njand's picture
Upload scripts/build_dataset.py
6c2380d verified
Raw
History Blame Contribute Delete
20.9 kB
#!/usr/bin/env python3
"""
Latin ASR Post-Processing Dataset Builder
Downloads the CLTK Latin Library in memory (ZIP) and LLPSI speech dataset,
normalizes text into pure classical i/u orthography, dynamically expands
indeclinable Roman numerals to Latin cardinal words, extracts word tokens and
casing/punctuation tags, and pushes the stratified dataset directly to Hugging Face Hub.
"""
import argparse
import io
import os
import random
import re
import sys
import unicodedata
import zipfile
from collections import Counter
import numpy as np
import requests
import nltk
from nltk.tokenize import sent_tokenize
from tqdm import tqdm
from datasets import Dataset, DatasetDict, load_dataset
from sklearn.model_selection import train_test_split
# Ensure NLTK tokenizers are available
for resource in ["punkt", "punkt_tab"]:
try:
nltk.data.find(f"tokenizers/{resource}")
except LookUpError:
nltk.download(resource, quiet=True)
# =====================================================================
# CONFIGURATION & LOOKUP DATASETS
# =====================================================================
# Standard English stopwords excluding Latin false-positives ('his', 'as')
ENGLISH_STOPWORDS = {
# Archive, Web Infrastructure & Editorial Terms
"library", "classics", "miscellany", "home", "homepage", "index", "latin",
"contents", "site", "html", "http", "https", "www", "com", "org", "edu",
"christian", "medieval", "neo-latin", "prepared", "proof", "read",
"proof-read", "proofread", "edited", "archive", "edition", "published",
"publisher", "press", "university", "translated", "transcribed",
"transcription", "scanned", "text", "texts", "source", "note", "notes",
"footnote", "volume", "vol", "book", "chapter", "section", "page", "pages",
"line", "lines", "version", "revised", "reprinted",
# High-Confidence English Function Words
"the", "of", "and", "to", "you", "that", "was", "for", "on", "are",
"with", "they", "this", "have", "from", "one", "had",
"by", "word", "but", "not", "what", "all", "were", "we", "when",
"your", "can", "said", "there", "use", "each", "which", "how", "their",
"if", "will", "up", "other", "about", "out", "many", "then", "them",
"these", "some", "would"
}
PRAENOMINA_1ST_2ND_STEMS = {
"A": "Aul", "Ap": "Appi", "C": "Gai", "Cn": "Gnae", "D": "Decim",
"F": "Faust", "H": "Host", "L": "Luci", "M": "Marc", "M'": "Mani",
"M′": "Mani", "M’": "Mani", "Mam": "Mamerc", "N": "Numeri", "Oct": "Octavi",
"P": "Publi", "Post": "Postum", "Pro": "Procul", "Q": "Quint", "S": "Spuri",
"Sec": "Secund", "Seq": "Secund", "Ser": "Servi", "Sex": "Sext", "Sp": "Spuri",
"St": "Stati", "T": "Tit", "Ti": "Tiberi", "V": "Vibi", "Vol": "Voles",
"Vop": "Vopisc"
}
DECLENSION_3RD_PRAENOMINA = {
"Opet": {"nom": "Opiter", "acc": "Opitrem", "gen": "Opitris", "dat": "Opitri", "abl": "Opitre"},
"Sert": {"nom": "Sertor", "acc": "Sertorem", "gen": "Sertoris", "dat": "Sertori", "abl": "Sertore"},
"Mai": {"nom": "Maio", "acc": "Maiorem", "gen": "Maioris", "dat": "Maiori", "abl": "Maiore"},
"Min": {"nom": "Mino", "acc": "Minorem", "gen": "Minoris", "dat": "Minori", "abl": "Minore"},
}
PUNCT_MAP = {
'': 'NONE',
'.': 'PERIOD',
',': 'COMMA',
':': 'COLON',
';': 'SEMICOLON',
'!': 'EXCLAMATION',
'?': 'QUESTION'
}
ROMAN_NUMERAL_REGEX = r"^M{0,4}(CM|CD|D?C{0,3})(XC|XL|L?X{0,3})(IX|IV|V?I{0,3})$"
