先请你猜一猜：在全部唐诗里，出现次数最多的字是哪一个？

月？酒？愁？还是“床前明月光”里的某个字？

都不是。答案是“不”。

这篇文章用的是开源的《全唐诗》数据，统计范围是康熙年间编定的九百卷：去掉重复后 49,037 首，256 万字。怎么数的、哪些不算，文末有一段“统计口径”。

《全唐诗》是康熙四十四年（1705 年）下令编的。彭定求等十位翰林奉旨编校，曹雪芹的祖父曹寅在扬州主持刊刻，不到两年就编成了九百卷，自称“得诗四万八千九百余首，凡二千二百余人”。

里面最常见的是哪种诗？按句数和每句的字数粗分一下：五言八句最多，占 28.8%；七言四句次之，占 20.3%。像《静夜思》那样的五言四句，只占 6.8%。

“不”字一共出现了 26,007 次，平均每 100 个字里就有 1 个；每 10 首唐诗里，大约有 4 首用到了它。

排在前面的，还有一批和它同类的字：“无”排第 4，“何”排第 12。把“不、无、未、莫、非”这几个表示否定的字算在一起，62.6% 的唐诗里至少有一个。

把两个字连起来看，就更明显了。全唐诗里最常见的两字组合，第一名是“何处”，出现在 1,498 首诗里。前十名里，“不知”“不见”“不可”“不得”占了四个，“万里”“千里”又占了两个。

把这些放在一起读：不知道，看不见，不可以，得不到；问在哪里；隔着万里千里。诗人最常写的，不是眼前有什么，而是想要却没有的：人不在身边，家回不去，时间留不住。

情绪字也是这样。“愁”出现了 4,245 次，是“喜”（1,288 次）的 3.3 倍。

诗，是从“得不到”里长出来的。

去掉“不、之、而、何”这些虚字，剩下的实字里排在前面的是：人、山、风、日、云、来、天、中、时、花、上、春、水、月。“月”排在全部字的第 19 位，每 5 首唐诗里就有 1 首写到月亮。

我又挑了三组字来看：季节、颜色和数字。

春天压倒性地多。“春”出现了 10,908 次，是“夏”的近 12 倍、“冬”的 21 倍；“秋”排第二，7,945 次。诗人偏爱的，是春和秋这两个变化的季节：一个开始，一个凋零。

颜色里，“白”遥遥领先，8,368 次：白云、白日、白发。最少的是“黑”，只有 506 次。

数字最有意思。除了“一”，用得最多的是“三”“千”“万”，三个字几乎一样多，都在 5,600 到 6,000 次之间，远远多过“二”“四”“七”“八”。诗人不爱精确，爱的是“三”这样的虚数，和“千”“万”这样的夸张：万里，千里。“二”只有 2,002 次，不过“两”和“双”分担了很多“二”的活儿，各有两千多次。

全唐诗一共用到了 8,653 个不同的字，但字和字的地位很不平等。最常用的 253 个字，就占了全部用字的一半；认识最常用的 917 个字，唐诗里八成的字你都认得；1,564 个字覆盖九成。另一头，有 1,506 个字在整部全唐诗里只出现过一次。

这种“少数字包办大部分”的规律，语言学里叫齐夫定律：一个字的排名越靠后，出现的次数就越少，而且按幂律递减。

比较两位诗人，不能只比次数。杜甫名下的字比李白多，约 10.6 万 字对 8.4 万 字，什么字都会多一些。我用了统计里常用来比较两份文本的一种方法，叫带先验的对数几率比（Monroe 等人，2008）。它会扣掉总字数的差别，也会照顾到每个字本来有多常见。结果是这样的：

李白的字典里，是月、海、天、金、玉、笑：37% 的李白诗写到了月亮，杜甫只有 15%。杜甫的字典里，是老、病、衰、江、峡、雨：近四分之一的杜甫诗里有“老”字；“病”字杜甫用了 162 次，李白只有 8 次。一个望着天上的月亮和大海，一个守着眼前的江水和病体。

