What you will get from this article
- What the two “tricks” of language are (word-meaning conventions, rules for combining words)
- Why a “word-chain device” cannot produce the sentences of a language
- What “phrase-structure grammar” solves, and what “a sentence is a tree” means
The answer first. According to Pinker a sentence is not a “chain” of words in a row but a nested “tree”, in which words group into phrases and phrases fit into bigger phrases. Because of this design, the same kind of phrase can be plugged in anywhere in a sentence, and a phrase can contain a phrase of its own kind, so sentences can be endlessly long.
A sentence is not a chain but a tree.
Two tricks (pp. 83–85)
Chapter 4 opens with the journalists’ saying that when a man bites a dog, that is news. Language can carry that “news,” Pinker says, because of two tricks (pp. 83–84).
1. The arbitrary pairing of words and meanings. This is what Saussure called “the arbitrariness of the sign” (see Arbitrariness of the sign for the original text). The word dog does not look, walk, or bark like a dog, yet it means “dog” because every English speaker went through the same rote learning as a child. In return the community gets the ability to convey a concept from mind to mind almost instantly.
2. Rules for combining words. Pinker brings in Humboldt’s famous phrase: language “makes infinite use of finite media.” The difference between Dog bites man and Man bites dog comes from the order of the words. The set of rules that turns order into meaning is a generative grammar. (Not to be confused with the prescriptive or stylistic “grammars” of school.)
On Humboldt see Energeia: language as activity.
A discrete combinatorial system (pp. 84–86)
Grammar is an example of a “discrete combinatorial system,” unusual in nature.
- A finite number of elements (here words) are sampled and permuted into larger structures (sentences).
- The structure’s properties are distinct from those of the elements. The meaning of Man bites dog differs from that of any of its three words and from the reverse order.
- Another example is the genetic code of DNA: four nucleotides yield sixty-four codons and unlimited genes.
- By contrast, in blending systems (paint, cooking, weather) the properties of the combination are the average or mixture of the elements. Red and white make pink.
Pinker says it is no coincidence that life and mind, the two most impressive systems in the universe, are discrete combinatorial; many biologists believe that if inheritance were not discrete, evolution could not have taken place (p. 85).
Sentences are effectively infinite (pp. 85–87)
The first consequence of a discrete combinatorial system is the sheer vastness of language.
- At a random point in a sentence there are on average about ten different words that could be inserted.
- For sentences up to twenty words long, the number is at least 10 to the 20th power (a hundred million trillion). At five seconds a sentence, memorizing them all would take about a hundred trillion years (pp. 85–86).
- Twenty words is far too strict. Shaw has a 110-word sentence; Faulkner’s Absalom, Absalom! has a sentence of about 1,300 words.
- And the number is infinite, by the same logic as for integers: if you think you have the largest, add 1. You can always extend: “Pinker wrote that Faulkner wrote…”, “Who cares that Pinker wrote that Faulkner wrote…?” (pp. 86–87).
This “infinite use of finite media” distinguishes the human brain from virtually all artificial language devices (pull-string dolls, cars that nag you to close the door, voice-mail instructions), which use a fixed list of prefabricated sentences (p. 87).
Grammar is autonomous from cognition (pp. 87–89)
The second consequence is that grammar is a code autonomous from meaning. It specifies how words may combine, independent of the particular meanings we convey.
Pinker lists strings that can be interpreted but are not well formed.
- It’s a flying finches, they are.
- The child seems sleeping.
- Sally poured the glass with water.
- Who did a book about impress you?
Any English speaker senses that something is wrong. “Ungrammatical” here means not the schoolmarm’s split infinitives or dangling participles but the result of having a fixed code for interpretation. Computers complain for the same reason (SYNTAX ERROR) (p. 88).
And the opposite exists: nonsense that is grammatical. Chomsky’s famous example:
Colorless green ideas sleep furiously.
Though senseless, it is well formed. The genre of nonsense verse and prose depends on grammar and sense being independent (Edward Lear, Mark Twain, Lewis Carroll’s “Jabberwocky”). Even when we cannot make sense of Carroll’s poem, like Alice we can tell “somebody killed something.” English speakers can answer questions like “What slept? How? Did one thing sleep, or several? What kind of ideas?” (p. 89).
The word-chain device: the simplest method (pp. 89–93)
How might one build a combinatorial grammar that puts words in order? The simplest is a word-chain device (technical name: finite-state or Markov model).