# Lowercase Roman numerals that collide with real Latin vocabulary (e.g. vi -> vī = "by force").
LOWERCASE_ROMAN_EXCLUSIONS = {"i", "vi"}
UNITS_4_TO_9 = {
4: "quattuor", 5: "quinque", 6: "sex", 7: "septem", 8: "octo", 9: "novem"
}
TEENS_AND_TENS = {
10: "decem", 11: "undecim", 12: "duodecim", 13: "tredecim", 14: "quattuordecim",
15: "quindecim", 16: "sedecim", 17: "septendecim", 18: "duodeviginti",
19: "undeviginti", 20: "viginti", 30: "triginta", 40: "quadraginta",
50: "quinquaginta", 60: "sexaginta", 70: "septuaginta", 80: "octoginta", 90: "nonaginta"
}
# Terminal punctuation allowed at the end of a sentence
TERMINAL_PUNCTUATION = {".", "?", "!"}
CLOSING_QUOTES = '"”»\'’'
# Citation and editorial abbreviations whose periods should NOT trigger sentence splits
SCHOLASTIC_ABBREVS = (
r"\b(Corinth|Cor|Gal|Eph|Phil|Col|Thess|Tim|Tit|Philem|Hebr|Pet|Joan|Apoc|"
r"Matt|Marc|Luc|Act|Rom|Gen|Exod|Lev|Num|Deut|Jos|Judic|Reg|Paral|Esd|Tob|"
r"Judith|Esth|Job|Ps|Prov|Eccl|Cant|Sap|Sir|Is|Jer|Lam|Bar|Ezech|Dan|Osee|"
r"Joel|Amos|Abd|Jon|Mich|Nah|Hab|Soph|Agg|Zach|Mal|Mach|cap|v|vv|f|fol|lib|"
r"p|pp|ibid|ca|seq|e\.g|i\.e|S|St|Th|q|a|art|ad|resp|dist|m|n)\."
)
# =====================================================================
# EDITORIAL & METADATA FILTERING
# =====================================================================
def strip_diacritics(text: str) -> str:
"""Strips macrons, accents, and converts ligatures (æ/œ -> ae/oe)."""
text = text.replace("æ", "ae").replace("œ", "oe").replace("Æ", "Ae").replace("Œ", "Oe")
nfd = unicodedata.normalize("NFD", text)
filtered = "".join(c for c in nfd if unicodedata.category(c) != "Mn")
return unicodedata.normalize("NFC", filtered)
def strip_section_numbers(text: str) -> str:
"""Strips bracketed or leading section numbers e.g. [1], [1.1], 1."""
text = re.sub(r"\[\s*[\d\s.,IVXLCDM]+\s*\]", "", text)
return re.sub(r"^\s*\d+\b\.?\s*", "", text)
def is_editorial_or_metadata(text: str) -> bool:
"""Identifies editorial headnotes, dates, and apparatus criticus entries."""
clean = text.strip()
if not clean:
return True
if re.search(r"\b(Scr|ep|epp|cod|codd|pag|v|vv|a\.u\.c|ed)\b\.", clean, re.IGNORECASE):
return True
if re.search(r"^\s*([ivxlcdm\d]+\s+)?(K|Kal|Nones|Non|Ibus|Id)\b", clean, re.IGNORECASE):
return True
return False
def has_all_caps_or_unexpanded_roman(sentence: str) -> bool:
"""Rejects sentences containing ALL-CAPS words or unexpanded Roman numerals."""
words = re.findall(r"\b[A-Z]+\b", sentence)
for w in words:
if len(w) > 1 or re.match(ROMAN_NUMERAL_REGEX, w):
return True
return False
def contains_english(text_line: str) -> bool:
words = set(re.findall(r"\b[a-zA-Z]+\b", text_line.lower()))
return bool(words.intersection(ENGLISH_STOPWORDS))
# =====================================================================
# LATIN NORMALIZATION & DYNAMIC ROMAN NUMERAL HELPERS
# =====================================================================
def roman_to_int(roman: str) -> int:
"""Parses a Roman numeral string into an integer."""
roman_dict = {'I': 1, 'V': 5, 'X': 10, 'L': 50, 'C': 100, 'D': 500, 'M': 1000}
total = 0
prev_val = 0
for char in reversed(roman):
val = roman_dict.get(char, 0)
if val < prev_val:
total -= val
else:
total += val
prev_val = val
return total
def int_to_indeclinable_latin(n: int) -> str | None:
"""
Converts an integer to Latin words ONLY if all constituent components
are strictly indeclinable. Returns None if any component declines.