还有一个意外：论写酒，白居易比李白更勤。20.6% 的白居易诗里有“酒”字，李白是 16.7%，全唐诗平均只有 9.4%。

上面那张图里，李白偏爱的字里有一个“言”。这个字李白用了 135 次，杜甫只用了 3 次。

这就怪了。按全唐诗的平均用字率，杜甫名下那十万多字里，“言”应该出现 145 次左右，实际只有 3 次。

杜甫的祖父，名叫杜审言。

古人有避讳的规矩：说话写字，要避开长辈的名字。杜甫避父亲的讳是有名的。他父亲叫杜闲，前人早就注意到，杜诗里几乎找不到“闲”字，个别例外也被考证为后来传刻的错字。数据也是这样：按平均应有约 190 次，实际只有 2 次。

“言”字不一样。唐律有“二名不偏讳”的规定：长辈的名字是两个字时，单用其中一个字并不算犯讳。也就是说，杜甫本来不必避“言”字。但在他的诗里，“言”几乎消失了。

唐人对家讳，常常比法律要求的更严。李贺的父亲名叫晋肃，就有人说李贺不该去考“进士”，韩愈为此专门写了一篇《讳辩》，引的正是“二名不偏讳”这一条。

杜甫是不是在避祖父的讳？数据只能给出线索，给不出定论。这一点我没有在常见的资料里查到专门的讨论，欢迎懂行的朋友指正。

作为对照，还有一个流传很广的说法：杜甫的母亲名叫海棠，所以他从不写海棠。数据查不出什么：整部全唐诗里“海棠”一共只出现了 41 次，按杜甫的字数，他本来也只该写到一两次。没写，并不反常。

说一个字“少”，要先问它本该有多少。

宋人笔记里有个故事，说白居易每写一首诗，先念给一位老太太听：

这个故事未必是真的，前人就有怀疑。不过如果它说得对，白居易用的字应该最“平常”。我拿全唐诗里最常用的 1,000 个字当作“常用字”，算出每位诗人有多少比例的字落在常用字之外，比较的是名下至少有 200 首诗的 61 位诗人。

结果有点意外：白居易是 17.9%，在 61 位诗人里排第 36，比中位数（17.0%）还“冷”一点，也比李白（16.2%）冷。最“平常”的是赵嘏（11.0%），最“冷”的是韩愈（28.2%），皮日休、陆龟蒙、元稹、李贺、杜甫紧随其后。

但只看比例会看错。把每个人在常用字之外用得最多的字拿出来，差别就清楚了：

白居易的“冷字”是睡、饱、味、妻、慵、酌：吃饭睡觉，喝酒犯懒，全是过日子的字。这些字在诗里少见，在生活里天天都用，老太太当然听得懂。韩愈的“冷字”是汝、余、谓、固，这些是写文章的字，正好应了后人说他“以文为诗”。李贺是铜、蛾、碎、虫，阴冷幽怪。杜甫是戈、甲、巫、涛，兵荒马乱，又困在巫峡。

所以，“老妪能解”量的是生活里的常用字，不是诗里的常用字。用诗的尺子量，白居易并不最浅。

白居易的“冷”，冷在把日子写进了诗。

数完这四万多首诗，说说我的几点看法。

我之前写过理性和感性。这一期是“理”的那一半：数据能告诉我们诗人写了什么，写了多少；至于为什么动人，还得靠读的人自己去感受。

四万多首诗，二百五十多万字，数到最后，排在第一的是一个“不”，最常问的是一句“何处”。

“诗人最常写的，
是他们没有的东西。

First, take a guess: across all Tang poems, which character appears most often?

月 (moon)? 酒 (wine)? 愁 (sorrow)? Or one of the characters in “床前明月光” (“Before my bed, the bright moonlight”)?

None of these. The answer is “不” (not).

This article uses open-source data for the “Complete Tang Poems” and covers the 900 volumes compiled under Kangxi: 49,037 poems once duplicates are removed, 2.56 million characters. How I counted and what I left out are set out in the “How I counted” section at the end.

The “Complete Tang Poems” was commissioned in the 44th year of the Kangxi reign (1705). Peng Dingqiu and nine other Hanlin academicians edited it by imperial order, and Cao Yin, grandfather of Cao Xueqin, oversaw the printing in Yangzhou. In under two years all 900 volumes were done, gathering, by its own count, “more than 48,900 poems by more than 2,200 poets.”

What kind of poem is most common? A rough sort by number of lines and characters per line: eight lines of five characters is the most common, at 28.8%, followed by four lines of seven characters, at 20.3%. Four lines of five characters, the form of “Quiet Night Thoughts,” make up only 6.8%.

