6. Writing clearly
This chapter is for everyone, but it is written with the many *ACL authors who work in English as a second, third, or fourth language especially in mind. Start with the good news, which is backed by the field's own rules: reviewers are explicitly told to judge your science, not your prose polish — "as long as the writing is clear enough, better scientific content should be more valuable than better journalistic skills" (Chapter 24). What you owe the reader is clarity, not native fluency. And clarity is a craft, not a gift you are born with: nearly all of it reduces to a handful of habits you can learn. Clear writing is not writing dumbed down — it is writing that costs the reader the least effort to understand a hard idea.
Write for the reader's effort, not your own
The deepest principle in scientific writing comes from George Gopen and Judith Swan's classic "The Science of Scientific Writing" (American Scientist, 1990): readers have fixed expectations about where in a sentence information will appear, and meeting those expectations is most of what we call clarity. Two rules of thumb follow:
- Old information first, new information last. Open a sentence with something the reader already knows, and land the new point at the end — the position of natural emphasis. The reader arrives with context and leaves with your point.
- Keep the subject next to its verb. A reader holds a sentence open, slightly straining, until the verb arrives; long detours between subject and verb are the most common source of "heavy" prose.
The reader's silent question is always "why are you telling me this, here?" Good structure answers it before they have to ask.
Tense: a small, learnable system
NLP prose uses a few tenses consistently, and choosing by rule rather than by ear removes a common error:
- Present for established knowledge and for what your paper does: "Transformers use self-attention." "We propose…", "Section 4 reports…", "the model learns…".
- Past for specific completed actions — your experiments and what happened: "We trained the model on 10k sentences." "Accuracy rose by 3 points."
- Present perfect for prior work that bears on the present: "Prior work has shown that…".
Drifting between tenses inside a paragraph is a frequent tell; the system above keeps it consistent.
Articles: a, an, the
The English article system has no equivalent in many languages, so it produces the single most common surface error in *ACL drafts. It rarely blocks meaning, so do not agonize — but the rules of thumb are simple:
- a / an introduces something for the first time, or names one of many: "We train a transformer."
- the points to something specific, already mentioned, or unique: "The transformer we train has 12 layers." "the highest score."
- no article for general plurals and uncountable nouns: "Transformers are widely used." "Data is noisy."
Watch the uncountable nouns NLP leans on — research, work, evidence, training data, progress. They take no "a" and no plural: not "a research" or "many researches," but "research" or "a study."
Hedge, but calibrate it
Academic English hedges claims on purpose — to match the strength of a statement to the strength of its evidence. Suggests, may, to our knowledge, under these conditions are precision tools, not weakness. There are two ways to misuse them:
- Over-hedging — the hedge-stack: "the results may possibly suggest that the method could perhaps help in some cases." Pick one hedge and say it once.
- Under-hedging — the overclaim: "our method solves named-entity recognition." Scope it to what you showed.
Aim for a claim exactly as strong as your evidence, and no stronger. The project's STYLE.md bans the hedge-stack; the Results and Limitations chapters police the overclaim from the other side.
Prefer strong verbs and the active voice
- Active, not passive. "We compute the score," not "the score is computed." Active voice names the agent and is almost always shorter.
- Un-freeze your verbs. Nominalization buries a verb inside a noun: "we made a decision" → "we decided"; "we performed a comparison of A and B" → "we compared A and B."
- Drop the weak-verb padding. "provides an improvement to" → "improves"; "has the ability to" → "can."
A page rewritten from nominal-and-passive to verbal-and-active typically loses a third of its words and gains all of its energy.
Cut the clutter
William Zinsser's whole argument in On Writing Well is that "clutter is the disease" of prose, and the cure is to strip each sentence to its cleanest components. Concretely:
- "in order to" → "to"; "due to the fact that" → "because"; "a large number of" → "many"; "utilize" → "use".
- Delete empty intensifiers and throat-clearing: very, quite, really, clearly, it is important to note that.
- One idea per sentence. A sentence past roughly 25 words is usually two ideas wearing one coat — split it.
The full substitution table lives in STYLE.md; internalize it and your second drafts shrink on their own.
Build paragraphs that flow
Sentences are not enough; the paragraph is the unit of argument.
- Topic sentence first. Lead with the point so a skimming reviewer gets it from the first line and the rest of the paragraph supports it.
- One idea per paragraph. If a paragraph turns to a second idea, start a new one.
- Link old-to-new across sentences. End a sentence on an idea, then begin the next from it. This is the stress-position rule scaled up to the paragraph, and it is what makes prose feel like it flows rather than lists.
- Earn your transitions. However, therefore, in contrast each signals a specific logical move. Use the one that names what you are actually doing, and do not sprinkle them for decoration.
A before-and-after
The single most useful exercise is revising your own clutter. Here is a typical first-draft sentence and its repair.
Before. "Due to the fact that the utilization of large language models has been increasing across a large number of different tasks, it is very important to note that the evaluation of these models is something that is often performed in a way that is not particularly rigorous, and our paper is a paper that makes an attempt to provide an improvement to this situation through the proposing of a new benchmark."
After. "Large language models are now used across many tasks, but they are often evaluated unrigorously. We propose a new benchmark to fix this."
Sixty-one words became twenty-three, with nothing lost. What changed: "due to the fact that" → restructured; "utilization … has been increasing" → "are now used"; "it is very important to note that" → deleted; "a paper that makes an attempt to provide an improvement to" → "to fix"; one runaway sentence → two clean ones; passive and nominal → active and verbal.
Habits that build the skill
- Read your draft aloud. Your ear catches the tangles your eye slides over.
- Imitate structure, not words. Read the papers in Chapter 27 and steal how they move from sentence to sentence — never the sentences themselves.
- Get one late reading from a fluent colleague, after the science is settled, so the feedback is about prose and not content.
- Use tools honestly. A grammar checker, or a model used to polish your own sentences, is the one clearly-permitted AI use (Chapter 21) — never to generate claims or paragraphs you did not write.
- Revise. First drafts exist to get the ideas down; clarity arrives in the second and third pass. Expect to rewrite.
Common mistakes
- Translating your first language's sentence shape directly — over-long sentences, late verbs, dropped articles. Write English sentences, not English words in another language's order.
- Hedge-stacks and overclaims. Say it once, and exactly as strongly as the evidence allows.
- Nominalization and passive voice everywhere. Name the agent; use the verb.
- Mistaking "clear" for "simple-minded." Explaining a hard idea so a tired reviewer grasps it on the first read is the highest writing skill, not the lowest.
- Polishing before the science is settled. You will rewrite that paragraph; do not buff it first.
Further reading
This is a craft with a deep literature; these are the books and essays worth your time.
- George Gopen and Judith Swan, "The Science of Scientific Writing" (American Scientist, 1990). The reader-expectation approach, in a dozen pages — the highest-value short read here.
- Joshua Schimel, Writing Science (Oxford, 2012). How to structure a paper as a story, from the whole down to the sentence; written for scientists.
- William Zinsser, On Writing Well. The classic campaign against clutter; not academic, and better for it.
- Steven Pinker, The Sense of Style (2014). A linguist's modern guide, including why expertise makes us harder to understand — the "curse of knowledge."
- Strunk and White, The Elements of Style. Terse, old, and still right about most things.
- Joseph M. Williams, Style: Lessons in Clarity and Grace. The deepest treatment of cohesion and the old-to-new principle.
- Justin Zobel, Writing for Computer Science (Springer). CS-specific, practical, and close to our conventions.
- The project's own STYLE.md — the house style this guide is written in and enforces.