How Chinese Reading Difficulty Is Calculated

Reading difficulty in Chinese is harder to measure than in English. Average sentence length and vocabulary size still matter, but Chinese has features that change the math: no spaces between words, characters that vary in difficulty independently of the words they form, function words that carry grammatical load, and a writing system where one rare character can derail an otherwise-easy passage.

The standard English formulas -- Flesch-Kincaid, Dale-Chall, SMOG -- don't translate cleanly. They lean on syllable counts, spaces between words, and word-frequency lists that have no direct Chinese equivalent. A useful Chinese readability metric has to be built differently.

Flesch-Kincaid, the most familiar English readability score, uses two inputs: average sentence length and average syllables per word. Both rely on visible word boundaries -- spaces -- which Chinese doesn't have. Determining where one word ends and the next begins is itself a non-trivial NLP problem (called segmentation), and segmentation choices change the score before any difficulty calculation begins.

Character complexity is another input English doesn't have. The character 一 (yī, one) and the character 餐 (cān, meal) are both single tokens, but they share almost nothing in difficulty for a learner. Vocabulary frequency lists handle this for words, but not for the underlying script -- and in Chinese, the script is where most of the difficulty actually lives.

Chinese readability research splits into two camps. L1 (first-language) systems target native readers and predict things like grade level for school texts. L2 (second-language) systems target learners and predict how hard a text is given a known proficiency level, typically expressed in HSK bands.

The two camps measure different things. L1 readability assumes the reader has full vocabulary access and asks how cognitively demanding the syntax is -- clause embedding, semantic abstraction, rare-word density. L2 readability mostly asks: what fraction of this passage will the reader actually be able to decode?

A score from one camp doesn't translate to the other. An L1 tool can call a passage easy because its sentences are short and its vocabulary is high-frequency among natives -- but if a third of the characters are off the HSK list, an L2 learner is stuck. The reverse also happens: a passage with simple HSK 1 vocabulary but heavy clause embedding can read easily for an HSK 4 learner and still rate as L1-difficult.

For learner-facing readability, the inputs that carry the most signal are well-established in the L2 readability literature. Linear regression with these features hits roughly 85-95% of the predictive accuracy of small neural models, with the practical advantage of being inspectable -- you can debug a misclassified passage by looking at the inputs.

HanziFeed grades the news passages in its reading feed using a hybrid lexical-syntactic score, computed at pipeline time and stored alongside each passage. The full formula is published below in case other tools want to reproduce it.

For every unique CJK character in the passage, look up its HSK 3.0 (2026) level (1-9). Characters not in the HSK list -- usually rare or technical -- default to level 9. Take the weighted mean across the unique characters: each level is weighted by how many distinct characters at that level appear in the passage.

Add a small penalty for variance in the distribution. A passage with a flat HSK 2 vocabulary is genuinely easier than one that averages HSK 2 but spikes to HSK 7 in places -- the mean alone hides this. The penalty is computed as 0.3 × the standard deviation of the HSK distribution.

Run a longest-match-greedy scan against the HSK 3.0 multi-character vocabulary list. For every match, record the word's HSK level. If at least 5 unique words are found, take the mean of the word-level distribution and apply a bounded bonus.

This bonus only fires upward. It catches compound words that are harder than their characters suggest -- 新能源 (xīn néngyuán, new energy) is HSK 6, but the individual characters are HSK 1-3. Without a word-level pass, character-only scoring would underrate passages that are heavy on technical compounds.

Split the passage on Chinese sentence-ending punctuation (。 ！ ？) and compute the average CJK character count per sentence. For passages with average sentence length above 20 characters, add a small bonus capped at 0.15.

The cap is deliberate. Sentence length is a real signal but a noisy one -- news writing in particular can stack noun phrases without genuinely raising difficulty. Letting it dominate the score would over-penalize informational text that's actually quite readable.

The final score is the sum of the three components: final = charScore + wordBonus + sentenceBonus. The score is then bucketed into one of five tiers using fixed thresholds.

Passages must also pass a separate quality gate: at least 40% of unique characters must be at or below the learner's target HSK level. A passage with the right average difficulty but bad coverage -- lots of unknown characters concentrated in the unfamiliar 60% -- is rejected before it reaches a learner, even if its tier looks correct.

Small neural models with the same input features hit roughly 95% of the predictive accuracy of the linear formula above. The remaining 5% wasn't worth the cost of an opaque model that's hard to debug.

When a passage gets miscategorized, a linear formula tells you exactly why -- one input was higher or lower than expected. With ML, the same diagnostic takes feature-attribution work and rarely points to a clean fix. For a pipeline that grades a few hundred passages a day, transparency wins.

This matches the broader L2 readability literature. L1 research has moved toward deep models because the syntactic signal is genuinely complex; L2 research mostly stays with linear regression because the dominant signal -- vocabulary coverage -- is already well-modeled by linear features.

The big graded reader brands -- Mandarin Companion, Imagin8Press, DuChinese, The Chairman's Bao -- all publish leveled content but rarely explain how the levels are set. From what's publicly visible, most use a vocabulary list per level and rely on editorial judgment to write or select content that fits.

That approach has a real strength: human editors catch things a formula misses, like a familiar word used in an unusual sense, or a passage that's lexically simple but culturally opaque. The weakness is that the bucket boundaries are not formalized, which makes it hard to compare Level 2 across two publishers, or to know exactly what a reader needs to know to attempt a given level.

Algorithmic grading and editorial grading aren't really competing methods -- they answer slightly different questions. An algorithm answers "how hard is this passage given the inputs?" An editor answers "is this passage actually appropriate for this audience?" Production-grade reading apps tend to combine both: an algorithm for scale, editorial review for the cases the algorithm gets wrong.