Datasets:

ArXiv:
megawika-2 / rehydrate-citations.py
ccmaymay's picture
Add files using upload-large-folder tool
5889057 verified
raw
history blame
8.83 kB
from collections import defaultdict
from collections.abc import Iterable
import copy
from io import TextIOBase
from functools import lru_cache
import json
from typing import Any, Optional
import click
import datetime
import ftfy # type: ignore
import requests
from tqdm import tqdm
import trafilatura # type: ignore
from humanfriendly import format_number
# Download parameters
DOWNLOAD_TIMEOUT = 10 # Timeout in seconds of each HTTP request
DOWNLOAD_BUFFER_SIZE = 65_536 # Size of each HTTP request chunk in bytes
CODE_MAX_NUM_CHARS = 1_000_000 # Downloaded source code character threshold (applied after decoding to text)
# Extraction parameters
INCLUDE_COMMENTS = False # Whether to include WP user comments in extracted text
INCLUDE_TABLES = True # Whether to include markdown for HTML tables in extracted text
INCLUDE_LINKS = True # Whether to include markdown for HTML links in extracted text
TEXT_MIN_NUM_WORDS = 100 # Extracted text word (whitespace-delimited token) threshold; None to disable
TEXT_MIN_NUM_CHARS = None # Extracted text character threshold; None to disable
class DownloadTooLarge(Exception):
pass
def download_url(url: str) -> tuple[Optional[str], dict[str, Any]]:
"""
Fetch content at URL, returning downloaded source code and metadata.
Metadata is represented as a dictionary containing Citation keys and values.
"""
source_code: Optional[str] = None
content_type: Optional[str] = None
num_chars: Optional[int] = None
error: Optional[str] = None
download_date = datetime.datetime.now().astimezone().isoformat()
try:
with requests.get(url, stream=True, timeout=DOWNLOAD_TIMEOUT) as response:
content_type = response.headers.get('Content-Type')
try:
content = ''
for chunk in response.iter_content(chunk_size=DOWNLOAD_BUFFER_SIZE, decode_unicode=True):
content += chunk
if len(content) > CODE_MAX_NUM_CHARS:
raise DownloadTooLarge()
source_code = content
num_chars = len(content)
except DownloadTooLarge:
source_code = None
error = 'Download is too large'
except Exception as ex:
source_code = None
error = f'{type(ex).__name__}: {ex}'
if not source_code:
source_code = None
error = 'Download is empty'
metadata = dict(
source_code_content_type=content_type,
source_code_num_chars=num_chars,
source_download_date=download_date,
source_download_error=error,
)
return (source_code, metadata)
def count_words(text: str) -> int:
return len(text.split())
def has_sufficient_word_count(source_text: str) -> bool:
return not TEXT_MIN_NUM_WORDS or count_words(source_text) > TEXT_MIN_NUM_WORDS
def has_sufficient_char_count(source_text: str) -> bool:
return not TEXT_MIN_NUM_CHARS or len(source_text) > TEXT_MIN_NUM_CHARS
def extract_source_text(source_code: str) -> dict[str, Any]:
"""
Given a source's HTML, extract and return textual content and error information
as a dictionary containing citation keys and values.
"""
text: Optional[str] = None
error: Optional[str] = None
try:
text = trafilatura.extract(
source_code,
include_comments=INCLUDE_COMMENTS,
include_tables=INCLUDE_TABLES,
include_links=INCLUDE_LINKS,
)
if text is not None:
text = ftfy.fix_text(text)
except Exception as ex:
error = f'{type(ex).__name__}: {ex}'
text = None
else:
if not text:
error = 'Text is empty'
text = None
elif not has_sufficient_word_count(text):
error = f'Text is too short ({format_number(count_words(text))} words)'
text = None
elif not has_sufficient_char_count(text):
error = f'Text is too short ({format_number(len(text))} characters)'
text = None
return dict(
source_text=text,
source_extract_error=error,
)
@lru_cache
def scrape_source_fields(url: str) -> dict[str, Any]:
"""
Download source code at URL and extract text from the downloaded source code,
returning a dictionary of fields to add to the corresponding citation.
