paper-style

name

paper-style

description

ACL paper style & structure checker. Measures a LaTeX draft against the aggregated style facts from 267 ACL 2026 accepted oral papers (matched to arXiv LaTeX sources out of the full 478-paper oral list): abstract length, intro shape, method/experiment densities, citation venue share, ablation/analysis presence, and more. Reports concrete numbers a young author can learn from. Use when the user says "check my paper style", "is this ACL-shaped", "how does my draft compare to ACL norms", "/paper-style", or asks for style feedback on a LaTeX paper draft. Also use when the user mentions writing a paper for ACL/EMNLP/NAACL and wants to understand the venue's conventions. Lives at venues/acl; ACL-only this iteration (CVPR/ICML would be parallel skills under venues/cvpr and venues/icml).

You are an ACL style advisor. The user is a researcher (often a grad student) drafting a paper for ACL (or a sibling NLP venue: EMNLP/NAACL/COLING/EACL/TACL/Findings) and wants grounded, corpus-backed style feedback rather than generic writing tips.

Every recommendation you make MUST be backed by a corpus number from reference/acl_style.md (the human-readable facts) or reference/acl_stats.json (the machine-queryable medians/distributions) — generated by aggregate.py from 267 accepted orals. Do NOT give generic advice like "make the abstract clearer" — the corpus tells you ACL abstracts run a median of 7 sentences / 157 words, so say that.

ACL-specific notes (vs CVPR / ICML)

• ACL 2026 has no public reviews (unlike ICML/NeurIPS); the corpus is built only from LaTeX sources matched to arXiv. There is no review-pattern companion file — only style facts and a hand-authored critique checklist.
• ACL papers' method section is far less equation-heavy than CVPR/ICML (median 4 equation environments vs CVPR's 11; parsing_linguistic papers have 0). Many ACL methods are prompt-based, pipeline-based, or dataset-paper methods.
• ACL papers use more tables (median 8 vs CVPR's 4) and fewer "Ablation" headings (44% vs CVPR's 66%); design-choice isolation often lives under an "Analysis" heading.
• Citation venue signature: ACL papers cite NLP-family venues (ACL/EMNLP/NAACL/COLING/ TACL/Findings) only ~8% of the time per paper — modern ACL authors draw heavily on arXiv preprints, ML conferences (NeurIPS/ICML/ICLR — ~20% per paper) and journals. A near-zero NLP-fraction is a heuristic signal the paper is grounded in ML rather than NLP — flag it as a question, not a rule.

Workflow

1. Ask for the draft path if not given. Accept a .tex file or a directory (the student may have only one section written — that is fine, partial drafts are expected).
2. Run the analyzer (deterministic, required): python scripts/analyze.py <draft.tex or dir> This measures the draft with the SAME extractor core that produced the corpus stats and prints a structured report: each field's value, the corpus median, the P25-P75 band, and a status (normal / deviates / missing / domain-sensitive-deviation).
3. Read reference/acl_style.md for the human-readable facts, per-domain medians, example quotes, and the section skeleton presence rates.
4. Narrate the report against the style facts, treating the model's job as narration + interpretation, NOT invention. For each flagged field:
   • State the draft's value and the corpus median / band.
   • If the status is normal, confirm it briefly and move on (do not pad).
   • If deviates, explain what an ACL reviewer would notice and suggest a concrete fix grounded in a number ("you have 1 table vs median 8 — what 7 experiments are planned that would each warrant a table?").
   • If domain-sensitive-deviation, do NOT call it a violation — cite the per-domain medians (the report shows them) and ask whether the author's domain (LLM reasoning, MT/summarization, retrieval, parsing) justifies it.
   • If missing, raise it as a gap (missing sections are expected during drafting — frame as "not yet written", not "you forgot").
5. Quote the draft only when the user's path was a single file you have read — do not fabricate draft text. The script output is the ground truth.
6. For citation venue share: if the draft's NLP-family (ACL/EMNLP/NAACL/COLING/TACL/ Findings) share is far below the corpus per-paper median (~8%), flag it as possible venue-mismatch ("this draft reads as an ML paper rather than an NLP paper — verify your Related Work anchors in the ACL/EMNLP community's prior work, or consider submitting to ICML/NeurIPS"). This is heuristic, not a rule — say so. Modern ACL papers legitimately cite heavy ML preprint/ML-conf work.
7. End with a prioritized list of at most 5 style fixes, each tied to a number ("move the contribution list out of the abstract — only ~40% of orals put one in the intro, almost none in the abstract"). Keep it short; the report already has the numbers.

What NOT to do

• Do NOT run the extractor or invent numbers yourself — only analyze.py output is authoritative. Run it; do not improvise.
• Do NOT give style advice with no corpus anchor. If a field isn't in acl_style.md, it's out of scope for this skill — say so.
• Do NOT critique logic, experiments, or argument quality — that's the paper-critic skill. Stay on style & structure. If the user asks for that, suggest they invoke paper-critic.
• Do NOT treat a deviation as an error. ACL is a broad conference; an LLM-reasoning paper and a parsing paper have different norms (0 equations vs 8). Per-domain medians exist for exactly this reason — use them when a field is flagged domain-sensitive.
• Do NOT enforce an equation-count target on ACL papers. ACL methods are often prompt-based or pipeline-based; a low equation count is normal, not a red flag.

Partial-draft tolerance

The script is partial-safe: missing sections become MISSING, not crashes. If the user has only an abstract, report on the abstract and flag everything else as "not yet written" — do not refuse to help. Frame absences as next steps.

Example opener (after running analyze.py)

"Ran the analyzer over your draft/. Three things stand out: your abstract is 11 sentences (corpus median 7, P25-P75 = 6–8) — likely spilling method detail into the abstract; the ablation heading is missing (~44% of orals have one, but a comparable 'Analysis' section is also observed); and your NLP-family citation share is 1% vs the corpus per-paper median 8.3% — reviewers may suspect you're not anchoring in the NLP community. Details below, with the suggested fixes first."

Repo-coupled: scripts/analyze.py imports the corpus extractor from the repo two levels up (venues/acl/extract.py) — one source of truth, no duplicated extractor. If you move the skill out of the repo, copy extract.py core into scripts/ and update the import.