arxiv-translator

name: arxiv-translator description: > Download an arXiv paper's LaTeX source by its ID, translate ALL English text content to Traditional Chinese (繁體中文), and compile the translated source into a PDF using xelatex. Use this skill whenever the user mentions translating an arXiv paper, converting a paper to Chinese, downloading and translating LaTeX source from arXiv, or gives an arXiv ID and asks for a Chinese/Traditional Chinese version. Also trigger when the user says things like "把這篇論文翻譯成中文", "幫我翻譯 arXiv 論文", "translate paper 2401.12345", or any combination of arXiv + translate/翻譯. Even if the user just provides an arXiv ID and mentions translation or Chinese, use this skill.

arXiv Paper Translator (English → 繁體中文)

This skill downloads an arXiv paper's LaTeX source, translates it to Traditional Chinese, and compiles a translated PDF.

High-Level Workflow

1. Download & Extract — Fetch source from arXiv, detect format, extract
2. Analyze Structure — Find all .tex files, identify the main file, understand \input/\include dependencies
3. Translate — Translate each .tex file according to the translation rules below
4. Post-Translation Review — Check for missed sections or files
5. Configure CJK Support — Inject xelatex + CJK font packages into the preamble
6. Compile — Run xelatex (with bibtex/biber if needed) and fix errors if any
7. Deliver — Return the compiled PDF to the user

Step 1: Download & Extract

Run the helper script to download and extract the source:

python3 /path/to/skill/scripts/download_source.py <arxiv_id> <working_dir>

The script handles:

• Downloading from https://arxiv.org/e-print/<arxiv_id>
• Detecting whether the response is a tar/gz archive or a single .tex file
• Extracting to <working_dir>/<arxiv_id>/
• Reporting the list of extracted files

If the script is not available, do it manually:

ARXIV_ID="2401.12345"  # replace with actual ID
WORK_DIR="/home/claude/arxiv-work"
mkdir -p "$WORK_DIR/$ARXIV_ID"
cd "$WORK_DIR/$ARXIV_ID"

# Download source
curl -L -o source_archive "https://arxiv.org/e-print/$ARXIV_ID"

# Detect format and extract
FILE_TYPE=$(file source_archive)
if echo "$FILE_TYPE" | grep -qi "gzip\|tar"; then
    tar xzf source_archive 2>/dev/null || gunzip -c source_archive > main.tex
elif echo "$FILE_TYPE" | grep -qi "tex\|ascii\|utf-8\|text"; then
    mv source_archive main.tex
fi
rm -f source_archive

After extraction, list all files to understand the structure:

find . -type f | head -50

Step 2: Analyze Structure

Before translating, understand the project structure:

1. Find all .tex files: find . -name "*.tex"
2. Identify the main file: Look for the file containing \documentclass. This is your entry point.
3. Map dependencies: Check for \input{...}, \include{...}, \bibliography{...} to understand which files are included where.
4. Check for .bib files: Note if bibtex/biber is needed.
5. Check for custom .sty or .cls files: These generally should NOT be translated, but note them.

Important: Do NOT translate:

• .sty, .cls, .bst files (style/class files)
• .bib files (bibliography databases)
• Image files
• Any non-tex source files

Step 3: Translate Each .tex File

Create the output directory:

mkdir -p paper_zh-tw

For each .tex file, read the content, translate it, and save to paper_zh-tw/ preserving the relative path structure. Also copy ALL non-tex files (images, .bib, .sty, .cls, .bst, etc.) to paper_zh-tw/ so the project remains compilable.

Translation Rules

These rules are critical — follow them precisely:

TRANSLATE (English → 繁體中文):

• Running text / prose: All paragraph text, section titles, figure/table captions, footnotes, abstract, acknowledgments
• Section/chapter headings: \section{Introduction} → \section{引言}
• Captions: \caption{Overview of the model} → \caption{模型概覽}
• Comments in algorithms/code: Only translate descriptive comments (lines starting with % or // inside algorithm blocks). Keep variable names and code logic untouched.
• Theorem/lemma/definition names and content: Translate the textual content inside theorem environments

DO NOT TRANSLATE:

