File size: 3,083 Bytes
5daac29 | 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 | """Text cleaning for corpus building.
clean_text() keeps its original signature/behavior for existing callers
(Datasets.Pipelines.ingestion_pipeline, Datasets.__init__) -- the new
Gutenberg-boilerplate stripping and unicode normalization are opt-in via
new parameters, not silently applied to every caller.
"""
from __future__ import annotations
import re
import unicodedata
# Project Gutenberg's start/end markers have used a few different exact
# wordings over the decades; this matches all the common variants (both the
# newer "*** START OF THE PROJECT GUTENBERG EBOOK <title> ***" style and the
# older "*** START OF THIS PROJECT GUTENBERG EBOOK <title> ***" / numbered
# "*** START OF THE PROJECT GUTENBERG EBOOK, <NN> ***" style).
_GUTENBERG_START_RE = re.compile(
r"\*\*\*\s*START OF (?:THE|THIS) PROJECT GUTENBERG EBOOK.*?\*\*\*",
re.IGNORECASE | re.DOTALL,
)
_GUTENBERG_END_RE = re.compile(
r"\*\*\*\s*END OF (?:THE|THIS) PROJECT GUTENBERG EBOOK.*?\*\*\*",
re.IGNORECASE | re.DOTALL,
)
def strip_gutenberg_boilerplate(text: str) -> str:
"""Cut everything up to and including the START marker, and everything
from the END marker onward -- these wrap Project Gutenberg's legal
boilerplate, not the book's actual content. If a marker isn't found
(a minority of older-format texts), that side is left untouched rather
than guessing, since guessing risks silently truncating real content.
"""
start_match = _GUTENBERG_START_RE.search(text)
if start_match:
text = text[start_match.end():]
end_match = _GUTENBERG_END_RE.search(text)
if end_match:
text = text[:end_match.start()]
return text
def normalize_unicode(text: str) -> str:
"""NFKC normalization only -- deliberately does NOT rewrite smart quotes,
em-dashes, etc. to ASCII equivalents. Those are legitimate characters a
tokenizer should learn, not noise; over-aggressive normalization here
would make the corpus less representative of real text, not cleaner.
"""
return unicodedata.normalize("NFKC", text)
def clean_text(
text: str,
strip_gutenberg: bool = False,
normalize: bool = False,
) -> str:
text = (text or "").replace("\x00", " ")
if strip_gutenberg:
text = strip_gutenberg_boilerplate(text)
if normalize:
text = normalize_unicode(text)
text = re.sub(r"[ \t]+", " ", text)
text = re.sub(r"\n{3,}", "\n\n", text)
return text.strip()
def is_low_quality(text: str, min_chars: int = 500, min_alpha_ratio: float = 0.6) -> bool:
"""Cheap quality filter: too short, or too little alphabetic content
(catches OCR garbage, tables-of-numbers, near-empty files) to be worth
keeping in a text corpus. Deliberately conservative -- false negatives
(letting some junk through) are cheaper than false positives (throwing
away real text), since dedup/downstream filtering catches more later.
"""
if len(text) < min_chars:
return True
alpha = sum(1 for c in text if c.isalpha())
return (alpha / len(text)) < min_alpha_ratio
|