词的意思由它的邻居决定

你没查过 tezgüino，却知道它能喝、还会醉人

把那四句话拆开，线索一条条都在：「一瓶」说明它是液体、有容量；「桌上」把它放进吃饭的场合；「人人都喜欢」说明是消费品不是工具；「会让你喝醉」一句就推进了含酒精那一类；「用玉米做」给了原料。五条合起来圈出一小块位置，而啤酒、米酒也在那块位置上。你知道的其实是「它跟谁站在一起」。

1954 年，语言学家哈里斯（Zellig Harris）把这件事写成一条主张：两个词的意思差多远，可以从它们出现的环境差多远看出来。三年后，英国语言学家弗斯（J. R. Firth）留下了那句被引用无数次的话——你要认识一个词，看它跟谁来往。两句都在说：别问这个词指向世界上的什么，去看它的邻居。这在当时只是个操作技巧，却有个六十年后才显出分量的副作用——邻居可以数，「指向世界上的什么」不可以。

现在看盲点。把这句补完：「杯子里的水是＿＿的。」热、冷都填得进去。英文同样：the coffee is hot 和 the coffee is cold 一样自然。「这件衣服很＿＿」，贵、便宜都行；「他往＿＿开了两小时」，东西南北随便挑。意思正相反的一对词，偏偏最能互相替换——因为它们要求的邻居几乎完全一样。

所以分布假说测的东西得说准确点：它测「能不能换到同一个位置上」，不测「意思是不是一样」。近义词能互换，反义词也能互换，这套办法就把它们判成一家。这不是哪个程序写坏了，是这个想法的必然推论；后来有一整支研究专门在算完之后往结果里硬塞「这两个是反义词」，才把它按下去一点。

EN Unpack the four sentences and every cue is visible: "a bottle of" makes it a liquid with a volume, "on the table" places it at meals, "everybody likes" makes it a consumable rather than a tool, "makes you drunk" puts it among alcoholic drinks, "out of corn" gives the raw material. Together they fence off a small region — and beer and rice wine sit in that same region. In 1954 Zellig Harris turned this into a methodological claim: difference of meaning correlates with difference of distribution. Three years later J. R. Firth left the much-quoted line that you shall know a word by the company it keeps. Both say: stop asking what a word points to in the world, and look at its neighbours. At the time this was mostly a fieldwork trick, but it had a consequence that took sixty years to matter — neighbours can be counted, and pointing-at-the-world cannot. Now the blind spot. Fill in the blank: "the coffee is ___." Hot and cold both fit. "This coat is very ___": expensive and inexpensive both fit. "He drove ___ for two hours": east, west, north, south. Words with opposite meanings are precisely the words that swap most freely, because they demand almost identical neighbours. So what the distributional hypothesis measures is substitutability of position, not sameness of meaning — and antonyms are maximally substitutable. This is not a bug in some implementation; it follows from the idea itself, it has recurred for decades, and a whole line of work exists that injects antonym information back into the result afterwards to suppress it.

把一张几万乘几万、格子大多是零的表，压成每个词三百个数

怎么把「邻居」变成机器能算的东西？最笨的办法就是最早的办法：开一张大表，每行一个词、每列也一个词，格子里填这两个词在同一小段文字里一起出现过多少次。tezgüino 那行在「瓶」「醉」「玉米」几列上有数，在「螺丝刀」那列是零；酒那行长得跟它很像。相似度就是两行有多像。

问题是这张表既巨大又几乎全空：绝大多数词很少出现，绝大多数词对从没一起出现过，一张五万乘五万的表有用的格子可能不到百分之一。

所以要压。1990 年一组做文献检索的人提出潜在语义分析（latent semantic analysis）：把那张又大又空的表挤成每个词几百个数，挤的时候「汽车」和「轿车」这种总跟同一批词做邻居的行会被合并到相近位置，哪怕它俩从没同句出现过。1997 年兰道尔（Thomas Landauer）和杜迈（Susan Dumais）拿它去考试：用一套百科全书的三万多篇文章（约四百六十万词）训练，做托福同义词选择题，八十道四选一。它答对 64.4%，而同一份题目当年来自非英语国家的申请者平均约 65%。这个数要读得小心：它说明只数共现能在某个特定题型上达到某个人群的平均分，不说明程序「懂英语」——同义词选择题恰好是分布法的主场。