“不” (not) appears 26,007 times in all, about once in every 100 characters, and roughly 4 in every 10 Tang poems use it.

Higher up the list are more characters of the same kind: “无” (without) ranks 4th and “何” (what, where) 12th. Taken together, the negating characters 不 (not), 无 (without), 未 (not yet), 莫 (do not) and 非 (is not) appear at least once in 62.6% of Tang poems.

Look at pairs of characters and it becomes even clearer. The most common two-character combination in the “Complete Tang Poems” is “何处” (where), found in 1,498 poems. In the top ten, “不知” (don’t know), “不见” (can’t see), “不可” (cannot) and “不得” (can’t get) take four places, and “万里” (ten thousand miles) and “千里” (a thousand miles) take two more.

Read them together: don’t know, can’t see, cannot, can’t get; asking where; ten thousand miles, a thousand miles away. What poets write about most is not what is in front of them, but what they want and do not have: someone who isn’t there, a home they can’t go back to, time that won’t stay.

Words for feelings tell the same story. “愁” (sorrow) appears 4,245 times, 3.3 times as often as “喜” (joy, 1,288 times).

Poetry grows out of what we cannot have.

Set aside function characters such as 不, 之, 而 and 何, and the content characters at the top are 人 (person), 山 (mountain), 风 (wind), 日 (sun), 云 (cloud), 来 (come), 天 (sky), 中 (in), 时 (time), 花 (flower), 上 (on), 春 (spring), 水 (water) and 月 (moon). “月” ranks 19th among all characters: one Tang poem in five mentions the moon.

I also picked three groups of characters to look at: seasons, colors and numbers.

Spring wins by a landslide. “春” (spring) appears 10,908 times, nearly 12 times as often as “夏” (summer) and 21 times as often as “冬” (winter); “秋” (autumn) comes second, with 7,945. What poets love are the two seasons of change, spring and autumn: one a season of beginnings, the other of withering.

Among colors, “白” (white) is far ahead, at 8,368: white clouds, the white sun, white hair. The least used is “黑” (black), with only 506.

Numbers are the most interesting. Apart from “一” (one), the most used are “三” (three), “千” (thousand) and “万” (ten thousand), almost neck and neck at between 5,600 and 6,000 each, and far ahead of “二” (two), “四” (four), “七” (seven) and “八” (eight). Poets have no love for precision. What they love is a loose “three,” meaning “several,” and the exaggeration of “thousand” and “ten thousand”: ten thousand miles, a thousand miles. “二” appears only 2,002 times, though “两” and “双” (both meaning “two” or “a pair”) take on much of its work, with more than two thousand occurrences each.

The “Complete Tang Poems” uses 8,653 distinct characters in all, but characters are far from equal. The 253 most common account for half of all the text; know the 917 most common, and you will recognize eight in ten of the characters you meet in Tang poetry; 1,564 characters cover nine in ten. At the other end, 1,506 characters appear exactly once in the whole collection.

This pattern, in which a few characters do most of the work, is known in linguistics as Zipf’s law: the further down the ranking a character sits, the less often it appears, and the counts fall off as a power law.

To compare two poets, it isn’t enough to compare counts. Du Fu left more text than Li Bai, about 106,000 characters against 84,000, so nearly every character shows up more often in his poems. I used a method that statisticians often use to compare two texts, the log-odds ratio with an informative prior (Monroe et al., 2008). It corrects for the difference in total length and also allows for how common each character is to begin with. Here is the result:

Li Bai’s dictionary holds 月 (moon), 海 (sea), 天 (sky), 金 (gold), 玉 (jade) and 笑 (laugh): 37% of Li Bai’s poems mention the moon, against only 15% of Du Fu’s. Du Fu’s dictionary holds 老 (old), 病 (ill), 衰 (decline), 江 (river), 峡 (gorge) and 雨 (rain): nearly a quarter of Du Fu’s poems contain “老,” and Du Fu used “病” 162 times, Li Bai only 8. One gazes at the moon in the sky and at the sea; the other keeps watch over the river before his eyes and his own failing body.

One more surprise: when it comes to wine, Bai Juyi wrote about it more often than Li Bai. “酒” (wine) appears in 20.6% of Bai Juyi’s poems, against 16.7% of Li Bai’s and just 9.4% across the “Complete Tang Poems” as a whole.

