What Makes a Text Hard? Measuring Difficulty by Vocabulary, Not Grammar
Ask most people what makes a text hard, and they will describe long, tangled sentences — the kind you find in legal documents or Victorian novels. Grammar complexity feels like difficulty because it is visible: you can see the subordinate clauses piling up. But three decades of second-language acquisition research tell a different story. For a learner reading in a foreign language, the single most powerful predictor of whether they will understand a passage is not sentence length. It is the proportion of words they do not know. That number — unknown-word density, or its inverse, lexical coverage — governs comprehension more directly than any structural measure, and knowing how to interpret it is one of the most practical skills an adult language learner can develop.
The Grammar Intuition and Why It Misleads
Readability formulas have been around since the 1940s. The Flesch Reading Ease and Flesch–Kincaid Grade Level tests are widely used metrics for English readability. Each test applies the same core measures — word length and sentence length — to a text, giving different critical weight to those features. These formulas were designed for native-speaker populations: schoolchildren learning to read, Navy recruits processing technical manuals. They work reasonably well in that context because, for fluent readers, long sentences are genuinely harder to parse. But they were never designed for people learning a second language, and they have a critical blind spot.
Readability formulas measure two surface features of text that correlate with difficulty across large samples: sentence length and word length. But "correlates with difficulty on average" is a much weaker claim than "measures how hard this specific passage is to understand." Consider the difference between a short, eight-word sentence like "The physician cauterized the subclavian laceration promptly" and a twenty-word sentence built entirely from the thousand most common words of a language. The formula scores the second as harder. The language learner experiences it as far easier. The formula misses vocabulary difficulty beyond length: "bane" and "cane" are equally easy by the formula, yet one is far less common.
For L2 readers, what matters is not the architecture of a sentence but whether the bricks are familiar. And that is precisely what lexical coverage measures.
Lexical Coverage: The Number That Actually Predicts Comprehension
Lexical coverage answers a simple question: what percentage of the running words in this text are known to this reader? If a 500-word passage contains 25 unknown word tokens, coverage is about 95%. If only 10 tokens are unknown, coverage is about 98%. Those two numbers — 95% and 98% — are not arbitrary. They appear repeatedly in the research literature as thresholds around which comprehension changes in meaningful ways.
The landmark study was conducted by Marcella Hu and Paul Nation. Hu and Nation's (2000) study used a narrative text of 633 words with four different percentage coverages of known words to assess comprehension: 80%, 90%, 95%, and 100%. Their aims were to investigate whether and how different unknown-word densities affect comprehension and whether there was a threshold coverage level required for adequate comprehension.
The results were stark. Where 80% of the running words in the text were familiar to the readers, none gained adequate comprehension on either measure. Where 90% and 95% of the tokens were familiar to the readers, some gained adequate comprehension but most did not. Unknown-word density explained substantial variance — 62.2% — in comprehension scores. The results suggest a mostly linear relationship between unknown-word density and reading comprehension: participants' total scores on the reading tasks increased as the density of unknown vocabulary decreased.
To put the thresholds in concrete terms: at 95% lexical coverage, roughly 1 word in every 20 running words is unknown. At 98% lexical coverage, roughly 1 word in every 50 is unknown. One unknown word in twenty does not sound catastrophic until you are reading a dense paragraph and stumble four times in a single page. The cognitive load of repeated unknown-word encounters disrupts the flow of meaning-building, and comprehension collapses.
The Two Thresholds: 95% and 98%
The literature converges on two practical benchmarks. Researchers have suggested that 95% coverage might be needed for minimal comprehension, while 98% might be necessary for optimal comprehension. Laufer and Ravenhorst-Kalovski (2010) suggested that 95% coverage may be sufficient for reading comprehension with support, and that at 98% coverage, support may not be necessary.
The 2023 replication of Hu and Nation by Kremmel and colleagues adds important nuance. That pre-registered replication poses some important caution: the proportion of vocabulary you should know to get adequate comprehension is not necessarily 98%. The honest answer is that the threshold varies with text type, reader purpose, background knowledge, and what we mean by "adequate." For reading purely for pleasure without a dictionary, 98% is a reasonable target. For reading with a glossary or in a structured learning environment, 95% can sustain useful comprehension. What is not in dispute is the direction of the effect: research on second-language reading has repeatedly found that comprehension tends to improve as lexical coverage rises.