- A bunch of lists of words (or prefabricated phrases) and directions for going from list to list.
- To produce a sentence, choose a word from the first list, then from the second, and so on.
- To recognize a sentence, check each word against each list in order.
Pinker notes this is a staple of satire. Michael Frayn’s novel The Tin Men (a program that composes royal-occasion stories) and the Social Science Jargon Generator (pick one word from each of three columns to make an impressive term like inductive aggregating interdependence) are examples (pp. 90–91).
A word-chain device is the simplest discrete combinatorial system. A scheme such as the / a / one → (loop on happy) → boy / girl / dog → eats → ice cream / hot dogs / candy generates an unlimited number of sentences such as A girl eats ice cream, The happy dog eats candy, The happy happy dog eats ice cream (p. 92). So does a recorded phone-number voice.
It is congenial to stimulus-response theories (a stimulus elicits a word, which becomes the next stimulus), and some psychologists held that human language is based on a huge word chain (pp. 92–93).
Chomsky’s refutation (pp. 93–96)
But Pinker says the modern study of grammar began when Chomsky showed that word-chain devices are not just suspicious but fundamentally the wrong way to think about how language works (p. 93). There are three reasons.
1. A sentence is a different thing from a string of transition probabilities
Colorless green ideas sleep furiously is senseless but grammatical. In English text the probability that colorless is followed by green is zero. Building strings word by word from probability tables yields Englishy but not English output, such as Pinker’s example House to ask for is to earn our living by working towards a goal for his team in old New-York was a wonderful place wasn’t it even pleasant to talk about and laugh hard when he tells lies he should not tell me the reason why you are is evident. (p. 93).
Two lessons follow.
- Learners of a language record not “which word follows which” but “which word category (noun, verb, …) follows which category.” So we can recognize colorless green ideas as adjective–adjective–noun, like strapless black dresses.
- Nouns, verbs, and adjectives are not hitched end to end in one long chain; there is an overarching blueprint that puts each word in a specific slot (pp. 93–94).
2. Long-distance dependencies
Chomsky showed mathematically that certain sentences cannot in principle be produced by a word-chain device.
- Either the girl eats ice cream, or the girl eats candy.
- If the girl eats ice cream, then the boy eats hot dogs.
Either must later be followed by or; if requires then. A word-chain device does not remember what it chose at the start (it is “an amnesiac”). When it reaches the or/then list it cannot recall whether it said either or if (pp. 94–95).
Could one duplicate the either…or and if…then chains separately? That is a disturbing solution, needing three identical subnetworks. And Chomsky went further: each of these sentences can be embedded in the others, including itself.
- If either the girl eats ice cream or the girl eats candy, then the boy eats hot dogs.
- Either the girl eats ice cream then the boy eats ice cream, or if the girl eats ice cream then the boy eats candy.
Since there is no limit in principle to the number of if’s and either’s, spelling out a duplicate chain for each order would need an infinite number of chains, which won’t fit inside a finite brain (pp. 95–96).
“No real person begins a sentence with Either either if either if…” Pinker replies that Chomsky was just adopting the mathematician’s esthetic, using either–or and if–then as the simplest example of long-distance dependencies.
3. Everyday sentences overflow with them too
Pinker’s example is an old grammarian’s joke about a sentence ending in five prepositions. Daddy trudges upstairs to Junior’s bedroom to read him a bedtime story; Junior spots the book, scowls, and asks, Daddy, what did you bring that book that I don’t want to be read to out of up for? By the final for, Junior has committed himself to four dependencies (to be read demands to, that book requires out of, bring requires up, what requires for) (p. 97). A real letter to TV Guide held four grammatical commitments at once. To handle these in a chain, an identical chain must be duplicated for each combination of dependencies, and “one quickly runs out of brain.”
Here Pinker quotes the Joyce Kilmer line: “Only God can make a tree.”
Phrase-structure grammar: the invention of the tree (pp. 97–101)
In a human grammar, a sentence is not a chain but a tree. Words are grouped into branches, phrases, and small phrases are joined into bigger ones.
Take The happy boy eats ice cream.
- The first three words (the happy boy) form a unit, a noun phrase (NP).