"""
if n <= 0:
return None
# Hundreds 200-900 decline (ducenti, trecenti, etc.)
hundreds = (n % 1000) // 100
if 2 <= hundreds <= 9:
return None
# Thousands > 1000 use 'milia' which declines as a neuter noun
thousands = n // 1000
if thousands > 1:
return None
parts = []
if thousands == 1:
parts.append("mille")
if hundreds == 1:
parts.append("centum")
rem = n % 100
if rem > 0:
if rem in TEENS_AND_TENS:
parts.append(TEENS_AND_TENS[rem])
else:
tens_val = (rem // 10) * 10
unit_val = rem % 10
# Reject if units are 1, 2, or 3 (unus, duo, tres decline)
if unit_val in (1, 2, 3) or tens_val not in TEENS_AND_TENS or unit_val not in UNITS_4_TO_9:
return None
parts.append(f"{TEENS_AND_TENS[tens_val]} {UNITS_4_TO_9[unit_val]}")
return " ".join(parts) if parts else None
def expand_safe_roman_numerals(text: str) -> str:
"""
Dynamically converts valid indeclinable Roman numerals (upper and safe lower)
to Latin cardinal words.
"""
def replacer(match):
token = match.group(0)
# 1. Protect real Latin words that look like lowercase Roman numerals
if token.islower() and token in LOWERCASE_ROMAN_EXCLUSIONS:
return token
upper_token = token.upper()
# 2. Validate Roman numeral syntax (e.g. XIV is valid, VICI is not)
if not re.match(ROMAN_NUMERAL_REGEX, upper_token):
return token
# 3. Convert and check if grammatically indeclinable
val = roman_to_int(upper_token)
latin_words = int_to_indeclinable_latin(val)
if latin_words is None:
return token
# 4. Preserve matching casing
if token.isupper():
return latin_words.upper()
elif token.istitle():
return latin_words.capitalize()
else:
return latin_words.lower()
# Match tokens consisting entirely of Roman numeral characters (case-insensitive)
return re.sub(r"\b[a-zA-Z]+\b", replacer, text)
def inflect_praenomen(abbrev: str, next_word: str) -> str | None:
clean_abbrev = abbrev.rstrip(".")
target = next_word.lower()
if clean_abbrev == "Agr":
if target.endswith("am"): return "Agrippam"
if target.endswith("ae"): return "Agrippae"
return "Agrippa"
if clean_abbrev in DECLENSION_3RD_PRAENOMINA:
rules = DECLENSION_3RD_PRAENOMINA[clean_abbrev]
if target.endswith(("em", "am", "um")): return rules["acc"]
if target.endswith("is"): return rules["gen"]
if target.endswith("i"): return rules["dat"]
if target.endswith("e"): return rules["abl"]
return rules["nom"]
if clean_abbrev in PRAENOMINA_1ST_2ND_STEMS:
stem = PRAENOMINA_1ST_2ND_STEMS[clean_abbrev]
if target.endswith("am"): suffix = "am"
elif target.endswith(("um", "em")): suffix = "um"
elif target.endswith("ae"): suffix = "ae"
elif target.endswith(("i", "is")): suffix = "i"
elif target.endswith(("o", "e")): suffix = "o"
elif target.endswith("a") and not target.endswith("ma"): suffix = "a"
else: suffix = "us"
return stem + suffix
return None
def expand_praenomina(text: str) -> str:
pattern = r"\b([A-Z][a-z]{0,3}['′’]?)\.\s+([A-Z][a-z]+)"
def replacer(match):
abbrev, next_word = match.group(1), match.group(2)
expanded = inflect_praenomen(abbrev, next_word)
if expanded is None:
return match.group(0)
return f"{expanded} {next_word}"
return re.sub(pattern, replacer, text)
def normalize_iu(text: str) -> str:
"""
Normalizes Latin text to standard classical i/u orthography:
- uva -> uua, virgo -> uirgo, jam -> iam.
- Compound -iacere forms: ejicio -> eicio, conjicio -> conicio, objicit -> obicit.
