| """Punctuation + truecasing: joint per-word labels = the punctuation mark that follows the word (punct/common.py: 0 . , ? - :) x the |
| word's casing. Casing classes from the original text: 'lower' (no upper-case first letter), 'Cap' (first cased letter upper, not all |
| upper), 'UPPER' (all cased letters upper, >= 2 of them). Words without cased letters (CJK, Arabic, ...) are 'lower'. |
| Inputs are always lowercased (ASR-style); the model restores marks and casing together.""" |
| import os, sys |
| sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "punct")) |
| from common import LABELS as PUNCT, NOSPACE, lower_keep_len |
| CASES = ["lower", "Cap", "UPPER"] |
| LABELS = [f"{p}|{c}" for p in PUNCT for c in CASES] |
| L2I = {l: i for i, l in enumerate(LABELS)} |
|
|
|
|
| def case_of(word): |
| cased = [ch for ch in word if ch.lower() != ch.upper()] |
| if not cased: |
| return 0 |
| if len(cased) >= 2 and all(ch.isupper() for ch in cased): |
| return 2 |
| return 1 if cased[0].isupper() else 0 |
|
|
|
|
| def unit_cases(text, ends, nospace): |
| """case class per unit of a cased text; units end at `ends` (exclusive offsets).""" |
| out, prev = [], 0 |
| for e in ends: |
| if nospace: |
| s = e - 1 |
| else: |
| s = text.rfind(" ", 0, e) + 1 |
| out.append(case_of(text[max(s, prev):e])); prev = e |
| return out |
|
|
|
|
| def apply_case(word, c): |
| if c == 2: |
| return word.upper() |
| if c == 1: |
| for i, ch in enumerate(word): |
| if ch.lower() != ch.upper(): |
| return word[:i] + ch.upper() + word[i + 1:] |
| return word |
|
|
