16 KiB
Fallback Translation and Glossary Bridge · Global Fix Map
🧭 Quick Return to Map
You are in a sub-page of Language.
To reorient, go back here:
- Language — multilingual processing and semantic alignment
- WFGY Global Fix Map — main Emergency Room, 300+ structured fixes
- WFGY Problem Map 1.0 — 16 reproducible failure modes
Think of this page as a desk within a ward.
If you need the full triage and all prescriptions, return to the Emergency Room lobby.
When native-language recall keeps missing the right snippet, switch to a controlled translation bridge with a domain glossary and alias shield. Translate only where needed, protect entities and negations, and verify improvement with ΔS, λ, and coverage.
Open these first
- Visual map and recovery → rag-architecture-and-recovery.md
- End to end retrieval knobs → retrieval-playbook.md
- Traceability schema → retrieval-traceability.md
- Contract the payload → data-contracts.md
- Language overview → multilingual_guide.md
- Tokenizer variance → tokenizer_mismatch.md
- Mixed scripts → script_mixing.md
- Locale normalization → locale_drift.md
- Romanization rules → romanization_transliteration.md
- Proper nouns and aliases → proper_noun_aliases.md
- Language detection → query_language_detection.md
- Analyzer routing → query_routing_and_analyzers.md
- Multilingual ranking → hybrid_ranking_multilingual.md
- Bilingual eval sets → code_switching_eval.md
- Stopwords and morphology → stopword_and_morphology_controls.md
Core acceptance targets
- ΔS(question, retrieved) ≤ 0.45 on three paraphrases and two seeds
- Coverage of target section ≥ 0.70
- λ convergent after the bridge, across native vs pivot language
- No entity corruption or negation loss in the final citation set
- Rank@k improves or remains flat after the bridge is enabled
When to enable the bridge
Enable only if all three hold:
- Native path shows flat-high ΔS across k settings.
- Query language and corpus language differ or the corpus is mixed locale.
- Entity recall improves during a quick pivot test without harming citations.
If any native pipeline item is obviously wrong, fix that first. See tokenizer, analyzer, or morphology pages above.
What usually breaks
| Symptom | Likely cause | Open this |
|---|---|---|
| Correct doc exists yet never ranks in top k | analyzer or tokenizer mismatch between query and store | tokenizer_mismatch.md · query_routing_and_analyzers.md |
| Names translate or transliterate inconsistently | missing alias shield or mixed romanization | proper_noun_aliases.md · romanization_transliteration.md |
| Negations flip meaning after MT | no do-not-translate list for negation tokens | stopword_and_morphology_controls.md |
| CJK queries degrade when pivoting via English | script segmentation and width rules differ by stage | script_mixing.md |
| Turkish/Greek accent fold changes matches | locale normalization not pinned per stage | locale_drift.md |
| Good recall but order is noisy across languages | reranker trained mono-lingual or features not aligned | hybrid_ranking_multilingual.md |
Design: glossary bridge in two modes
Mode A — Query-side pivot Translate the query to the corpus language with a glossary and alias shield. Run retrieval native to the store, then reason in user language.
Mode B — Corpus-side pivot Keep query in user language, retrieve in native, but translate candidate snippets to the user language for reranking and reasoning. Never re-index on the pivot.
Glossary components
- do_not_translate: names, products, codes, unit strings, legal terms.
- preferred_terms: enforce a deterministic mapping for domain words.
- romanization_map: stable transliteration table with 1-to-N aliases.
- negation_and_modality: tokens that must survive intact.
- protected_char_classes: width, diacritics, punctuation class locks.
Trace fields to log
{
"bridge_mode": "A|B",
"pivot_lang": "en|zh|..",
"glossary_hash": "sha256:...",
"alias_set_hash": "sha256:...",
"ΔS_before": 0.xx,
"ΔS_after": 0.yy,
"coverage_before": 0.xx,
"coverage_after": 0.yy
}
Minimal implementation steps
-
Detect language Use the contract from query_language_detection.md. Refuse fallback if detection is unstable.
