Spoken vs written Chinese

You can hold a conversation in Chinese and still not get through a news article.

It is not only that the topics are harder. The news is built from a different, and much larger, set of words than speech is.

Dengdeng is our reading app for learners who want to read about China but find the original articles too difficult. Each story comes in Easy, Medium, and Hard, so you can change the difficulty without changing the subject.

The numbers below are from public corpora, not from our stories.

For speech we used SUBTLEX-CH, a frequency list from 33.5 million words of film and television subtitles — the closest large, public stand-in for how people talk. For writing we used 100,000 news sentences from the Leipzig Corpora Collection, segmented with jieba.

One is conversation. The other is news writing.

Two corpora, read with care

• Two imperfect stand-ins. Subtitles are not speech — they are written-down speech, cleaned up a little. News scraped from the web is not all of written Chinese — it leans on current events, politics, and the economy. Each is a good proxy for its register, not a perfect one.
• The corpora differ in size. SUBTLEX is far larger, so its long tail of rare words is better sampled. We compare coverage as a share of running text and word frequencies per million tokens, both of which are size-neutral, rather than raw counts.
• Segmentation is automatic. "Words" come from jieba; part of speech comes from jieba's tagger at the dictionary level. Both are approximate and aggregate. The news corpus mixed simplified and traditional sources, so we normalised everything to simplified first — otherwise the same word in two scripts would have looked like two different, register-specific words.
• Frequency is not difficulty. A word being common in the news does not make it easy, and recognising every word in a sentence is not the same as understanding the sentence.

The size of the gap

How many distinct words must you know to cover a given share of everything written in a register? Sort each corpus from most common word to least, and walk down the list.

To cover 80% of running speech you need about 844 distinct words. To cover the same 80% of the news you need roughly 6,692 — nearly eight times as many.

Push to a comfortable reading threshold of 95% and conversation asks for about 6,887 words. The news asks for 40,367. The gap does not close as you climb. It widens in absolute terms.

At 90% coverage the difference is already about 16,500 words. That is the extra vocabulary between holding a conversation and reading a paper without constant stumbling.

To read the news comfortably you need on the order of fifteen-to-thirty thousand more words than you do to follow a conversation. That is most of a second vocabulary.

Function words and nouns

Part of the answer is what kind of words each register is made of. Tag every token and weigh the parts of speech by how often they occur.

In speech, function words — pronouns, particles, adverbs, prepositions — make up about 44% of everything said. Pronouns alone are nearly 19% of all tokens. Conversation runs on a small, closed set of grammar: I, you, he, this, that, a handful of particles.

The news inverts it. Function words fall to about 26%. Pronouns collapse to 4%. Nouns double, from roughly 16% of speech to 32% of the news — one written word in three is a noun. Verbs hold roughly steady.

The written register spends its words on naming things. There is an endless supply of things to name.

The words you are missing

We ranked each corpus and pulled out the words that are common in the news but rare or absent in speech — the vocabulary a conversational learner has almost never seen.

The list reads like the table of contents of a newspaper: 疫情 (epidemic), 群众 (the masses), 干部 (cadre), 防控 (prevention and control), 制度 (system), 建设 (construction), 强调 (to emphasise), 落实 (to implement).

These are not exotic or literary. They are ordinary news words. They barely surface in subtitles, where 疫情 sits past spoken rank 20,000 and 干部 past 25,000. A learner who built their vocabulary from conversation never had a reason to pick them up.

This is the bulk of that second vocabulary. It clusters around governance, the economy, and public life. It is learnable as a set.

The reverse list is the texture of talk: greetings and acknowledgements (谢谢 thanks, 对不起 sorry, 再见 goodbye), particles and fillers (嗯, 嘛, 噢), kinship and endearment (妈 mum, 爸 dad, 宝贝 darling), and blunt feeling (讨厌 ugh, 他妈的 — the swear). A newspaper has no use for any of them.

The same idea, two words

The split is not only in which words appear. It is also in which of two near-synonyms a register prefers.

Chinese is full of pairs that mean almost the same thing but sit at different levels of formality. The news and the street pick opposite members of each pair.

For the future, the news reaches for the formal 将 (jiāng, "shall") often enough to edge out the everyday 会 (huì, "will"). In speech, 会 outnumbers 将 by more than ten to one.

The inclusive "we," 咱们 (zánmen), is a fixture of conversation and almost absent from reporting, which uses 我们. The formal "and," 及 (jí), is roughly thirty times more common in the news than in speech, where 和 does all the work.

The everyday intensifier 很 ("very") is more than six times as frequent in speech as in writing, where the stiffer 十分 picks up some of the slack.

Learn only the colloquial member of each pair — as you would from talking — and you will read the formal one as a small stumble every time it appears, several times a paragraph.

What to do

Conversation was not wasted. It is the foundation. The function-word core you built from talking still applies.

The difference between conversation and news is mostly a few thousand nouns, plus formal synonyms for grammar you already know.

The way to learn those words is to read the register that uses them.

Common questions

Why can I speak Chinese but not read the news?

The news uses a larger, more formal set of words than speech does. Comfortable news reading needs on the order of fifteen to thirty thousand more words than conversation.

How many words do I need to read Chinese news?

About 6,700 distinct words cover 80% of news text. Conversation needs about 840 for the same share.

Method

Spoken baseline: SUBTLEX-CH (Cai & Brysbaert 2010), a word-frequency list from ~33.5M subtitle tokens; we used its word and count columns directly. Written register: the Leipzig Corpora Collection file zho_news_2020_100K (99,993 news sentences), normalised traditional→simplified with OpenCC, then word-segmented and POS-tagged with jieba. Coverage curves are token-weighted (share of running text covered as word types are added most-common first). The register-gap tables compare per-million-token frequencies and within-corpus ranks; "absent in speech" means a word falls below SUBTLEX rank 6,000 or is unlisted. POS shares are token-weighted over jieba's dictionary-level tags, bucketed into coarse categories. Two caveats bear repeating: subtitles and news are proxies for their registers, not the whole of either, and automatic segmentation introduces some noise — but both effects are small against a gap this size.

If a Chinese article feels too big a jump from conversation, compare the levels in a Dengdeng preview. Start with the version you can follow, then try a harder one to see how the wording changes.