Install
openclaw skills install @vittta/translate-paper-zhTranslate full papers into searchable Chinese PDFs
openclaw skills install @vittta/translate-paper-zhProduce a faithful, complete translation with verifiable coverage. Prefer the original paper's two-column class and visual hierarchy. Never call an abstract, selected sections, text-only extraction, or a figure-label-incomplete version a full translation.
The bundled helpers need Python 3.10+ and the dependencies in requirements.txt. Use an existing compatible runtime, or create a task-local virtual environment and install with python -m pip install -r "$SKILL_DIR/requirements.txt". Commands below assume python resolves to that runtime. Composition also needs an embeddable Simplified Chinese font; the TeX route needs an appropriate Unicode-capable TeX toolchain. The helpers do not call a translation API.
Set SKILL_DIR to this skill's resolved directory and use a task-local working directory for artifacts. Run:
python "$SKILL_DIR/scripts/paper_inventory.py" source.pdf work/inventory \
--source-url 'AUTHORITATIVE_URL' --rights-basis 'VERIFIED_LICENSE_OR_AUTHORIZATION'
Inspect every rendered source page. Extracted blocks are candidates, not guaranteed reading order or complete content: two columns can merge, math can split, and raster/vector labels may be absent. Use OCR only where needed, then compare to the page. Add otherwise invisible labels as visual_units in the plan and complete source_reviews honestly. Never mark these complete solely because an extractor ran.
Use the manifest reference for fields, coverage records and audit commands. Keep the original inventory immutable; reconcile IDs and checksum deliberately if repairing an extraction. Maintain a shared terminology table with first-use English expansions for important technical terms. Preserve hedging, limitations and distinctions such as training/test-time adaptation, online versus standard variants, loss functions and update rules.
Choose the least destructive approach in layout and fidelity:
overlay_regions.py replaces planned text units and preserves vector/image artwork, but does not translate, discover labels, protect mathematical meaning automatically, or edit raster labels. Use it only for suitable regions. For mixed prose/math, dense tables or crowded charts, prefer re-typesetting the affected component and compare against the source.Translate in section-sized chunks, giving workers the source pages/text, glossary, equations context and precise coverage. Keep a model-authored translation path; do not make completion depend on an unverified external translation API. Never transmit private manuscripts to a third-party service without appropriate authorization. Review each chunk against the source before marking fidelity_checked.
Preserve figures as vectors where possible. Translate labels, legends and panel annotations as well as captions; keep every tick value, color mapping, point, line and panel. Retain full tables and footnotes. Do not treat original English within a figure as translated because its caption is Chinese.
If embedded CJK fonts make the PDF unnecessarily large, optionally run optimize_pdf_fonts.py input.pdf optimized.pdf before final review. It uses a font-subsetting fallback and saves only when every page’s extracted text and rendered pixels remain identical. Keep the original if it rejects the optimization.
audit_translation.py to render every output page, extract searchable Chinese, check font embedding, check coverage, detect remaining English and create review templates. Audit failure is expected before manual review.