"""
# 1. Handle compound verbs from -iacere (-jic- / -jici- after prefix/vowel -> -ic- / -ici-)
text = re.sub(r'([a-zA-Z])j[iI]', r'\1i', text)
text = re.sub(r'([a-zA-Z])J[iI]', r'\1I', text)
# 2. Target only compound -iic- verb forms (eiicio -> eicio)
text = re.sub(r'([aeiouAEIOU])ii([cC])', r'\1i\2', text)
# 3. Convert all remaining J/j to I/i
text = text.replace('j', 'i').replace('J', 'I')
# 4. Convert all V/v to U/u
text = text.replace('v', 'u').replace('V', 'U')
return text
def clean_punctuation(text: str) -> str:
cleaned = re.sub(r"[^\w\s.,?!:;]", "", text)
cleaned = re.sub(r"\s+([.,?!:;])", r"\1", cleaned)
return re.sub(r"\s+", " ", cleaned).strip()
def mask_citation_periods(text: str) -> tuple[str, dict[str, str]]:
"""Masks periods in citation abbreviations so NLTK sent_tokenize ignores them."""
placeholder_map = {}
def repl(match):
key = f"__ABBR_{len(placeholder_map)}__"
placeholder_map[key] = match.group(0)
return key
masked_text = re.sub(SCHOLASTIC_ABBREVS, repl, text, flags=re.IGNORECASE)
return masked_text, placeholder_map
def unmask_citation_periods(text: str, placeholder_map: dict[str, str]) -> str:
"""Restores original citation abbreviations after sentence splitting."""
for key, orig in placeholder_map.items():
text = text.replace(key, orig)
return text
def is_valid_sentence(sentence_text: str) -> bool:
text = sentence_text.strip()
if not text or len(text.split()) < 3:
return False
# Strip trailing quotes using the constant
core_text = text.rstrip(CLOSING_QUOTES)
if not core_text or core_text[-1] not in TERMINAL_PUNCTUATION:
return False
if has_all_caps_or_unexpanded_roman(text):
return False
return True
def capitalize_first_letter(s: str) -> str:
"""Capitalizes the first alphabetic character in the string, skipping leading punctuation/whitespace."""
for i, char in enumerate(s):
if char.isalpha():
return s[:i] + char.upper() + s[i + 1 :]
return s
def normalize_sentence(sentence: str) -> str:
text = sentence.strip()
text = strip_diacritics(text)
text = normalize_iu(text)
text = clean_punctuation(text)
if not text:
return ""
# Ensure sentence ends in true terminal punctuation
if text[-1] not in TERMINAL_PUNCTUATION:
text += "."
# Force the first actual letter to uppercase
return capitalize_first_letter(text)
# =====================================================================
# CORPUS PROCESSING & TOKEN EXTRACTION
# =====================================================================
def process_latin_corpus(
raw_text: str, max_merge_len: int = 1000, p_merge: float = 0.50
) -> str:
raw_paragraphs = re.split(r"\n\s*\n+", raw_text.strip())
cleaned_paragraphs = []
for block in raw_paragraphs:
lines = [line.strip() for line in block.splitlines() if line.strip()]
if not lines:
continue
single_line_paragraph = " ".join(lines)
# 1. Strip diacritics and ligatures early
prep_paragraph = strip_diacritics(single_line_paragraph)
if contains_english(prep_paragraph) or is_editorial_or_metadata(prep_paragraph):
continue
# 2. Strip section numbers, expand praenomina & expand safe Roman numerals
prep_paragraph = strip_section_numbers(prep_paragraph)
prep_paragraph = expand_praenomina(prep_paragraph)
prep_paragraph = expand_safe_roman_numerals(prep_paragraph)
raw_sentences = sent_tokenize(prep_paragraph)
valid_normalized_sentences = []
for sentence_str in raw_sentences:
sentence_clean = sentence_str.strip()
if is_valid_sentence(sentence_clean) and not is_editorial_or_metadata(sentence_clean):
norm_sent = normalize_sentence(sentence_clean)
if norm_sent:
valid_normalized_sentences.append(norm_sent)
if valid_normalized_sentences:
merged = []
current = valid_normalized_sentences[0]
for nxt in valid_normalized_sentences[1:]:
if len(current) + 1 + len(nxt) <= max_merge_len and random.random() < p_merge:
current = f"{current} {nxt}"
else:
merged.append(current)
current = nxt
merged.append(current)
cleaned_paragraphs.append("\n\n".join(merged))
return "\n\n".join(cleaned_paragraphs)
def build_latin_dataset(limit_files=None) -> dict[str, list[str]]:
"""Downloads the CLTK repo as an in-memory ZIP archive and processes text files."""
