语言相遇的时候

借来的词身上带着一个时间戳

回到「明」那三个读音。最早那层跟着佛经从南朝一带过去，读 myō，今天还留在光明（kōmyō）这类词里；第二层是遣唐使那批人从长安带回去的，读 mei，明治（Meiji）、説明（setsumei）都是它；最晚那层读 min，留在明朝体（minchōtai，一种字体）这种词里。「行」也一样：修行（shugyō）、旅行（ryokō）、行灯（andon，纸灯笼）。

关键在于没有哪门语言会先把旧借词清干净再借新的。新的一层直接压上去，旧的留在它当年落脚的那几个词里。所以读音"不统一"不是乱，是历史没被抹掉的样子。

英语有同样的地层。shirt（衬衫）是本土词，skirt（裙子）是从古诺斯语（维京人说的那门语言）借来的同一个词根——英语本土那条路把这个位置的音变成了 sh，诺斯语那边没变，于是两个版本都留下，各分走一半意思。连代词都能借：they、them、their 来自古诺斯语，本土那套第三人称复数被挤掉了。汉语同样有三层：刹那（梵语 kṣaṇa）、塔（stūpa 的省译）跟佛经进来；文化、革命是明治时期日本人用汉字造出来、再借回中文的。

什么好借、什么难借，是有等级的。世界借词数据库（World Loanword Database）把 41 种语言里 1460 个常用概念一个一个查过来源，借词比例从普通话的 1.2% 一直排到 Selice Romani（斯洛伐克一个罗姆人社区的话）的 62.7%。名词最容易借走，动词次之，功能词最难——功能词得插进一整套已经排好的格子里，名词只要有地方放就行。而这个跨度主要不由语言本身决定：普通话那头是多数人日常用不上第二门语言、又有强势的书面标准，Selice Romani 那头是人人多语、少数语言、没有书面标准。借得多不等于这门语言不够用或者不纯——那个百分比是社会处境留下的读数，不是语言质量的评分。

EN Japanese 明 has three on'yomi — myō, mei, min — because Chinese readings were borrowed in three separate waves: with Buddhist texts, with Tang-era embassies, and later still. Nothing cleans out the old layer first, so pronunciation "inconsistency" is history left visible: 光明 kōmyō, 明治 Meiji, 明朝体 minchōtai. English has the same stratigraphy. Native shirt and borrowed skirt come from one Germanic root, split because Old Norse did not undergo the change that gave English sh-; even the pronouns they, them, their are Norse. Chinese shows three layers too: 刹那 (chànà, from Sanskrit kṣaṇa, "an instant"), 文化 (wénhuà, "culture") and 革命 (gémìng, "revolution") coined in Meiji Japan and borrowed back, and later phonetic loans like 沙发 (shāfā, "sofa"). Borrowability is ranked. The World Loanword Database traced 1460 everyday meanings in 41 languages; borrowed shares run from 1.2% in Mandarin to 62.7% in Selice Romani. Nouns go first, verbs next, function words last, because a function word has to fit an existing grammatical slot. That spread mostly reflects social setting — universal multilingualism versus a dominant written standard — not any property of the languages themselves.

词从这边来，语法从那边来

巴布亚新几内亚的通用语托克皮辛语（Tok Pisin），几百万人在用。把它的词一个个拆开，几乎全是英语：mi（我）、yu（你）、gutpela man（好人，来自 good fellow）、kukim（煮某样东西，词尾 -im 来自 him，在这里是个及物标记）。

可它的"我们"有两个。yumi（you + me）包括正在听你说话的那个人，mipela（me + fellow）不包括。这个区分英语完全没有，当地的美拉尼西亚语言有——比如 Kuanua（也叫 Tolai，新不列颠岛东部的话）的代词系统就是这么分的。词壳是英语的，格子是本地的。海地克里奥尔语是同一回事：ap 表示正在进行、pral 表示马上要，语音外壳来自法语的 après 和 après aller，可整套时体标记的排列和分工跟西非 Gbe 语群里的丰语（Fongbe）对得上，不跟法语对。

这就出了个麻烦。「给语言认亲戚」那期的手艺有个默认前提：词表和语法指向同一个祖先。在这里它们指向两边。所以这类语言至今不被塞进普通的语系树——不是因为它们特殊，是因为那棵树只画得出单一来源。

它们是不是自成一类，学界吵了二十多年。McWhorter 提出过"克里奥尔原型"：屈折词缀极少、不用声调区别词义、派生透明，理由是这些语言只有几百年历史。反对一方（DeGraff、Mufwene 等）叫它"克里奥尔例外论"：这些语言走的是和别的语言一样的演变过程，单列一类这件事历史上常和对使用者的贬低绑在一起。证据强度：两边都有大量描写材料，近年有人用定量方法检验语法特征在这类语言形成时是否照常传递，结果偏向"照常"；争论没有结束。要紧的是这里的"简单"只是个可数的技术指标——数屈折词缀有几个——不是"表达力差"，任何一门克里奥尔语都能说任何事。

