8. Writing clearly
We wrote this chapter for every author, and above all for the many *ACL authors who work in English as a second, third or fourth language. Start with the good news, backed by the field's own rules. Reviewers are told to judge your science over your polish: "as long as the writing is clear enough, better scientific content should be more valuable than better journalistic skills" (Chapter 27). You owe the reader clarity. You do not owe them native fluency. Clarity is a craft, and nearly all of it comes down to a handful of habits you can learn. Clear writing explains a hard idea at the least cost to the reader.
Write for the reader's effort, not your own
George Gopen and Judith Swan set out the core idea in their classic "The Science of Scientific Writing" (American Scientist, 1990). Readers expect information in fixed places in a sentence, 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 put the new point at the end, where emphasis falls. The reader arrives with context and leaves with your point.
- Keep the subject next to its verb. A reader holds a sentence open until the verb arrives. A long detour between subject and verb is the most common cause of "heavy" prose.
The reader keeps asking "why are you telling me this, here?" Good structure answers before they have to ask.
Tense: a small, learnable system
NLP prose uses three tenses, each for a fixed job. Choose by rule instead of by ear:
- 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, such as 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…".
Reviewers notice tense drift inside a paragraph. The three rules prevent it.
Articles: a, an, the
Many languages have nothing like the English article system, and articles cause the most common surface error in *ACL drafts. A wrong article rarely blocks meaning, so do not agonize. 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 on purpose, to match the strength of a claim to the strength of its evidence. Suggests, may, to our knowledge and under these conditions are precision tools. Authors misuse them in two ways:
- Over-hedging stacks hedges: "the results may possibly suggest that the method could perhaps help in some cases." Pick one hedge and say it once.
- Under-hedging overclaims: "our method solves named-entity recognition." Scope the claim to what you showed.
Make each claim 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.
Prefer strong verbs and the active voice
- Use the active voice. Write "We compute the score," not "the score is computed." The active 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."
Rewrite a page from nouns and passives into verbs and actives, and it typically loses a third of its words.
Cut the clutter
In On Writing Well, William Zinsser argues that "clutter is the disease" of prose. His cure is to strip each sentence to its cleanest parts:
- "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.
- Put one idea in each sentence. A sentence past roughly 25 words usually holds two ideas. Split it.
STYLE.md holds the full substitution table. Learn it and your second drafts will shrink.
Build paragraphs that flow
The paragraph is the unit of argument.
- Topic sentence first. Lead with the point, so a skimming reviewer gets it from the first line. 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 sentence from it. This is the stress-position rule applied to the paragraph. It makes prose flow where a list would only stack.
- Earn your transitions. However, therefore and in contrast each signal a specific logical move. Use the one that names the move you are making. Never add one for decoration.
A before-and-after
Revising your own clutter teaches more than any rule. 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. "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". The runaway sentence became two clean ones, and passive nouns became active verbs.
Habits that build the skill
- Read your draft aloud. Your ear catches the tangles your eye slides over.
- Imitate structure, never words. Read the papers in Chapter 30 and copy how they move from sentence to sentence. Never copy the sentences.
- Get one late reading from a fluent colleague after the science is settled, so the feedback is about prose.
- Use tools honestly. A grammar checker, or a model that polishes your own sentences, is the one clearly permitted AI use (Chapter 24). Never use one to generate claims or paragraphs you did not write.
- Revise. A first draft gets the ideas down. Clarity comes in the second and third pass, so expect to rewrite.
Common mistakes
- Translating your first language's sentence shape directly. The result is over-long sentences, late verbs and dropped articles. Write English sentences, not English words in another language's order.
- Hedge-stacks and overclaims. Say it once, 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 hardest writing skill.
- Polishing before the science is settled. You will rewrite that paragraph, so do not buff it yet.
Further reading
These books and essays repay the time.
- George Gopen and Judith Swan, "The Science of Scientific Writing" (American Scientist, 1990). The reader-expectation approach in a dozen pages, and the best 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. It is not academic, and is better for it.
- Steven Pinker, The Sense of Style (2014). A linguist's modern guide. It explains 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). Practical, written for computer science, and close to our conventions.
- The project's own STYLE.md, the house style this guide follows.