scholarly-translation

name

scholarly-translation

description

Produce a scholarly, annotated, bilingual HTML "critical edition" translation of an academic text (journal article, book chapter, essay), chapter by chapter. Use this skill whenever the user wants to translate a research paper, academic article, scholarly book, or long intellectual text into another language — especially Chinese — and wants it as a polished reading document rather than a bare translation. Trigger this for any request like "translate this paper / article / chapter", "做成中英对照 / 带注释的译本", "翻译这篇论文并做成 HTML", or when the user uploads a scholarly PDF and asks for a translation, even if they don't explicitly say "HTML" or "skill". Handles unified terminology, key-term annotation with original-language glosses, translator's notes (校注) for proper nouns and historical context, faithful figure extraction from PDFs, native rebuilding of important tables, and reading navigation (table of contents, hover footnotes). Default target language is Chinese; adapt for others.

Scholarly Translation → Annotated Bilingual HTML

This skill turns a dense academic source into a self-contained HTML reading document in the style of an annotated critical edition: fluent target-language prose, original-language glosses on key terms, translator's notes in a side margin, faithfully extracted figures, natively rebuilt tables, and smooth navigation. Defaults assume English → Chinese; the structure generalizes to other language pairs by swapping fonts and the glossary.

Use assets/template.html as the starting scaffold — it already contains the full CSS/JS design system. Use references/figure-extraction.md whenever the source has figures in a PDF.

Core principles (agreed conventions)

1. Chapter by chapter, quality over speed. Translate one coherent section at a time. Append each new chapter to the same HTML file so the document grows into one deliverable. End each chapter with a short "续译提示" (continue) block summarizing what was done and what comes next.
2. Fluency over literalism. Never translate word-by-word. Recast verbs and word order so the result reads as natural target-language prose. (Canonical failure to avoid: rendering "Kent's book is by intention a comparative study…" as "肯特的著作有意识地是一项…比较研究" — wrong; write "肯特这部著作， 本意就是一项…比较研究".) Faithfulness is to meaning, not to surface syntax. Recurring calque traps, all observed in real edits — hunt them in a dedicated readability pass (see step 8):
   • Register/tone, not just denotation. Translate the attitude. "breathlessly mistaken for innovation" is excited, not panting → 兴冲冲 (not 气喘吁吁); "in the short/long run" is everyday → 短期／长期 (not the techy-sounding 短程／长程).
   • Recast calqued idioms & abstract-noun chains. "one place to look" → 不妨到… 去找 (not 是一处可去之地); "starting inspiration" → 起步的灵感; "vulnerability to tipping" → 在「翻转」面前的脆弱性 / verbalize it; "nonlinear rates of tipping" → 使翻转以非线性的速率发生; "an equilibrium of reproducing transformations" → unpack into a clause, don't stack 的-modifiers.
   • 量词 / collocations. Check measure words (梯子 takes 一架, not 一具).
3. Unified terminology, fixed up front. Before translating prose, build a key-terms glossary (§0) and reuse it across every chapter. Confirm contested renderings with the user; record the rationale for non-obvious choices in a 校注. The contract binds the prose, not just §0: once a recurring concept word is fixed (e.g. biography → 生平), never let it drift to a synonym (履历) later in the body — but keep deliberate authorial contrasts distinct (careers 职业生涯 vs biographies 生平). After translating, grep the body for each §0 term to confirm one rendering, one meaning.
4. Annotate, don't just translate. Key terms get the original gloss inline; proper nouns, institutions, and historical events get explanatory 校注.

Workflow

0. Read the source and scope it

• If a PDF: read it (rasterize/extract). Identify the chapter/section structure, the figures, the tables, and the footnote numbering. Confirm the chapter plan.
• Note: translating a document the user provided for their own study is a normal scholarly task. Quote/translate freely from it; embedding its own figures into the user's translation is fine.

1. Build the glossary (§0) first

A 3-column table — 英文术语 | 中文译法 | 简注. Include every recurring technical term, method name, and institution. This table is the contract for all later chapters; keep it stable. Surface 2–4 genuinely contested choices to the user for sign-off before proceeding.

2. Translate the chapter as fluent prose

• Wrap key terms at first occurrence: …<span class="t">稳健行动<span class="en">robust action</span></span>…
• Wrap names / minor original-language phrases (people, places, Latin/Italian institutional terms) in <span class="lat">Cosimo de' Medici</span>.
• Translate all footnotes; put them in a per-section <div class="footnotes"> block whose <li value="N"> numbers match the original exactly. The JS turns every <sup class="fnref">N</sup> into a hover-tooltip + jump link automatically.
• Use <blockquote> for block quotations (with <cite> for attribution).

3. Add translator's notes (校注)

Use <aside class="jiaozhu"> with a <span class="tag">校注 · 【topic】</span> header. Add them for: proper nouns and institutions (what was the Catasto? who were the magnates?), historical events, and the reasoning behind a terminology choice. Keep each note tight (2–5 sentences). On wide screens they float into the right margin; on narrow screens they inline.