2013 年换了思路：不再先数完再压，而是边猜边学。遮住一句话里的一个词，让程序根据左右邻居猜它是谁，猜错就调整每个词身上那串数。猜了几十亿次之后，那串数就是这个词的位置。这套做法叫 word2vec，出名是因为一个像魔术的现象：「国王」那串数减去「男人」、加上「女人」，落点离「女王」最近。

这个魔术被高估了。2016 年林岑（Tal Linzen）指出一个朴素的问题：标准做法会先把题目里那三个词从候选里排除掉——不排除的话，最近的往往就是「国王」自己。很大一部分成功来自「女王本来就离这三个词很近」，而不是向量里真记着一条从男到女的路。现象是真的，解释被夸大了。

同年还有个更硬的结果：英语和西班牙语各自训练出的空间里，数字词的排布形状差不多，动物词也是，一个统一的变换就能大致对上，拿来猜词典里缺的对应词。「从规则到统计」那期的对齐难题在这里换了面目：那边靠数成对的句子，这边靠两个空间的形状。

EN The crudest way to turn neighbours into arithmetic is also the earliest: build a huge table, one row and one column per word, each cell holding how often the two co-occur within a short span of text. Tezgüino's row has counts under "bottle", "drunk" and "corn" and a zero under "screwdriver"; wine's row looks much the same. But the table is enormous and almost entirely empty — word frequencies are wildly uneven, most words are rare, and most word pairs never co-occur at all, so under one percent of a fifty-thousand-square table carries anything. Hence compression. In 1990 a group working on document retrieval proposed latent semantic analysis, squeezing the table down to a few hundred numbers per word; the squeeze pulls together rows that keep the same company, so "car" and "automobile" land near each other even if they never share a sentence. In 1997 Landauer and Dumais trained it on some thirty thousand encyclopedia articles, about 4.6 million words, and set it the synonym section of the TOEFL — eighty four-way multiple-choice items. It scored 64.4%; applicants from non-English-speaking countries averaged around 65% on the same items. Read that number carefully: it shows that counting co-occurrence alone can reach a certain population's average on one particular item type, not that the program knows English — synonym choice is the home ground of a distributional method. In 2013 the approach changed from count-then-compress to guess-and-adjust: hide a word, have the program guess it from its neighbours, and correct the string of numbers attached to each word when it guesses wrong. After billions of guesses those numbers are the word's position. This is word2vec, famous for a trick that looks like magic: king minus man plus woman lands nearest to queen. The trick is overrated. Tal Linzen pointed out in 2016 that the standard procedure excludes the three input words from the candidate answers — because otherwise the nearest position is usually "king" itself — so much of the success comes from queen already being close to those three, not from a stored route from male to female. A sturdier result from the same year: the spaces trained separately on English and Spanish have similar internal shapes for number words and for animal words, so a single transformation lines them up well enough to guess dictionary entries that were missing. The alignment problem from the machine-translation story reappears here in a different guise — there it was counted sentence pairs, here it is the shape of two spaces.

它学到的是词的意思，还是人用词的习惯

2017 年，卡利斯坎（Aylin Caliskan）和同事在《科学》上把心理学测内隐联想的那套题目原样搬到从网络文本训练的词向量上。结果一条条对上——花类词离愉快的词更近、昆虫类词离不愉快的词更近，这是人身上测过很多次的老结果；而职业词跟女性称谓的接近程度，和该职业实际的女性比例正相关。

关键的读法不是「向量有偏见」，而是它只有共现，所以分不开两件事：「护士跟女性一起出现得多」里，一半是劳动统计上的事实，一半是叙述上的惯性，共现次数对两者一视同仁。要分开，需要的恰好是这套办法从设计上就没有的——词表之外的世界。

还有个更基础的限制：一个词只给一串数，而这串数只能是它所有用法的平均。汉语的「打」就很难办——打电话（dǎ diànhuà，通话）、打水（dǎ shuǐ，取水）、打毛衣（dǎ máoyī，织毛衣）、打车（dǎ chē，叫出租车），邻居分成好几簇，平均到一个位置上谁也不像。英语 bank 同理，河岸和银行共用一个位置。后来的模型改成每次出现单算一个位置，同一个「打」在不同句子里落在不同地方。