How Coverage Translates to Vocabulary Size
The thresholds become more concrete when you map them onto vocabulary size. Nation (2006) conducted a corpus analysis and found that the first 1,000 most frequent word families in the British National Corpus provide coverage of 78% to 81% of a written text, the second 1,000 add another 8% to 9%. The gains diminish steeply after that, which is why learning common words first gives such a large return.
| Word families known | Approximate written-text coverage | What that means in practice |
|---|---|---|
| 1,000 | ~78–81% | 1 unknown word in ~5; heavy guesswork required |
| 2,000 | ~87–88% | 1 unknown word in ~8; reading is effortful |
| 3,000–4,000 | ~92–95% | 1 unknown word in 13–20; minimally supported comprehension |
| 5,000–6,000 | ~95–96% | Approaching independent reading with effort |
| 8,000–9,000 | ~98% | Comfortable, unassisted reading of authentic texts |
If 98% coverage of a text is needed for unassisted comprehension, then an 8,000 to 9,000 word-family vocabulary is needed for comprehension of written text. That is a daunting figure for most learners. The practical implication is that reading authentic, unmodified prose comfortably in a foreign language requires sustained, long-term vocabulary building — and that knowing which words to build matters enormously. High-frequency words unlock coverage quickly; rare words barely move the needle.
Measuring a Real Text: What 1,500 Words of Portuguese Gets You
Abstract percentages are easier to understand with a concrete example. Here is a short passage in Portuguese, pitched at everyday topics. Before reading further, notice that a learner who knows the 1,500 most common words of Portuguese already knows every token in it — coverage is 100% at this frequency band. That is exactly the kind of information Beeblio computes for every passage it generates.
Hoje é um dia bonito. Vou à cidade com o meu amigo e depois comemos em casa. À noite leio um livro.
The result — 100% of tokens known at the 1,500-word band — is what makes this passage genuinely readable rather than merely grammatically simple. The sentences are not short or stripped of structure; a few are fairly complex. What keeps difficulty low is that every word is already in a learner's active recognition vocabulary. Add a handful of words from the 1,600–1,700 band and coverage drops just enough to introduce new items without overwhelming comprehension. That is the zone where incidental vocabulary learning is most likely to occur: enough known context to infer meaning, enough novelty to push the lexicon forward.
Why Grammar Complexity Is a Secondary Signal
None of this means grammar is irrelevant. A sentence with five embedded relative clauses in an unfamiliar syntactic pattern will slow any reader. But in practice, for adult learners with even modest exposure to a language, structural patterns become automatic long before vocabulary is mastered. You learn a handful of core constructions in the first few hundred hours; vocabulary learning never really stops.
The more fundamental point is that grammar complexity and vocabulary difficulty interact asymmetrically. A syntactically simple sentence built from unknown words is incomprehensible. A syntactically complex sentence built from known words is merely slow. Adequate vocabulary provides a basis for comprehension of language input; Nation found that the vocabulary required for comprehending an ordinary text should cover at least 95–98% of the vocabulary of the text, or else learners' comprehension will be interrupted frequently due to the interference of unfamiliar words. The interruptions are the problem. Grammar is the road; vocabulary is whether you can read the signs.
The Extensive Reading Connection
The practical upshot of the coverage research is a strong argument for reading at your level rather than above it. Extensive reading is a way of language learning through substantial amounts of reading, chosen and mastered by the learners themselves. Long-term language development requires exposure to understandable input, autonomy, and enjoyment, all of which are fostered by extensive reading.
During extensive reading sessions, learners read at their own pace, individually and silently, and the focus is on reading for enjoyment. Day and Bamford's first principle is that the reading material is easy — attuned to learners' reading proficiency. This is not a concession to laziness. It is an acknowledgment that comprehension is the precondition for acquisition. When coverage falls below 95%, the reader is spending cognitive resources on decoding rather than meaning-building, and the conditions for incidental word learning deteriorate. Nation (2006) reports that 98% coverage is most appropriate for incidental vocabulary learning.
The practical challenge is that most learners cannot reliably estimate their own coverage on a given text. They feel frustrated without being able to diagnose why. A short passage from a language-learning app can feel impenetrable not because the grammar is advanced but because six or eight words per hundred are unknown — enough to shatter the coherence of every other sentence.
Frequency Bands as a Progress Measure
Frequency bands are categories of vocabulary organized by how frequently words occur in large corpora of natural language. The logic of organizing learning around frequency bands follows directly from the coverage data. Frequency band membership determines the learning priority of a word: high-frequency words provide the greatest return on learning investment because they appear everywhere, while low-frequency words are less essential and appear rarely.