zip_url = "https://github.com/cltk/lat_text_latin_library/archive/refs/heads/master.zip"
print("Downloading CLTK Latin Library archive into RAM...")
response = requests.get(zip_url)
response.raise_for_status()
dataset = {}
with zipfile.ZipFile(io.BytesIO(response.content)) as z:
txt_files = [f for f in z.namelist() if f.endswith(".txt")]
if limit_files is not None:
txt_files = txt_files[:limit_files]
print(f"Processing {len(txt_files)} files from memory...")
for file_path in tqdm(txt_files):
with z.open(file_path) as f:
raw_text = f.read().decode("utf-8", errors="ignore")
cleaned_text = process_latin_corpus(raw_text)
sentences = [s.strip() for s in cleaned_text.splitlines() if s.strip()]
if sentences:
dataset[file_path] = sentences
return dataset
def extract_token_features(text: str) -> dict[str, list[str]]:
pattern = r'([A-Za-z]+)([\.,:;!\?]?)'
matches = re.findall(pattern, text)
tokens = []
tags = []
for word, punct in matches:
if not word:
continue
casing = "TITLE" if word[0].isupper() else "LOWER"
p_label = PUNCT_MAP.get(punct, 'NONE')
tokens.append(word.lower())
tags.append(f"{casing}_{p_label}")
return {"tokens": tokens, "tags": tags}
# =====================================================================
# MAIN PIPELINE
# =====================================================================
def main():
parser = argparse.ArgumentParser(description="Compile and push normalized Latin ASR post-processing dataset.")
parser.add_argument("--repo-id", type=str, default="njand/latin-asr-post-processing-dataset", help="Hugging Face repo ID")
parser.add_argument("--limit-files", type=int, default=None, help="Limit number of CLTK files processed (for testing)")
parser.add_argument("--test-ratio", type=float, default=0.05, help="Test split ratio")
parser.add_argument("--seed", type=int, default=42, help="Random seed")
parser.add_argument("--no-push", action="store_true", help="Do not push dataset to Hugging Face Hub")
args = parser.parse_args()
np.random.seed(args.seed)
random.seed(args.seed)
# 1. Download & Process CLTK Corpus in RAM
latin_dataset = build_latin_dataset(limit_files=args.limit_files)
# 2. Load & Process LLPSI Speech Dataset
print("Loading and normalizing LLPSI speech dataset...")
llpsi_ds = load_dataset("njand/llpsi-speech-dataset", split="train", columns=["text"])
llpsi_sents = [normalize_sentence(row["text"]) for row in llpsi_ds if row.get("text")]
latin_dataset["llpsi"] = llpsi_sents
# 3. Extract Tokens and Casing/Punctuation Tags
print("Extracting token features and target tags...")
structured_data = []
for source, sentences in tqdm(latin_dataset.items()):
for sentence in sentences:
features = extract_token_features(sentence)
if features["tokens"]:
structured_data.append({
"source": source,
"tokens": features["tokens"],
"tags": features["tags"]
})
total_tokens = sum(len(item["tokens"]) for item in structured_data)
print(f"Total samples (lines): {len(structured_data):,}")
print(f"Total tokens: {total_tokens:,}")
# 4. Stratified Train/Test Split (handling singletons)
source_counts = Counter(item["source"] for item in structured_data)
stratifiable_items = []
stratifiable_labels = []
train_data = []
test_data = []
for item in structured_data:
source = item["source"]
if source_counts[source] < 2:
if np.random.rand() < args.test_ratio:
test_data.append(item)
else:
train_data.append(item)
else:
stratifiable_items.append(item)
stratifiable_labels.append(source)
strat_train, strat_test = train_test_split(
stratifiable_items,
test_size=args.test_ratio,
random_state=args.seed,
stratify=stratifiable_labels
)
train_data.extend(strat_train)
test_data.extend(strat_test)
print(f"Train size: {len(train_data):,} samples | Test size: {len(test_data):,} samples")
# 5. Build HF DatasetDict
dataset_dict = DatasetDict({
"train": Dataset.from_list(train_data),
"test": Dataset.from_list(test_data)
})
# 6. Push to Hugging Face Hub
if not args.no_push:
print(f"Uploading dataset to Hugging Face Hub: {args.repo_id}")
dataset_dict.push_to_hub(args.repo_id, private=False)
print("Upload complete!")
else:
print("Skipping Hub upload (--no-push flag active).")
if __name__ == "__main__":
main()