True)
if __name__=='__main__':
if sys.argv[1]=='prepare':
prepare(Path(sys.argv[2]))
elif sys.argv[1]=='download':
download()
|