Among the characters Li Bai favors in that chart is “言” (speak). Li Bai used it 135 times; Du Fu used it only 3 times.

That is odd. At the average rate for the “Complete Tang Poems,” the hundred thousand or so characters under Du Fu’s name should include “言” about 145 times. The actual count is 3.

Du Fu’s grandfather was named Du Shenyan (杜审言).

In old China there was a rule of name taboo: in speech and in writing, you avoided the names of your elders. Du Fu’s avoidance of his father’s name is well known. His father was called Du Xian (杜闲), and scholars noticed long ago that the character “闲” (idle) is almost nowhere to be found in Du Fu’s poems; the few exceptions have been traced to errors made in later copying and printing. The data agree: at the average rate it should appear about 190 times, but it appears only twice.

“言” is a different matter. Tang law had a rule that “a two-character name is not tabooed one character at a time”: when an elder’s name had two characters, using just one of them was no offense. In other words, Du Fu had no need to avoid “言.” Yet in his poems, “言” all but vanishes.

Tang people were often stricter about family taboos than the law required. Because Li He’s father was named Jinsu (晋肃), some said Li He should not sit for the jinshi (进士) examination, and Han Yu wrote a whole essay, “On Taboo Names,” in his defense, citing precisely this rule about two-character names.

Was Du Fu avoiding his grandfather’s name? The data can offer a clue, but not a verdict. I haven’t found any specific discussion of this in the usual sources, and I would welcome corrections from readers who know the field.

For contrast, there is a widely repeated story that Du Fu’s mother was named Haitang (海棠, crabapple blossom), so he never wrote about crabapples. The data find nothing to support it: “海棠” appears only 41 times in the whole “Complete Tang Poems,” and for the amount Du Fu wrote, he would only be expected to mention it once or twice. Not writing it is nothing unusual.

Before calling a character “rare,” ask how often it ought to appear.

A Song-dynasty book of jottings tells a story: whenever Bai Juyi wrote a poem, he would first read it to an old woman.

The story may not be true; it was doubted long ago. But if it were right, Bai Juyi’s characters should be the most “ordinary.” I took the 1,000 most common characters in the “Complete Tang Poems” as the “common characters,” worked out what share of each poet’s characters fall outside that set, and compared the 61 poets with at least 200 poems to their name.

The result is a bit of a surprise. Bai Juyi comes in at 17.9%, 36th of the 61 poets: slightly “rarer” than the median (17.0%), and rarer than Li Bai (16.2%) too. The most “ordinary” is Zhao Gu (11.0%), and the “rarest” is Han Yu (28.2%), followed closely by Pi Rixiu, Lu Guimeng, Yuan Zhen, Li He and Du Fu.

But the share alone can mislead. Take the characters each poet uses most outside the common set, and the differences become clear:

Bai Juyi’s “rare characters” are 睡 (sleep), 饱 (full), 味 (taste), 妻 (wife), 慵 (lazy) and 酌 (pour wine): eating and sleeping, drinking and lazing about, all words for getting through the day. They are rare in poetry but used every day in life, so of course an old woman could understand them. Han Yu’s are 汝 (you), 余 (I), 谓 (say) and 固 (indeed): words from prose essays, which bears out what later critics said of him, that he “wrote poetry as prose.” Li He’s are 铜 (bronze), 蛾 (moth), 碎 (shattered) and 虫 (insect): gloomy, cold and eerie. Du Fu’s are 戈 (spear), 甲 (armor), 巫 (Wu, as in the Wu Gorge) and 涛 (waves): war and upheaval, and a man stranded in the Wu Gorge.

So “even an old woman could understand” measures the characters common in daily life, not the ones common in poetry. Measured with poetry’s ruler, Bai Juyi is not the plainest poet.

Bai Juyi’s “rare” characters are rare because he wrote everyday life into his poems.

Having counted these forty-odd thousand poems, here are a few thoughts of my own.

I wrote before about reason and emotion. This issue is the “reason” half: data can tell us what the poets wrote and how much of it; as for why it moves us, that is still for each reader to feel.

Forty-odd thousand poems, more than two and a half million characters, and at the end of the count, the character in first place is “不” (not), and the question asked most often is “何处” (where).

“What poets write about most
is what they do not have.