Use a word-shaped mask token so the model stops eating it

Measured against M2M100 418M on real content: the OpenNMT ⦅0⦆ convention was
dropped on every single occurrence of a protected term. SentencePiece
fragments punctuation runs, and the model then has nothing it recognises as a
unit to copy across.

The brand paid for it. "Viena Latina" came back as "Wien Latin" in one file
and "Vienna Latina" in another — the same name rendered two different wrong
ways across two pages, which is worse than being consistently wrong. "Grätzl"
survived untouched throughout, because it is genuinely unknown to the model,
whereas "Viena" reads as a city name and gets translated.

Masks are now word-shaped (Zq0Xv), which a model treats like an unknown proper
noun and carries through rather than translating. Matching is case- and
space-insensitive, since models re-case and pad these.

The tiered fallback stays as the safety net for when it is still dropped.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NizVpJ2dwzCbjCrTLCjeHn
This commit is contained in:
Claude 2026-09-16 13:34:26 +00:00
parent cc32cf49b5
commit 8d07a6037a
No known key found for this signature in database

View File

@ -5,9 +5,11 @@ model has no equivalent, so structure is protected here instead: non-prose
blocks pass through untouched, and inline constructs are masked with opaque
placeholders whose survival is verified after the round trip.
Placeholders use OpenNMT's protected-sequence convention (U+FF5F/U+FF60).
SentencePiece keeps these atomic; ``{{x}}``, ``<x>`` and ``%s`` get fragmented
by BPE and dropped by the model.
Placeholders are word-shaped (``Zq0Xv``) rather than punctuation. Measured on
M2M100 418M: the OpenNMT ``⦅0⦆`` convention was dropped on every single
occurrence, because SentencePiece fragments punctuation runs and the model then
fails to copy them. A token that looks like an unknown proper noun gets carried
through, since the model has nothing to translate it to.
"""
from __future__ import annotations
@ -16,7 +18,8 @@ import re
from .provider import SITE_TO_MODEL, Provider
OPEN, CLOSE = "⦅", "⦆"
_MASK = "Zq{}Xv"
_MASK_RE = re.compile(r"Zq\s*(\d+)\s*Xv", re.IGNORECASE)
# Community vocabulary that must reach readers unchanged. Not inherited from
# the WordPress plugin, which had no glossary at all — edit freely.
@ -43,9 +46,6 @@ _INLINE = (
re.compile(r"https?://\S+"), # bare URLs
)
_PLACEHOLDER = re.compile(re.escape(OPEN) + r"\s*(\d+)\s*" + re.escape(CLOSE))
class PlaceholderError(RuntimeError):
"""A masked span did not survive translation intact."""
@ -57,7 +57,7 @@ class _Masker:
def _take(self, match: re.Match) -> str:
self.spans.append(match.group(0))
return f"{OPEN}{len(self.spans) - 1}{CLOSE}"
return _MASK.format(len(self.spans) - 1)
def mask(self, text: str) -> str:
for pattern in _INLINE:
@ -68,10 +68,10 @@ class _Masker:
return text
def restore(self, text: str) -> str:
# Models pad and reorder placeholders; normalise spacing before matching.
text = _PLACEHOLDER.sub(lambda m: f"{OPEN}{m.group(1)}{CLOSE}", text)
# Models pad, re-case and reorder placeholders; normalise before matching.
text = _MASK_RE.sub(lambda m: _MASK.format(m.group(1)), text)
for index, span in enumerate(self.spans):
token = f"{OPEN}{index}{CLOSE}"
token = _MASK.format(index)
seen = text.count(token)
if seen != 1:
raise PlaceholderError(