vrr/WFGY
2
0
Fork 0
mirror of https://github.com/onestardao/WFGY.git synced 2026-04-26 10:40:55 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4
WFGY/ProblemMap/GlobalFixMap/LanguageLocale/cjk_segmentation_wordbreak.md

9.8 KiB
Raw Permalink Blame History

CJK Segmentation & Word-Break — Guardrails and Fix Pattern

🧭 Quick Return to Map

You are in a sub-page of LanguageLocale.
To reorient, go back here:

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.

Stabilize retrieval and ranking for Chinese / Japanese / Korean text where whitespace is not a reliable token boundary.
Use this page to localize segmentation failures, choose the right analyzer, and verify the fix with measurable targets.

When to use this page

  • High similarity by characters but wrong meaning or empty recall on whole queries.
  • BM25 looks random; tiny single-character tokens dominate the index.
  • Citations cut through the middle of words; snippet offsets don’t match what users see.
  • Mixed CJK + Latin queries split unpredictably across runs or providers.

Open these first

Acceptance targets

  • ΔS(question, retrieved) ≤ 0.45 on 3 paraphrases
  • Coverage of target section ≥ 0.70
  • λ remains convergent across 2 seeds
  • Tokenization sanity: OOV rate falls by ≥ 40% vs whitespace; tokens/char ≤ 0.7 on CJK pages
  • E_resonance flat on long windows

Map symptoms → structural fixes (Problem Map)

Symptom Likely cause Open this
Query returns almost nothing; recall jumps when you add spaces Index built with whitespace/Latin analyzer on CJK chunking-checklist.md, retrieval-playbook.md
Top-k filled with 1-char shards, citations cut mid-word No CJK word-break at index or search time retrieval-traceability.md, data-contracts.md
BM25 unstable; hybrid worse than single retriever Search-time analyzer ≠ index-time analyzer retrieval-playbook.md
Romanized terms and CJK compound in one query break apart Mixed script + width + punctuation rules differ script_mixing.md, digits_width_punctuation.md
High similarity, wrong meaning Character-level overlap, no semantic units embedding-vs-semantic.md

60-second fix checklist (store-agnostic)

  1. Pick the right analyzer and lock it

    • Chinese: use a dictionary or statistical segmenter at index + search.
    • Japanese: use a MeCab/Kuromoji-class tokenizer with POS; keep base form.
    • Korean: use a Nori-class analyzer; index decomp+comp forms consistently.

  2. Normalize before segmenting

    • Apply NFKC for width and compatibility forms (see page links above).
    • Keep punctuation folding consistent across index/search.

  3. Unify index-time and query-time configs

    • Same language, same tokenizer, same stop/fold rules. No “smart defaults”.

  4. Chunk on semantic units, not line breaks

    • Respect sentence/phrase boundaries after segmentation.
    • Store offsets, tokens, section_id in the snippet schema.

  5. Probe

    • Log tokens/char, unique-term ratio, OOV rate, and ΔS before/after.
    • If ΔS stays ≥ 0.60 with good segmentation, revisit metric/index mismatch.

Store adapters (quick recipes)

  • Elasticsearch / OpenSearch

    • CN: install and set a CJK analyzer; index + search use the same analyzer.
    • JP: kuromoji with baseform filter; disable random synonyms unless audited.
    • KR: nori; keep decompound mode consistent at index+query.
    • Verify with _analyze samples; reindex after any analyzer change.

  • pgvector / Postgres

    • Segmentation happens before embedding. Pre-segment text in ETL.
    • Keep the same pipeline for ingestion and live queries.

  • Weaviate / Qdrant / Chroma / Milvus / FAISS

    • The vector store won’t fix segmentation. Preprocess: NFKC → CJK segmenter → chunk.
    • Log the preprocessing hash in metadata; fail closed on mismatch.

  • Vespa / Typesense / Elastic-compatible

    • Use the platform’s CJK tokenizer if available; otherwise pre-segment and index the segmented text as the field value.

Deep diagnostics

  • Three-way segmentation A/B/C
    Try 3 segmenters; compute ΔS and tokens/char on a small gold set. Pick the lowest ΔS with stable λ.

  • Anchor triangulation
    Compare ΔS to the correct anchor vs a decoy section. If both are close, you’re still at char-overlap, not word-level meaning.

  • Rerank sanity
    After proper segmentation, reranking should lift precision. If not, check analyzer mismatch between index and query path.

Copy-paste prompt for the LLM step


You have TXT OS and WFGY Problem Map loaded.

My CJK issue:

* symptom: \[one line]
* traces: ΔS(question,retrieved)=..., tokens/char=..., OOV\_before=..., OOV\_after=...

Tell me:

1. which layer failed (segmentation, normalization, index/search mismatch),
2. which exact WFGY page to open,
3. the minimal steps to push ΔS ≤ 0.45 and keep λ convergent,
4. a reproducible test (3 paraphrases × 2 seeds) to verify the fix.
   Use BBMC/BBCR/BBPF/BBAM when relevant.

Next planned page

rtl_bidi_directionality.md (Arabic/Hebrew mixing, mirroring, numerals)

🔗 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

Layer Page What it’s for
⭐ Proof WFGY Recognition Map External citations, integrations, and ecosystem proof
⚙️ Engine WFGY 1.0 Original PDF tension engine and early logic sketch (legacy reference)
⚙️ Engine WFGY 2.0 Production tension kernel for RAG and agent systems
⚙️ Engine WFGY 3.0 TXT based Singularity tension engine (131 S class set)
🗺️ Map Problem Map 1.0 Flagship 16 problem RAG failure taxonomy and fix map
🗺️ Map Problem Map 2.0 Global Debug Card for RAG and agent pipeline diagnosis
🗺️ Map Problem Map 3.0 Global AI troubleshooting atlas and failure pattern map
🧰 App TXT OS .txt semantic OS with fast bootstrap
🧰 App Blah Blah Blah Abstract and paradox Q&A built on TXT OS
🧰 App Blur Blur Blur Text to image generation with semantic control
🏡 Onboarding Starter Village Guided entry point for new users

If this repository helped, starring it improves discovery so more builders can find the docs and tools.
GitHub Repo stars