This abstraction is a little awkward, but it facilitates caching.
"""
download_fields: dict[str, Any] = {}
extraction_fields: dict[str, Any] = {}
if url:
(source_code, download_fields) = download_url(url)
if source_code:
extraction_fields = extract_source_text(source_code)
return download_fields | extraction_fields
def scrape_source(citation: dict[str, Any]) -> dict[str, Any]:
"""
Scrape sources for citation, storing scraped content in citation *in-place.*
"""
if citation['url']:
citation['dehydrated_citation'] = copy.deepcopy(citation)
citation.update(scrape_source_fields(citation['url']))
return citation
def update_citations_in_excerpts_with_citations(
excerpts_with_citations: list[dict[str, Any]],
new_citations: Iterable[dict[str, Any]],
) -> list[dict[str, Any]]:
"""
Replace citations in `excerpts_with_citations` with the corresponding citations from `new_citations`,
EXCEPT for the `char_index` field (leaving it as-is).
Citations are aligned/matched by their content field.
If a citation content string appears multiple times in `new_citations`, matching citations in
`excerpts_with_citations` will be replaced with the final match in `new_citations`.
`excerpts_with_citations` is modified *in-place,* and the updated list is returned.
"""
ewc_citation_indices_by_key: dict[str, set[tuple[int, int]]] = defaultdict(set)
for (i, ewc) in enumerate(excerpts_with_citations):
for (j, citation) in enumerate(ewc['citations']):
key = citation['content']
ewc_citation_indices_by_key[key].add((i, j))
for new_citation in new_citations:
key = new_citation['content']
for (i, j) in ewc_citation_indices_by_key[key]:
excerpts_with_citations[i]['citations'][j] |= {
k: v
for (k, v) in new_citation.items()
if k != 'char_index'
}
return excerpts_with_citations # was modified in-place
def scrape_article_sources(article: dict[str, Any]) -> dict[str, Any]:
"""
Scrape sources for all web citations in article, storing scraped content in article citation
objects *in-place.* Return modified article for convenience.
"""
all_citations = []
for element in article['elements']:
if element['type'] == 'paragraph':
for sentence in element['sentences']:
for citation in sentence['citations']:
scrape_source(citation) # modifies in-place
all_citations.append(citation)
elif element['type'] == 'heading':
for citation in element['citations']:
scrape_source(citation) # modifies in-place
all_citations.append(citation)
# modifies in-place
update_citations_in_excerpts_with_citations(article['excerpts_with_citations'], all_citations)
return article
@click.command(context_settings={'show_default': True})
@click.argument('input_file', type=click.File())
@click.argument('output_file', type=click.File(mode='w'))
@click.option('--stream', '-s', is_flag=True,
help='Read one article at a time, rehydrating it and writing the output article before reading the next'
' article. The default behavior is to read all articles into memory before processing them.')
def main(input_file: TextIOBase, output_file: TextIOBase, stream: bool) -> None:
"""
Scrape sources for all MegaWika 2 articles in JSON-lines (one JSON-encoded article per line) file INPUT_FILE,
updating each web citation with source-scraping content and metadata and writing the articles with source content
to JSON-lines file OUTPUT_FILE.
Each web citation will be updated with `source_text` *and* source-scraping metadata fields; the original
(dehydrated, metadata-only) web citation will be stored under the new citation field `dehydrated_citation`.
"""
articles: Iterable[dict[str, Any]] = (json.loads(line) for line in input_file)
if not stream:
click.echo('Reading all articles into memory')
articles = list(articles)
for article in tqdm(articles, desc='Rehydrating', unit='article'):
scrape_article_sources(article)
output_file.write(json.dumps(article) + '\n')
click.echo('Done')
if __name__ == '__main__':
main()