-
Assemble glossary
- Pull domain terms.
- Add aliases from proper_noun_aliases.md.
- Add romanization table from romanization_transliteration.md.
- Add negation and unit strings to do-not-translate.
-
Choose mode
- Mode A if store is single-locale and analyzers are correct.
- Mode B if store is mixed or analyzers cannot be changed.
-
Run retrieval Route analyzers per query_routing_and_analyzers.md. For Mode B, translate only candidates for reranking.
-
Verify Compute ΔS and coverage. Require λ convergent across two seeds and three paraphrases. Log trace fields.
-
Publish Keep the glossary versioned and pinned in eval reports. Guard with retrieval-traceability.md and data-contracts.md.
Spec: glossary JSON
{
"version": "glossary_acme_finance_2025_08_30",
"pivot_lang": "en",
"do_not_translate": ["Value at Risk", "CAGR", "ROE", "İstanbul", "北京市", "§"],
"preferred_terms": {
"账面价值": "book value",
"净现值": "net present value"
},
"romanization_map": {
"北京市": ["Beijing Shi", "Beijing City"],
"İstanbul": ["Istanbul", "Stamboul"]
},
"negation_and_modality": ["not", "never", "must", "should"],
"protected_char_classes": ["fullwidth_digit", "narrow_no_break_space"]
}
Copy-paste prompt for the LLM step
You have TXTOS and the WFGY Problem Map loaded.
My multilingual issue:
- native_lang: {xx}
- user_lang: {yy}
- mode: {A|B}
- glossary: {do_not_translate, preferred_terms, romanization_map, negation_and_modality}
- question: "{user_question}"
- candidates: [{snippet_id, text, source_url}...]
Do:
1) Apply the glossary strictly. Protect names, units, negations.
2) Perform cite-then-explain. If citations are weak, return the minimal fix and do not fabricate.
3) Return JSON:
{ "bridge_mode": "A|B", "pivot_lang": "en|...", "citations": [...],
"answer": "...", "ΔS": 0.xx, "coverage": 0.xx, "λ_state": "→|←|<>|×",
"next_fix": "..." }
Keep it auditable and short.
Eval protocol
- Use bilingual and code-switching sets from code_switching_eval.md.
- Compare native vs bridge on the same questions and seeds.
- Accept only if ΔS ≤ 0.45, coverage ≥ 0.70, λ convergent, and entity recall does not regress.
- Report deltas for Rank@k and citation accuracy.
Common gotchas
- Translating the index. Never translate and re-index as a “quick fix”. Pivot only at query or candidate stage.
- Letting the MT rewrite units or numbers. Add them to do-not-translate.
- Dropping diacritics or width during translation. Pin normalization from locale_drift.md and script_mixing.md.
- Reranking with a mono-lingual model. If scores are noisy across languages, follow hybrid_ranking_multilingual.md.
🔗 Quick-Start Downloads (60 sec)
| Tool | Link | 3-Step Setup |
|---|---|---|
| WFGY 1.0 PDF | Engine Paper | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>” |
| TXT OS (plain-text OS) | TXTOS.txt | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
🧭 Explore More
| Module | Description | Link |
|---|---|---|
| WFGY Core | WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack | View → |
| Problem Map 1.0 | Initial 16-mode diagnostic and symbolic fix framework | View → |
| Problem Map 2.0 | RAG-focused failure tree, modular fixes, and pipelines | View → |
| Semantic Clinic Index | Expanded failure catalog: prompt injection, memory bugs, logic drift | View → |
| Semantic Blueprint | Layer-based symbolic reasoning & semantic modulations | View → |
| Benchmark vs GPT-5 | Stress test GPT-5 with full WFGY reasoning suite | View → |
| 🧙♂️ Starter Village 🏡 | New here? Lost in symbols? Click here and let the wizard guide you through | Start → |
👑 Early Stargazers: See the Hall of Fame
⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.
