Orthography

Correct pinyin is not a run of syllables. It is word-spaced, capitalised and punctuated, and 我要去北京玩儿。is Wǒ yào qù Běijīng wánr. rather than wǒ yào qù běi jīng wán er.

convert(dictionary, "我要去北京玩儿。"); // "Wǒ yào qù Běijīng wánr."

Three separate things happen in that one example. 北京 groups into one word, it capitalises as a proper noun, and 儿 attaches as an r-suffix instead of surfacing as er. The reference standard is 《汉语拼音正词法基本规则》 GB/T 16159—2012, and this is a distinct pipeline stage rather than a side effect of the decoder.

Turn the whole spacing pass off with grouping: false; capitals and apostrophes have options of their own.

Word spacing (分词连写)

The decoder produces words; 分词连写 decides which of them are written together.

convert(dictionary, "他看了"); // "tā kànle" — aspect particle attaches
convert(dictionary, "走着"); // "zǒuzhe"
convert(dictionary, "我的"); // "wǒ de" — 的 stands alone
convert(dictionary, "桌子"); // "zhuōzi" — suffix attaches to its stem
convert(dictionary, "现代化"); // "xiàndàihuà"
convert(dictionary, "一个人"); // "yí gè rén" — measure word separates
convert(dictionary, "南京市"); // "Nánjīng Shì" — place generic separates
convert(dictionary, "南京市", { grouping: false }); // "Nánjīngshì"

Three rules survive being measured against the whole dictionary rather than against a corpus, and those three are what is implemented:

• the aspect particles 了, 着 and 过 attaching to a verb or adjective
• a k-tagged suffix attaching to its stem
• the generic half of an administrative place name, where the word is tagged ns and the part before the generic is itself a dictionary entry

That second condition on place names is not decoration. Without it 上山下乡 is tagged ns and comes apart as Shàngshānxià Xiāng.

The curated list

Some spacings no rule reaches, and those live in a curated word list instead:

convert(dictionary, "不是"); // "bú shì"
convert(dictionary, "一天"); // "yì tiān"
convert(dictionary, "黄河"); // "Huáng Hé"
convert(dictionary, "中国人"); // "Zhōngguórén"
convert(dictionary, "我还给你了。"); // "Wǒ huán gěi nǐ le."

Each of those defeated a rule for a specific reason:

Wanted	Why no rule reaches it
不是 bú shì	jieba tags it c, the same tag as 不但 bùdàn
一天 yì tiān	of 247 numeral+量词 candidates many are lexicalised: 大米, 层次
黄河 Huáng Hé	青海 is Qīnghǎi and 上海 is Shànghǎi; nothing separates them
中国人 Zhōngguórén	no suffix tag marks 人, and the decode splits it
还给 huán gěi	开会 and 睡觉 are also verb pairs and are written together

The list is curated, so words it does not cover are written the way the rules leave them:

convert(dictionary, "不但"); // "búdàn"
convert(dictionary, "大米"); // "dàmǐ"
convert(dictionary, "青海"); // "Qīnghǎi"

Why splitting is harder than joining

Joining what the decode separated overrides nothing — the dictionary had no entry for the joined form, so there was no evidence either way. Splitting a decoded word contradicts positive evidence that it is one word, so it needs a condition strong enough to survive the whole dictionary, or a listed entry with a reason attached. That asymmetry is why the rule set is short and the list exists.

Capitals

convert(dictionary, "银行"); // "yínháng", not "Yínháng"
convert(dictionary, "我要去北京。"); // "Wǒ yào qù Běijīng."
convert(dictionary, "李华"); // "Lǐ Huá"
convert(dictionary, "长江"); // "Cháng Jiāng"
convert(dictionary, "你好，世界"); // "nǐ hǎo, shìjiè"

Proper nouns always; the first word of a sentence only where the source is punctuated as one. That punctuation is the only thing separating 学生 looked up as a word from 这是我的书。written as a sentence, and a comma does not count.

Proper-noun detection comes from jieba’s part-of-speech tags — nr person, ns place, nt organisation, nz other. Personal names absent from every dictionary are recovered from a closed surname list of about 500 entries plus “the next one or two characters are the given name”, which is what gets Lǐ Huá.

Inheriting jieba’s tags means inheriting its mistakes. 无缝钢管 is tagged nz, so it converts as Wúfènggāngguǎn with a capital it has not earned.

Apostrophes (隔音符号)

convert(dictionary, "西安"); // "Xī'ān"
convert(dictionary, "天安门"); // "Tiān'ānmén"
convert(dictionary, "女儿"); // "nǚ'ér"
convert(dictionary, "海鸥"); // "hǎi'ōu"

An apostrophe goes before any syllable of a word that begins with a, o or e and is not the first. Those three are the complete trigger set: i, u and ü surface as y and w in that position and cannot create a boundary ambiguity.

apostrophe: "standard" writes it only where leaving it out would genuinely read as something else, which is what GB/T 16159 asks for. "always" is the default because that is what essentially every style guide does.

Punctuation

。，、；：？！ are rewritten as their Latin equivalents and take the space the full-width glyph carried:

convert(dictionary, "北京。"); // "Běijīng."
convert(dictionary, "北京。", { punctuation: "keep" }); // "Běijīng。"

Brackets and quotation marks are left alone under either setting.

儿 and 儿化

One character, three behaviours, so this is a per-word dictionary fact rather than a rule:

convert(dictionary, "玩儿"); // "wánr" — retroflex suffix, one syllable
convert(dictionary, "女儿"); // "nǚ'ér" — full syllable, needs an apostrophe
convert(dictionary, "儿子"); // "érzi" — full syllable, word-initial

The erhua flag comes from CC-CEDICT’s explicit r5 token. Source data is unreliable here — the phrase corpus writes 玩儿 as wán er — which is why the build refuses to write an artifact unless 儿化 round-trips both ways.

What is not implemented

The standard is larger than what is built. Known gaps, all of which degrade to something readable rather than something wrong:

GB/T 16159 says	What happens now
成语 hyphenate 2+2: 风平浪静 → fēngpíng-làngjìng	fēngpínglàngjìng, no hyphen
4+ syllable compounds split: 无缝钢管 → wúfèng gāngguǎn	Wúfènggāngguǎn, unsplit
AABB/ABAB reduplication hyphenates	spaced: 研究研究 → yánjiū yánjiū
老王 → Lǎo Wáng	lǎo Wáng — 老 is not treated as a prefix

A coverage caveat worth knowing before expecting more from the tag-conditioned rules: 487,552 of 721,718 Han words carry no part-of-speech tag at all, since only the jieba-sourced third of the dictionary has one. Any rule conditioned on a tag is silently inert on two thirds of the vocabulary.