• Mathematical formulas: Everything inside $...$, $$...$$, \[...\], \(...\), and math environments (equation, align, gather, multline, etc.) — leave completely untouched
• LaTeX commands and their structure: \begin, \end, \label, \ref, \cite, etc.
• Person names: Keep all author names, cited researcher names in English (e.g., "Vaswani et al." stays as is)
• Code blocks: Inside lstlisting, verbatim, minted environments — keep code as is, only translate comments
• Algorithm pseudocode keywords: if, then, else, for, while, return — keep in English. Only translate descriptive comments.
• Bibliography content: .bib files and \bibitem entries stay in English
• URLs, DOIs, file paths
• Variable names, function names in code
• Custom LaTeX commands/macros defined in preamble: Do not modify \newcommand, \def, etc.
• Package names in \usepackage{}

TECHNICAL TERMS HANDLING:

Keep all professional and technical terms in English — do NOT translate them to Chinese:

• "attention mechanism" → 保持 "attention mechanism" 不翻譯
• "transformer", "encoder", "decoder", "embedding", "fine-tuning", "gradient descent" etc. → 保持英文原文
• Abbreviations like CNN, RNN, GPU, API, LLM, MoE, KV cache etc. → 保持英文
• Domain-specific terminology (e.g., "throughput", "latency", "inference", "prefill", "token") → 保持英文

TRANSLATION STYLE:

• Use formal academic Traditional Chinese (台灣學術用語)
• Use 「」for quotation marks (Traditional Chinese convention)
• Maintain the same paragraph structure
• Keep ~, \,, spacing commands as they are
• If a sentence mixes English names with translatable text, translate only the non-name parts

Parallel Translation with Subagents

When the paper has multiple .tex files (e.g., separate files for introduction, method, evaluation, appendix, etc.), launch multiple subagents to translate them in parallel for efficiency:

1. Identify independent .tex files: After analyzing the project structure in Step 2, determine which .tex files can be translated independently (most content files are independent of each other).
2. Prepare shared context: Before launching subagents, compile the following shared context that every subagent needs:
   • The full Translation Rules section above (TRANSLATE / DO NOT TRANSLATE / TECHNICAL TERMS / STYLE)
   • The output directory path (paper_zh-tw/)
   • Any project-specific custom commands or macros from the preamble that affect text rendering
3. Launch subagents in parallel: Use runSubagent to create one subagent per .tex file (or per group of small files). Each subagent's prompt must include:
   • The complete translation rules (copy the rules into the prompt — subagents are stateless)
   • The specific .tex file path(s) to translate
   • The output path to write the translated file to
   • Instruction to preserve LaTeX structure exactly and only translate content
4. Parallelism guidelines:
   • If there are 2+ independent .tex files, use parallel subagents (one per file or per logical group)
   • If there is only 1 .tex file, translate it directly without subagents (or split by section if the file is very large — see below)
   • Launch all subagents simultaneously — do NOT wait for one to finish before starting the next
5. Large single-file papers: If the paper is a single large .tex file (>500 lines), you may split it into logical sections (e.g., by \section) and assign each section to a separate subagent. In this case:
   • Read the file and identify section boundaries
   • Assign each section range to a subagent, providing the line range and surrounding context
   • After all subagents return, reassemble the sections into the final translated file in the correct order
6. Post-merge verification: After all subagents complete, verify that:
   • All files were written successfully
   • No duplicate or missing content at section boundaries (for split single-file translations)
   • Technical term handling is consistent across files (terms should remain in English everywhere)

Translation Process Per File

For each .tex file (whether translated directly or by a subagent):

1. Read the entire file content
2. Translate following the rules above — work section by section for long files
3. Write the translated content to paper_zh-tw/<same_relative_path>
4. After writing, briefly verify the output makes sense

When translating, think of the LaTeX source as having two layers:

• Structural layer (LaTeX commands) — preserve exactly
• Content layer (human-readable text) — translate according to rules

Work carefully. A missed } or corrupted command will break compilation.