Knowing that you can read at the 1,500-word band means that texts built primarily from the 1,500 most common words of your target language will sit comfortably above the 95% coverage threshold for you. Knowing that the 2,000-word band still leaves you around 87–88% coverage on general written text gives you a concrete reason to push further. The band number is not an abstract CEFR level or a teacher's impression — it is a direct statement about the density of unknown words you will encounter in real reading.
This is why tracking known words is a more honest progress metric than tracking hours studied or lessons completed. Hours can be spent inefficiently. Lessons can be completed without retention. Known words, as measured against a frequency list and validated through reading performance, tell you what you can actually do with the language on the page.
FAQ
Does this mean I should ignore grammar study entirely?
No. Grammar study accelerates the point at which structures become automatic, reducing the cognitive load they impose during reading. The argument here is about priority and measurement, not exclusion. For most adult learners beyond the absolute beginner stage, vocabulary growth has a larger marginal return on comprehension than additional grammar study, and the coverage research is the evidence for that claim. Grammar matters most early on and in production; for reading comprehension, vocabulary density is the dominant variable.
Is 98% really the magic number, or is it more complicated?
It is more complicated. The 2023 replication of Hu and Nation cautions that the proportion of vocabulary you should know to get adequate comprehension is not necessarily 98%. The threshold varies by text type, reader background knowledge, purpose, and what counts as "adequate" comprehension. A thriller read for entertainment may need 98%+; a specialised text in a field you know well may be manageable at 95%. Think of 95% as the floor for useful reading and 98% as the target for comfortable, effortful-free comprehension.
How do I know my lexical coverage on a text before I read it?
Without a tool, you cannot — at least not precisely. A rough heuristic is to scan the first page and count how many words you do not recognise. Five unknowns in a 100-word stretch is 95% coverage; ten is 90%. Apps that measure your vocabulary band and generate level-matched texts bypass this problem by computing coverage before you read. The measured example in this article is precisely that: a passage where the coverage has been calculated against a frequency list, not estimated by feel.
What about cognates? Portuguese speakers learning Spanish, for example, recognise far more words than they have explicitly studied.
Cognate overlap inflates effective coverage significantly for closely related languages, and any honest frequency-band system should account for it. Transparent cognates function like known words for comprehension purposes, even if the learner has never formally studied them. The practical implication is that learners of closely related languages can often read at a higher coverage level than their studied vocabulary alone would predict. Beeblio calibrates bands empirically through reading performance, which captures this effect in the data rather than trying to model it theoretically.
Does listening have the same coverage thresholds as reading?
The thresholds are similar but slightly lower. Nation (2006) estimates an 8,000 to 9,000 word-family vocabulary for comprehension of written text and a 6,000 to 7,000 word-family vocabulary for spoken text, because speech draws on a narrower, higher-frequency range of vocabulary. Some researchers have found 95% sufficient for listening comprehension in contexts where the spoken register is informal and predictable. The key difference is that listening is transitory — you cannot stop and re-read — which may actually raise the effective threshold needed for real-time comprehension.
Sources
- Hu, M., & Nation, I.S.P. (2000). Unknown vocabulary density and reading comprehension. Reading in a Foreign Language, 13(1), 403–430.
- Kremmel, B., Indrarathne, B., Kormos, J., & Suzuki, S. (2023). Unknown vocabulary density and reading comprehension: Replicating Hu and Nation (2000). Language Learning, 73, 1127–1163.
- Nation, I.S.P. (2006). How large a vocabulary is needed for reading and listening? Canadian Modern Language Review, 63(1), 59–82.
- Schmitt, N., Jiang, X., & Grabe, W. (2011). The percentage of words known in a text and reading comprehension. Modern Language Journal, 95(1), 26–43.
- Laufer, B., & Ravenhorst-Kalovski, G.C. (2010). Lexical threshold revisited: Lexical text coverage, learners' vocabulary size and reading comprehension. Reading in a Foreign Language, 22(1), 15–30.
- Webb, S., et al. (2024). How does lexical coverage affect the processing of L2 texts? Applied Linguistics.
- Day, R.R., & Bamford, J. (1998). Extensive reading in the second language classroom. Cambridge University Press.
- Flesch–Kincaid readability tests (Wikipedia overview with primary formula references).
Every passage in Beeblio is pinned to your current frequency band, so the coverage calculation happens before you read a single word. When you tap an unknown item, it goes directly into spaced-repetition review and nudges your band forward. As your known-word count grows, the texts Beeblio writes for you shift — still comfortable, still at or just above the coverage sweet spot, but always pulling you toward the next band. Text difficulty, measured honestly, becomes your most reliable guide to texts worth reading right now.