emnlp-writing-style

EMNLP Writing Style

Use this during revision. EMNLP's house voice is that of a careful experimentalist: the

paper reports what was measured, in prose that never outruns the tables. The most common

style failure at this venue is not bad English — it is claim inflation, where the

abstract promises "understanding" and the evidence delivers accuracy on four datasets.

The scoping discipline

Every empirical claim carries an implicit quantifier, and EMNLP reviewers expand it:

| Draft sentence | What a reviewer hears | Scoped rewrite |

|---|---|---|

| "LLMs cannot track negation" | All LLMs, all negation, all languages | "The three 7-13B models we test fail on double negation in English and German (Table 4)" |

| "Our method improves summarization" | Everywhere, always | "…improves faithfulness on news summarization in two of three datasets; gains do not transfer to dialogue (§6.3)" |

| "This is a multilingual benchmark" | Typologically broad | "…covering five Indo-European languages; we discuss this limit in §Limitations" |

| "significantly better" | A statistical test exists | "+1.8 F1, paired bootstrap p<0.01 over 5 seeds" — or delete "significantly" |

The rewrites are not weaker papers; they are the same paper with its evidence and its

prose in agreement — which is precisely what the soundness score measures.

Examples must be load-bearing

NLP papers live on examples, and EMNLP reviewers distinguish two uses sharply.

Decorative: a cherry-picked output demonstrating the method at its best. Load-bearing:

an example representing a counted category — "the model hedges the modal verb, an

error type accounting for 34% of failures (Table 6, row 2)." Convert every decorative

example to a load-bearing one or cut it; an example without a frequency is an anecdote,

and the CFP-era reviewer culture treats anecdotes as filler.

First-page architecture

The opening should let a tired area chair reconstruct the paper: the phenomenon or task

failure (with a number), why existing evaluation misses it, what you did, what you

found, and where it breaks. Two structural rules follow. First, name the *language

behavior* before the model: "summarizers delete speaker commitments" precedes any

mention of architecture. Second, contribution bullets must be falsifiable — "we show X

holds in Y setting but not Z" survives review; "we propose a novel framework" is

unfalsifiable and reads as such.

Writing the Limitations section

The mandatory, uncounted Limitations section is the most misused free real estate in

the format. Ritual limitations ("results may not generalize") waste it. Content

limitations do reviewer-defusing work:

Weak:   "Our approach may not work for all languages."
Strong: "All six test languages are written in Latin or Hangul scripts with
        whitespace tokenization; our segmentation-dependent method has no
        evidence for Thai or Chinese. Annotator pool was US-based crowdworkers,
        so the politeness judgments encode one cultural frame (§5.4)."

The strong version names the boundary, the mechanism behind it, and where the paper

already discusses it. Reviewers routinely check whether the Limitations section admits

the weakness they independently found — matching it is credibility; missing it is a

soundness deduction.

Hedging calibration

Empirical prose needs exactly one hedging level per claim, chosen by evidence:

• Demonstrated on your data: indicative mood, no hedge — "the effect reverses on

dialogue data."

• Supported but unproven mechanism: "consistent with", "suggests" — once, not stacked.
• Speculation: fence it explicitly ("we speculate…") or move it to future work.

Stacked hedges ("may potentially suggest") and false confidence ("proves") are equal

and opposite failures; both misreport evidence strength.

Titles and abstracts, venue idiom

EMNLP titles carry information, and the program is searched by task: name the

phenomenon or task in the title, not just the system ("Pragmatic Drift in Multilingual

Summarizers" beats "PoliteSumm: A Novel Framework for Better Summaries"). System names

are fine as prefixes when the system is the release; they are dead weight when the

finding is the contribution. In abstracts, the venue idiom is numbers early: the

best EMNLP abstracts contain the headline measurement, the languages or datasets, and

the significance posture — a reader should be able to referee the claim's scope from

the abstract alone.

Terminology honesty

Two lexical habits mark careful empirical writing here:

• "Understanding," "reasoning," "knows" — mentalistic verbs applied to models get

flagged unless operationalized; prefer the measured behavior ("answers correctly

under paraphrase") or define the construct once and stick to it.

• "State-of-the-art" — a comparative claim against a moving population; if used,

date it and bound it ("among open models under 13B evaluated on X as of May 2026"),

or replace it with the actual comparison made.

Two-column craft, ACL flavor

• Long papers get 8 content pages, short papers 4, references free, Limitations and

ethics uncounted — design the argument for the cap instead of compressing at the end.

• Wide comparison tables dominate NLP papers: prefer one decision-relevant table in the

body and push completeness grids to the appendix, cross-referenced.

• Non-Latin script examples need declared fonts and romanization plus gloss; an example

the reviewer cannot read is an example that does not exist.

• Short papers: one claim, stated in the first paragraph, with the strongest single

table — a short paper that "surveys" its own contribution has already failed.

The revision pass, ordered

Run the passes separately — combined passes miss things: (1) claim-scope pass over

every abstract and section-opening sentence; (2) example audit tying each excerpt to a

counted category; (3) hedge calibration; (4) Limitations content check against the

weaknesses you privately know; (5) format sweep for scripts, table widths, and the

uncounted sections. The first pass matters most and is best done by the coauthor who

did not run the experiments — attachment to results is what inflates their

description.

Output format

[Voice diagnosis] scoped / inflated / over-hedged / anecdotal
[Claim-scope repairs] <sentence -> scoped rewrite>
[Example audit] <decorative examples -> counted category or cut>
[Limitations quality] ritual / content — with the missing admissions
[Format risks] <tables, scripts, page-cap pressure>