acl-writing-style

ACL Writing Style

Use this on the manuscript itself. ACL reviewers are NLP specialists who read

for whether the paper understands language as well as models; the style

that survives them is concrete, example-grounded, and precisely scoped.

First-page contract

• Open with the task or linguistic phenomenon, not the model family: what

goes in, what comes out, why it is hard, and for whom.

• State the contribution as a typed claim by paragraph two: new method, new

resource, new analysis, or new finding — ACL reviews are calibrated per type.

• Give one real example (input, desired output, failure of the status quo)

on page one; abstract problem statements without an example read as vague

at this venue.

• Say what languages the paper covers in the abstract if the answer is not

"English only" — and if it is, say that too.

Claim scoping in the LLM era

| Reflex phrasing | ACL-safe phrasing |

|---|---|

| "LLMs cannot do X" | "The five models tested fail X under these prompts" |

| "Our method understands Y" | "Improves the Y benchmark by n points; error classes A, B shrink" |

| "Works across languages" | "Evaluated on de/hi/sw/zh/ar; typological coverage discussed in §7" |

| "Significantly better" | Reserve for tested significance; give the test and p-value or interval |

| "State-of-the-art" | Scope to the exact setting, model scale, and date checked |

Reviewers increasingly ask whether a result is a property of the task, the

model snapshot, or the prompt; write so each claim names which.

Examples and error analysis as prose

• Every qualitative example must be attached to a number: how often the

illustrated behavior occurs, in which slice, under which condition.

Cherry-picked generations presented as evidence is a named reject pattern.

• Use interlinear glosses or transliteration conventions correctly for

non-English examples; sloppy linguistics costs credibility with exactly the

reviewers who like the paper's topic.

• Name error categories functionally ("negation-scope errors") rather than

narratively ("the model gets confused").

Anonymity-compatible voice

• Write self-reference in third person: "Smith (2024) introduced X," never

"In our previous work." Keep it in place until camera-ready.

• Do not cite "anonymous (under review)" material that reviewers cannot read;

ARR bars relying on documents unavailable to them.

• Acknowledgements, funding, and AI-assistance credits are omitted at

submission and added at camera-ready.

Compression into 8 (or 4) pages

• The short-paper form is a single sharp point with one strong experiment —

do not shrink a long paper into four pages; re-argue it.

• Push prompt dumps, per-language tables, and hyperparameter grids to the

appendix; keep one summary row of each in the body (see acl-supplementary).

• Kill the related-work-as-inventory section; two paragraphs of positioned

contrast beat a page of citations (see acl-related-work).

• Figures earn their space only when they carry an argument — pipeline

diagrams restating the text are the first cut.

Limitations and ethics prose

• Write Limitations as the referee brief against yourself: scope, data

coverage, model dependence, evaluation validity. Specificity here is

protected — ACL instructs reviewers not to penalize honest limitations.

• The optional ethics statement is for real stakes: human data, dual use,

representational harm. A boilerplate ethics paragraph is worse than none.

Micro-edit pass

weak:   "We leverage powerful LLMs to achieve impressive gains."
strong: "Reranking with a 7B model cuts negation-scope errors from
         31% to 12% of sampled failures (Table 4)."

weak:   "Performance is good across all settings."
strong: "Gains hold on 4 of 5 languages; Swahili degrades (-1.2 F1),
         which §7 traces to tokenizer fragmentation."

Terminology and notation discipline

• Pick one name per concept and hold it: a system called "our reranker,"

"the verifier," and "the LLM judge" in three sections reads as three

systems to a tired reviewer.

• Define task-specific terms at first use, even standard-seeming ones —

"hallucination," "faithfulness," and "robustness" each have three

incompatible literatures behind them.

• Dataset names get their citation at first mention and exact split names

thereafter ("XNLI dev-matched," not "the dev set").

• Numbers in prose match tables to the decimal; reviewers diff them.
• Language codes: introduce once (ISO 639), then use consistently in

tables, figures, and prose alike.

Section-level failure smells

• An introduction with no example → underspecified task (fix first).
• A method section narrating engineering chronology ("we first tried...")

→ rewrite as design with rationale.

• A results section that re-reads the table aloud → replace with claims the

table supports plus pointers into it.

• A conclusion introducing new claims → move them into results or delete;

ACL reviewers treat conclusions as summaries under oath.

Output format

[Style diagnosis] task-first / model-first / survey-ish / underspecified
[First-page fix] <one concrete rewrite>
[Overclaim list] <claim -> scoped version>
[Example-evidence gaps] <anecdotes lacking counts>
[Compression plan] <cut / move / merge>