那么人呢？人也在用分布——「婴儿是统计学家」那期讲过，还不会说话的婴儿就在数哪些音节总是挨着出现。区别在别处，而且可以拿出来比：1978 年凯里（Susan Carey）和巴特利特（Elsa Bartlett）跟三四岁的孩子说「把那个 chromium 的托盘拿给我，不是蓝色那个，是 chromium 那个」——孩子拿对了橄榄绿那个，一周后还记得这词跟那个颜色有关。这叫快速映射，靠的不是出现次数，是那一次里他能看见两个托盘、能排除掉已经有名字的那个。差异因此可检验：一个词要在共现空间里有个像样的位置通常得出现几十上百次，孩子一次就够；反过来，上百万篇文档里的关联人一辈子也读不完。（复制情况要说清：一次接触学到的是部分信息，能保持多久仍有分歧。）

2020 年本德（Emily Bender）和科勒（Alexander Koller）用一个寓言把界线画得更狠：一只章鱼截住海底电缆偷听岸上两人聊天，久了能接得天衣无缝，但它从没见过岸上任何东西；有一天对方问它怎么用几根棍子挡熊，它就答不上来。主张是：光靠形式学不到意思。反方也不弱——「通货膨胀」「正义」这类词，绝大多数人同样从没见过它指的东西，照样用得好好的。争论今天还在。能确定的只有一句：分布假说给的是一个词在语言内部的位置，不告诉你这个位置挂在世界的哪里。

EN In 2017 Aylin Caliskan and colleagues did something clean in Science: they took the standard psychological test for implicit association and ran it, unchanged, on word vectors trained from web text. The results lined up one by one — flower words sit closer to pleasant words and insect words closer to unpleasant ones, a much-replicated human finding; and the closeness of occupation words to female terms correlates positively with the actual share of women in those occupations. The right reading is not "vectors are biased" but that co-occurrence is all they have, so they cannot separate two things. "Nurse occurs near female terms" is half a fact about the labour force and half an inertia of how people write, and counting treats both alike. Telling them apart requires exactly what the method lacks by design: a world outside the word list. A more basic limit: one word gets one string of numbers, and that string can only be the average of all its uses. Mandarin 打 (dǎ) is a hard case — 打电话 (dǎ diànhuà, "make a phone call"), 打水 (dǎ shuǐ, "fetch water"), 打毛衣 (dǎ máoyī, "knit a sweater"), 打车 (dǎ chē, "hail a taxi"). The neighbours fall into several clusters and one averaged position resembles none of them. English "bank" has the same problem, the riverside and the institution sharing a spot. Later models give every occurrence its own position instead, which is where context begins. What about people? People use distribution too: pre-verbal infants already track which syllables keep each other's company, as the statistical-learning story described. The difference lies elsewhere, and it is concrete enough to test. In 1978 Susan Carey and Elsa Bartlett told three- and four-year-olds "bring me the chromium tray, not the blue one, the chromium one"; the children brought the olive one, and a week later still linked the word to that colour. This is fast mapping, and it does not run on counts — it runs on the child seeing two trays and ruling out the one that already has a name. So the testable difference: a word needs dozens to hundreds of occurrences before it has a decent position in a co-occurrence space, and a child needs one. The reverse holds too — no person will ever read the millions of documents a count can cover. On replication: a single exposure yields partial knowledge, and how long it lasts is still disputed, with reports both of a month's retention and of rapid loss. In 2020 Emily Bender and Alexander Koller drew the line harder with a fable: an octopus taps an undersea cable, eavesdrops on two people ashore, and eventually holds up its end of the conversation seamlessly — having never seen anything on land. Asked one day how to fend off a bear with a few sticks, it has nothing. Their claim is that meaning cannot be learned from form alone. The counter is not weak either: plenty of words that people share are learned purely from how others use them — most speakers of "inflation" or "justice" have never seen what those point at. The argument is unsettled. What is certain is narrower: the distributional hypothesis gives a word's position inside language, and says nothing about where that position is pinned to the world.