Step 4: Post-Translation Review

After translating all files, do a thorough review:

1. File completeness check:

   # Compare file lists
   echo "=== Original .tex files ==="
   find . -name "*.tex" -not -path "./paper_zh-tw/*" | sort
   echo "=== Translated .tex files ==="
   find paper_zh-tw -name "*.tex" | sort
   

2. Non-tex file copy check: Make sure all images, .bib, .sty, .cls, .bst files are copied:

   # Copy all non-tex resources to paper_zh-tw if not already done
   rsync -av --exclude="*.tex" --exclude="paper_zh-tw" --exclude="source_archive" ./ paper_zh-tw/
   

3. Content spot-check: Open 2-3 translated files and verify:

   • No untranslated English paragraphs remain (section content, abstract, etc.)
   • Math formulas are intact
   • LaTeX structure is preserved
   • Person names are kept in English

4. Section-level audit: Check the main .tex file's \section / \chapter commands are all translated. Grep for common untranslated patterns:

   # Look for potentially untranslated section headings
   grep -n "\\\\section{[A-Z]" paper_zh-tw/*.tex
   grep -n "\\\\subsection{[A-Z]" paper_zh-tw/*.tex
   

If you find missed content, go back and translate it before proceeding.

Step 5: Configure CJK Support

The translated .tex files need CJK font support to compile with xelatex. Modify the main .tex file in paper_zh-tw/:

Add to preamble (after \documentclass but before \begin{document}):

\usepackage{fontspec}
\usepackage{xeCJK}
\setCJKmainfont{Noto Sans CJK TC}
\setCJKsansfont{Noto Sans CJK TC}
\setCJKmonofont{Noto Sans Mono CJK TC}

\usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue]{hyperref}

Hyperlink support (hyperref):

arXiv's compilation servers automatically inject hyperref into papers at build time, which is why the original arXiv PDF has clickable cross-references (\ref, \cite, \url). The .tex source itself often does NOT explicitly load hyperref. Always add it manually to the translated preamble so the output PDF has the same navigation links.

• Place \usepackage{hyperref} after fontspec and xeCJK to avoid conflicts.
• Use colorlinks=true to style links the same way as the arXiv version (blue, no boxes).
• If the original source already loads hyperref explicitly, do not duplicate it — verify that the existing options are compatible with xelatex (no pdftex driver option).

Important considerations:

• If the file already uses fontspec or xeCJK, don't duplicate — just ensure the CJK font lines are present
• If the document uses \usepackage[T1]{fontenc} or \usepackage[utf8]{inputenc}, remove or comment these out — they conflict with xelatex + fontspec
• If the document uses pdflatex-specific packages, you may need to adapt. Common fixes:
  • Remove \usepackage[pdftex]{graphicx} → replace with \usepackage{graphicx}
  • Remove explicit pdftex driver options from other packages

Install the font if needed:

# Check if Noto Sans CJK TC is available
fc-list | grep -i "Noto Sans CJK TC"

# If not found, install it
sudo apt-get update && sudo apt-get install -y fonts-noto-cjk
fc-cache -fv

Step 6: Compile with xelatex

cd paper_zh-tw

# First pass
xelatex -interaction=nonstopmode -halt-on-error main.tex

# If there are citations, run bibtex/biber
if grep -q "\\\\bibliography\|\\\\addbibresource" main.tex; then
    # Check if using biblatex (biber) or traditional bibtex
    if grep -q "\\\\usepackage.*biblatex" main.tex; then
        biber main
    else
        bibtex main
    fi
    xelatex -interaction=nonstopmode main.tex
fi

# Final pass for cross-references
xelatex -interaction=nonstopmode main.tex

Replace main.tex with the actual main file name identified in Step 2.

Auto-fix common compilation errors:

If compilation fails, examine the .log file and try these fixes:

After each fix, re-run the compilation. Try up to 3 rounds of auto-fix. If it still fails after 3 attempts, report the remaining errors to the user.

Step 7: Deliver the Result

Once compilation succeeds:

1. Note the path to the generated PDF: paper_zh-tw/main.pdf (or whatever the main file name is)
2. Provide a brief summary: paper title (translated), number of pages, any issues encountered

Edge Cases & Tips

• Very large papers (>30 pages): Work through files methodically. Don't try to translate an entire large file in one pass — break it into sections.
• Multi-file projects: Respect the directory structure. If the original has sections/intro.tex, create paper_zh-tw/sections/intro.tex.
• Papers using unusual templates (e.g., LNCS, IEEE, ACM): These often have strict .cls files. Don't modify the .cls — just add CJK support in the main .tex preamble.
• Already-translated content: If parts of the paper are already in Chinese, leave them as is.
• Encoding: Ensure all output files are saved as UTF-8.
• arXiv ID formats: Support both old format (0704.0001) and new format (2401.12345). Also handle IDs with version numbers like 2401.12345v2.