brycewang-stanford NAACL Writing Style
The house register at NAACL is precise about language and modest about generality. The chapter's renaming to "Nations of the Americas" sharpened a tendency its reviewers already had: sentences that quietly equate "language" with "English" or "text" with "standardized text" get flagged, and claims are expected to wear their linguistic scope on the surface.
Scope discipline: say which languages, every time
The most common NAACL-bound prose failure is the unscoped universal:
| Draft sentence | What a reviewer writes | Repaired sentence |
|---|---|---|
| "Our method improves summarization." | "On which languages?" | "improves summarization on English and Brazilian Portuguese news" |
| "LLMs struggle with negation." | "All LLMs? All constructions?" | "the four evaluated models mishandle sentential negation in Spanish" |
| "This transfers to low-resource settings." | "You tested one." | "transfers to Guaraní; other low-resource claims remain untested" |
| "Speakers prefer our outputs." | "Which speakers, recruited how?" | "22 L1 Mexican-Spanish annotators preferred outputs 68% of the time" |
The repair is never hedging for its own sake — it is replacing a category ("low-resource languages") with the sample you measured.
Examples as argument, not decoration
- Choose examples your quantitative analysis actually turns on: the error class that drives the gap, the construction the model misparses.
- Gloss every non-English example in three lines — original, aligned gloss, free translation — so a reviewer who reads none of your languages can still verify the reasoning.
(3) axolotl o- k- ita in siwatl [Nahuatl]
axolotl PST-3SG.OBJ- see DET woman
'The woman saw the axolotl.'
-> the model tags 'axolotl' as agent: the error class in §5.2
- After each example, one sentence must state what it demonstrates; an example without a claim attached is a page cost with no return.
First page architecture
State, in order and within the first column and a half: the phenomenon or task; the gap in current handling; what you built or measured; the headline result with its scope; why it matters to this community. A NAACL reader should be able to quote your contribution accurately from the first page alone — including its language coverage — without meeting a model name before the problem.
Limitations as anticipatory prose
Write the Limitations section by simulating the three most likely reviewer objections and answering them in your own words first. Genuinely useful NAACL-flavored entries: variety coverage ("results are for Rioplatense Spanish; peninsular varieties untested"), annotator population limits, data-governance constraints on release, and evaluation-metric validity for morphologically rich languages. The section is uncounted; verbosity there is free, evasiveness is not.
Compression order for the content-page budget
- Cut roadmap paragraphs and per-section previews — structure should show, not narrate itself.
- Merge related-work prose into contrastive sentences attached to your contributions.
- Collapse repeated table commentary into one analytical paragraph.
- Move example galleries and full prompt text to the appendix, keeping the load-bearing instances.
- Only then consider shrinking figures — never below caption legibility.
Terminology and notation consistency
Multilingual papers accumulate inconsistency faster than monolingual ones because every concept has more surface forms. Fix a house convention in a notes file before drafting:
- One term per concept: pick "code-switching" or "code-mixing" and hold it; alternating reads as two half-edited drafts merged.
- One romanization/orthography per language throughout, matching what the data actually uses; if the corpus mixes orthographies, that fact is content and belongs in the data description, not in silent inconsistency.
- Variety codes stated once and reused (
es-MX,pt-BR,gn), never informal synonyms ("Mexican", "the Spanish data") drifting through sections. - Metric names with their implementation on first mention, then the bare name.
A ten-minute consistency sweep with grep -c over the LaTeX source for
each term variant catches most of it mechanically.
Register calibration
- Prefer measured verbs: indicates, is consistent with, fails on — over promotional ones: revolutionizes, unlocks, dramatically.
- Name languages with their standard names and ISO codes on first mention
(Guaraní,
gn); respect community-preferred names for varieties. - Keep diacritics correct in every language name and citation — misspelling Wixárika in a paper about it is a credibility wound at this venue.
Output format
[Scope audit] <unscoped claims found -> repaired versions>
[Example audit] <each example -> claim it carries; gloss ok?>
[First-page test] pass / rewrite needed (what's missing)
[Limitations preview] <predicted objections vs section coverage>
[Compression plan] <cuts in order, estimated page recovery>