有条更大的背景让这场争论没那么孤立。Lupyan 和 Dale 2010 年扫了两千多门语言，发现屈折形态的复杂度跟使用人口、接触程度、分布面积都呈负相关：小社区、少有外人来学的语言形态更繁，大范围、很多成年人半路来学的更简。若这条成立，接触重的语言形态偏简就是连续谱上的一个位置，克里奥尔语是谱的一端，不是谱外的另一类。至于这些语法怎么在一代人里被装出来，「孩子能凭空造出语法」那期给的是另一半答案。

EN Take Tok Pisin, spoken by millions in Papua New Guinea. Word by word it looks like English: mi "I", yu "you", gutpela man "a good man" (from good fellow), bilong "of" (from belong), kukim "cook something", where -im (from him) marks a transitive verb. But it has two words for "we": yumi includes the person you are talking to, mipela does not. English has no such distinction; Melanesian languages such as Kuanua (Tolai) do. The shells are English, the categories are local. Haitian Creole is the same story: the progressive ap and prospective pral come phonologically from French après and après aller, while the ordering and division of labour in the whole tense-aspect system patterns after Fongbe, a Gbe language of West Africa. This breaks an assumption behind the comparative method — that vocabulary and grammar point to one ancestor. Whether creoles form a type of their own has been argued for decades: McWhorter's "creole prototype" (minimal inflection, no lexical tone, transparent derivation) against DeGraff's and Mufwene's charge of creole exceptionalism, that these are ordinary languages and the label has historically travelled with contempt for their speakers. Strength: rich description on both sides, recent quantitative work favouring ordinary transmission, no settled verdict. "Simple" here is a countable technical measure, not expressive poverty. Lupyan and Dale's 2010 survey of over 2000 languages found inflectional complexity negatively correlated with population size and contact, which places contact languages at one end of a continuum rather than outside it.

邻居住久了会长得像，哪怕不是亲戚

巴尔干半岛上有件长期让人不解的事。阿尔巴尼亚语、罗马尼亚语、保加利亚语、马其顿语都把定冠词粘在名词后面——英语的 the 站在名词前，它们挂在后面：罗马尼亚语 lup（狼）加上去成 lupul（那只狼），保加利亚语 masa（桌子）成 masata（那张桌子）。这四门语言在印欧语系里分属不同的分支，亲缘关系远；而它们各自最近的亲戚——意大利语、俄语——都不这么干。

还不止这一条。不定式在这一带集体退场：马其顿语和现代希腊语完全不用了，保加利亚语几乎不用，改成"我想我去"这样的从句。将来时则普遍拿"想要"那个动词改造而成，希腊语、罗马尼亚语、保加利亚语、阿尔巴尼亚语的将来标记都这么来的。希腊语没跟上定冠词后置那条，却跟上了这两条——特征是一条一条传的，不打包。

这种情形叫语言联盟（Sprachbund）：一批语言因为挨着住、通婚、做生意，在语法上慢慢趋同。传递者是双语的人——一个人脑子里同时跑着两套系统，把一边的搭法顺手带进另一边，而词不必跟着走。

区域特征还会把人骗得很惨。越南语有声调、词多是单音节、名词前要加量词，看着像汉语像泰语，1912 年 Maspero 就据此把它划给台语那一支。1954 年 Haudricourt 指出，越南语的声调是从丢掉的音节末尾辅音里长出来的：原来收 -h 的那批字变成一组调，收喉塞音（你说"啊——"突然掐断时喉咙里那个封口动作）的变成另一组，时间大约在六到八世纪，这套对应跟孟高棉语族的材料严丝合缝。越南语属南亚语系，最近的亲戚是芒语；声调不是祖先给的，是整个东南亚大陆共用的一件区域行头。教训很直接：最扎眼的那个特征，常常正是借来的那个。

EN Albanian, Romanian, Bulgarian and Macedonian all attach the definite article to the end of the noun — Romanian lup "wolf" becomes lupul "the wolf", Bulgarian masa "table" becomes masata "the table" — although they sit on different branches of Indo-European and their own closest relatives, Italian and Russian, do nothing of the kind. The infinitive has also vanished across the area, completely in Macedonian and Modern Greek and nearly so in Bulgarian, replaced by subjunctive clauses; and the future is built from a verb meaning "want" in Greek, Romanian, Bulgarian and Albanian alike. Greek missed the article feature but shares the other two: features travel one at a time. This is a Sprachbund, a linguistic area, and what carries structure across it is bilingual people, not common descent — patterns move while vocabulary stays put. The extreme case is Kupwar, an Indian village described by Gumperz and Wilson in 1971, where Marathi, Kannada and Hindi-Urdu — two Indo-European, one Dravidian — converged over centuries onto near sentence-by-sentence intertranslatability while keeping separate lexicons; a forty-year restudy in 2016 judged the original estimate of convergence too strong but upheld the direction. Areal features also mislead. Because Vietnamese has tone, mostly monosyllabic words and classifiers, Maspero assigned it to Tai in 1912. In 1954 Haudricourt showed its tones grew out of lost final consonants — syllables once ending in -h fell into one tone class, those ending in a glottal stop into another, around the sixth to eighth centuries — matching Mon-Khmer correspondences exactly. Vietnamese is Austroasiatic; its closest relative is Mường. The loudest feature is often the borrowed one.