- An English noun phrase is “an optional determiner, followed by any number of adjectives, followed by a noun.” As a rule:
NP → (det) A* N(parentheses = optional, asterisk = as many as you want). - A sentence (S) is a noun phrase followed by a verb phrase (VP):
S → NP VP. A verb phrase is a verb followed by a noun phrase:VP → V NP. - And a dictionary: nouns (boy, girl, dog, ice cream …), verbs (eats, likes, bites), adjectives (happy, lucky, tall), determiners (a, the, one).
Combining these rules gives a tree like this (a simplified redrawing of Pinker’s figure).
S
├─ NP
│ ├─ det: the
│ ├─ A: happy
│ └─ N: boy
└─ VP
├─ V: eats
└─ NP
└─ N: ice cream
This invisible superstructure eliminates all the problems of word-chain devices. The key is that a tree is modular. A symbol like “NP” is like a connector that lets one component (a phrase) snap into any of several positions inside larger components. Once a phrase is defined by a rule it never has to be defined again; it can be plugged in anywhere (p. 99).
- Subject ([The happy happy boy] eats ice cream.)
- Object (I like [the happy happy boy].)
- Indirect object (I gave [the happy happy boy] a cookie.)
- Possessor ([The happy happy boy]‘s cat eats ice cream.)
All use the same NP. There is no need to learn separately, for subject, object, indirect object, and so on, that the adjective precedes the noun (p. 100). And since this is independent of word meaning, it makes colorless green ideas and The girl bites the dog (a newsworthy sentence) with the same rules (p. 100).
Recursion
Either…or and if…then can be written as rules that include a sentence in the definition of a sentence.
S → either S or SS → if S then S
One instance of a symbol embedded inside another instance of the same symbol: logicians call this recursion. A single node holds together each either with its or and each if with its then, so long-distance dependencies are handled by nesting (pp. 100–101).
A tree also shows which words go together in meaning (pp. 101–103)
A tree also helps tie grammar to meaning, because it shows which words belong together, not merely that they are strung in a row. In Either the girl eats ice cream or the girl eats candy, the girl and the ice cream, and the boy and the hot dogs, each form a unit, and in the tree they are branches (p. 101). In a chain it was “just one damn word after another.”
Ambiguous sentences show the tree’s role. A real newspaper example:
- discuss sex with Dick Cavett (“sex” is what is to be discussed, and it is to be discussed with Dick Cavett, or: “sex with Dick Cavett” is what is to be discussed)
The two meanings arise from different ways of joining the words, that is, different tree shapes. If sex with Dick Cavett forms one branch, “sex with Dick Cavett” is the topic. If with Dick Cavett attaches to the verb phrase, the discussing is done with him (pp. 102–103). Pinker says Groucho Marx’s joke works the same way (I once shot an elephant in my pajamas).
Parts of speech are not kinds of meaning (pp. 105–106)
The school definition “a noun is the name of a thing” is not right, Pinker says. Nouns come with all kinds of meanings.
| Noun | Meaning |
|---|---|
| the destruction of the city | an action |
| the way to San Jose | a path |
| whiteness | a quality |
| three miles / three hours | a measurement of space or time |
| the answer | an answer to a question |
| a meeting | an event |
| the square root of minus two | an abstract concept |
| He finally kicked the bucket | no meaning at all (idiom) |
Likewise verbs are not only actions but also mental states (know, like), possession (own, have), and abstract relations (falsify, prove). Conversely a single concept (“being interested”) can be expressed by a noun, verb, adjective, or adverb (her interest in fungi / Fungi are starting to interest her / She seems interested in fungi / Interestingly, the fungi grew…) (pp. 105–106).
So a part of speech is not a kind of meaning but a kind of token obeying formal rules, like a chess piece or a poker chip, which is Pinker’s point.
Cautions
- The scope here is the first part of chapter 4 (pp. 83–106). What follows (the “head” of a phrase, the common anatomy of phrases in all languages, Universal Grammar) was not read.
- The “tree” is ToL’s text rendering of Pinker’s figures.
- Phrase-structure theory has changed greatly since Chomsky. What is presented here is the introductory version Pinker gives to general readers.
Summary
- Language has two tricks: word-meaning conventions and rules for combining words.
- The combining rules are discrete, so the number of sentences is effectively infinite.
- Word-chain devices cannot make the sentences: they break on long-distance dependencies (either…or).
- The phrase-structure “tree” treats phrases as parts that plug in anywhere and solves long-distance dependencies by nesting (recursion).
Next
- The big picture: Is language learned or born?
- Previous: What language do we think in?
- The experiment showing children know “structure”: Why do children know grammar they were never taught?