Also add 大白话 "一句话" notes for genuinely abstract concepts — the ones a reader stalls on (autocatalysis, multifunctionality, co-evolution, "tipping" / poisedness, a coined two-stage process…). At the term's first definition, restate the idea in everyday words or a concrete mini-analogy (autocatalysis → "A 造 B，B 造 C，C 又回头造 A，整圈自我维持、自我修复"; co-evolution → "考官随考生一起变，'最优' 没了固定标尺"). When the source itself is dense, this unblocks comprehension far more than tweaking the prose does — readers reported these as the most helpful part. If the chapter relies on a recurring figure convention (e.g. Venn "b-figures"), add one note that teaches the reader how to read them.

4. Figures — extract faithfully, embed, translate captions

Follow references/figure-extraction.md. Key rules: render the PDF page at ~220 DPI; for landscape figures rotate the whole page first then crop; crop by the ink bounding box (not guessed pixels) so nothing is clipped; always view the full result to verify completeness before embedding; optimize to grayscale PNG and inline as base64. Put a Chinese <figcaption> that also glosses the in-image legend (line-style → tie-type). Place each figure right where the prose first discusses it.

5. Tables — rebuild important ones, summarize the rest

• Important tables (the empirical backbone, the ones the argument turns on): rebuild natively as <table class="dtable"> with a <caption> carrying the table number, right-aligned tabular-nums data cells, a bold total row, and a <p class="tbl-note"> underneath for medians / test statistics / source.
• Minor tables (e.g. tables that only report test statistics): fold their conclusions into a 校注 or a tbl-note rather than reproducing the matrix.
• For very large/complex tables, rebuild a focused excerpt of the most telling rows/columns and note in the tbl-note what the full original contains.

6. Navigation

• Give every <h2 class="ch"> / important <h3 class="sub"> an id, and add a matching <li><a href="#id"> to nav.toc. Update the TOC every chapter.
• Hover footnotes, footnote jump-and-highlight, and the back-to-top button are handled by the bundled JS — no per-chapter work needed beyond correct value numbers on the footnote <li>s.

7. Verify and deliver

Run structural checks (figures present and complete, all ids and TOC anchors match, <figure>/<aside>/<p>/<div> tags balanced, footnote values present, the fluency fixes landed). Two checks that catch silent content loss:

• Footnote ref ↔ def parity. Every defined <li value="N"> must have a matching inline <sup class="fnref">N</sup>, and vice versa. A def with no ref means you dropped the sentence/paragraph that anchored it — this is how an omitted paragraph surfaces (it happened: a whole modeling-chapters paragraph was caught only by its orphaned footnote 8).
• Term consistency grep. For each §0 term, grep the body to confirm a single rendering (and that deliberate near-synonyms stayed distinct).

Save to the outputs folder and present_files.

8. Readability pass (after the literal translation)

Do this as its own pass, with fresh eyes — stiffness hides while you're still matching the source. Dump every prose paragraph tag-stripped (regex the <p>s, drop .en glosses and <sup>s) and reread cold, hunting the principle-2 calque traps and any §0-term drift. The source's inherent density is not yours to fix by paraphrase; the translation's added friction is. Surface a candidate list (original → fix, grouped, marked must-fix vs. stylistic) to the user before applying — it's their voice. Then apply mechanically as (old,new) replacements with a per-string count==1 assertion so nothing misfires silently.

Editing the file mechanically (avoid huge tool payloads)

The file gets large (figures are base64). Prefer Python string/regex edits run in bash over pasting big strings:

• Append a chapter by replacing the current <div class="continue">…</div> block.
• Swap a figure by regex-replacing only the src (see figure-extraction.md).
• Apply fluency fixes as a list of (old, new) replacements and report misses. Read the file tail with grep -n/view to get exact anchor strings first.

The design system (in assets/template.html)

Fresco / quattrocento palette — adapt the accents to the subject, but keep the three semantic roles:

• --porphyry #7a2b25 — key-term English glosses (.t .en), term underlines.
• --gild #977028 — numerals, ornament, block-quote rule.
• --slate #3f5560 — translator's notes (.jiaozhu), table headers.
• --paper #ece8df ground, --ink #221d18 text, --rule #cdc6b7 hairlines.

Fonts: Noto Serif SC (CN body), EB Garamond (Latin display/italic), Noto Sans SC (UI/notes). For non-CN targets, swap the CJK serif for an appropriate body face.

Key classes (all defined in the template): .t+.en (key term + gloss), .lat (names/minor original), .jiaozhu (校注 margin aside), sup.fnref + .footnotes, figure.fig/figcaption, table.dtable + .tbl-note, table.gloss-table, nav.toc, #fn-tooltip, #to-top, article.reading (the ~44rem reading column). The signature element is the margin-note critical- edition layout plus a subject-appropriate SVG ornament in the masthead.

Quality bar / self-check

• Would a fluent native reader find any sentence stiff or "translated"? Fix it — watch the principle-2 calque traps (register, idiom, abstract-noun chains, 量词).
• Is every recurring key term rendered the same way as in §0 — in the prose, not just the glossary? (grep to confirm; keep deliberate contrasts distinct.)
• Does each genuinely abstract concept have a 大白话 "一句话" note, or is it left as a bare literal term the reader will stall on?
• Every defined footnote also referenced? (orphan def = a dropped paragraph.)
• Does each figure show its entire content (axes, all nodes, legend)?
• Do hover footnotes and TOC links actually resolve?
• Are sensitive choices (a debatable rendering, a factual correction, a simplified table) explained in a note rather than left silent?