Take the Mandarin string 咬死了猎人的狗 (yǎo sǐ le liè rén de gǒu), roughly bite-dead-hunter-of-dog. Read it once and you probably get: something bit the hunter's dog to death. Now read it the other way — the dog that bit the hunter to death — and the star of the sentence becomes a dog that killed a man.
Not one character added, not one moved. All that changed is where you drew the line: after 咬死了 (yǎo sǐ le, bit to death), or after 猎人的 (liè rén de, the hunter's). Chinese grammar textbooks have used this example for decades, because it forces something normally invisible into the open: a sentence is not a string of words. Words come in gangs, and gangs nest inside gangs.
This issue is about that skeleton: what it looks like, why one string can take two shapes, and its prettiest property — a word can be bound to a slot very far away, but there are places it cannot reach.
First, how to tell which words belong to the same gang. Three tests, all of which you can run on the spot.
One: can the stretch be swapped for a single word? 猎人的狗 (liè rén de gǒu, the hunter's dog) can be replaced by 它 (tā, it). 咬死了猎人 (yǎo sǐ le liè rén, bit the hunter to death) cannot be replaced by anything. Two: can it be moved as a block? Fronting 猎人的狗 (liè rén de gǒu, the hunter's dog) gives a fine sentence; fronting the other stretch gives a broken one. Three: can it be coordinated with something of the same kind? The hunter's dog and the neighbour's cat is fine; the other stretch has no partner.
A stretch that can be swapped, moved and coordinated is a gang, and linguistics calls it a constituent. The three tests are three ways of asking one question: is this stretch being treated as a single thing?
The awkward part is that you cannot say what you are doing. The dividing line is not pronounced and is not written down. Is there any way to see it?
In 2016 Nai Ding and colleagues ran a very clean experiment. They built four-word sentences out of equal-length monosyllables, synthesised them, and played the syllables at a steady four per second with no pauses between words and no rise or fall in pitch — the only rhythm physically present in the sound was the syllable, four times a second. The Mandarin and English materials had the same structure.
Then they recorded listeners' brain activity. It followed three rhythms at once: four per second (syllables), two per second (two-word phrases), one per second (four-word sentences). The last two are not in the acoustic signal at all; the listener builds them. Play the same materials to someone who does not know the language and those two rhythms disappear. Strength of evidence: replicated many times, but there is still an argument over whether it reflects grammatical assembly or the predictability of the next word.
The issue called Words Are Assembled was about assembly inside the word; this is assembly above the word. The same trick, used at two levels.
English has a famous sentence: The horse raced past the barn fell. Native speakers stall on that final fell, and most conclude the sentence is broken. It is not.
When you reach raced you naturally take it as the main verb — the horse ran past the barn. That reading survives until fell, where it collapses, because a sentence cannot have two main verbs. The right shape is: raced past the barn is only a small stretch modifying horse, the main verb is fell, and the whole thing means the horse that was raced past the barn fell over. Sentences like this are called garden path sentences: you are led down a side path and only discover the mistake at the end, then have to walk back to the fork.
This is not just introspection. Frazier and Rayner in 1982 watched people read such sentences with an eye tracker, and saw it: at the word that forces the reversal, fixations get noticeably longer and the eyes jump back to re-read the earlier stretch. Backing up costs something, and the cost can be timed.
Why go wrong at all? Because people commit to a structure before the sentence is over. Every incoming word gets hung on the skeleton built so far, and if it will not hang, the skeleton is taken apart. The benefit is speed — you can finish someone's sentence for them. The cost is the occasional teardown.
Speakers are quietly helping, without either side noticing. Snedeker and Trueswell in 2003 used a sentence you can actually act out: Tap the frog with the flower — either use the flower to tap the frog, or tap the frog that is holding a flower. Asked to say it in front of real objects, speakers put the break in different places: for the use-the-flower meaning they stretch out frog and leave a gap after it; for the frog-holding-a-flower meaning they stretch out tap instead. Listeners' eyes land on the right object before the ambiguous phrase has even arrived. And when the speaker does not know an ambiguity is present, the cues get much weaker — so this is cooperation, mostly unconscious, not a by-product of the mouth.
One last distinction. The Mandarin sentence 鸡不吃了 (jī bù chī le) also has two meanings — the chicken is not eating any more, or we are not eating chicken any more — but it is a different animal from the hunter's dog. There is no second grouping; what changes is the role 鸡 (jī, chicken) plays. Ambiguity has several sources, and structure is only one of them.
To ask what Lily bought, English says What did Lily buy? — what stands at the front, yet it is really the object of buy, and the object slot sits empty. The word is at one end, its real position at the other, and an invisible line runs between them.
That line can be very long. What did you say that Lily thought that John bought? has two layers of who-said-who-thought stuffed in the middle, and native speakers get it in a second. Add as many layers as you like; the sentence gets wordy, not broken.
But some regions cannot be entered. What did you meet the man who bought? — intended as asking about the thing the man you met had bought — is shorter than the previous one, and speakers agree it is broken. In 1967 Ross named these unreachable regions islands: once the far end of the line lands inside something like a relative clause, the line cannot be pulled out.
Mandarin takes the other route: the question word does not move. In 你觉得他说小李买了什么 (nǐ juéde tā shuō Xiǎo Lǐ mǎi le shénme) — what do you think he said Lily bought — the word 什么 (shénme, what) sits quietly in the object slot and never goes to the front, so asking how far it can travel does not apply. The strange part is that the islands are still there. Lu, Thompson and Yoshida ran a formal acceptability-rating study in 2020 and found that Mandarin question words inside a relative clause become hard to accept too — for both argument and adjunct question words, which is the opposite of the textbook generalisation that Mandarin argument question words ignore islands. Strength of evidence: one large-sample rating study, in conflict with a long-standing claim, still under discussion.
Islands are also not identical everywhere. In Swedish and the other Mainland Scandinavian languages, pulling a phrase out of a relative clause is accepted by native speakers and shows up in real corpora at a non-trivial rate; neighbouring Icelandic does not accept it. Nobody has settled how to explain the difference.
The bigger argument is whether an island is a ban written into the grammar or simply a sentence that costs too much to process. Sprouse, Wagers and Phillips tested this directly in 2012: measure each person's working-memory capacity, then see whether it predicts how strongly that person rejects island sentences. It does not. They took this as evidence that processing load does not explain islands away; the other camp (Hofmeister, Sag and colleagues) replies that load has many sources and one memory measure will not capture them. Strength of evidence: one well-designed large-sample study plus an unfinished conversation. Both sides agree on one thing: nobody teaches this constraint.
You can judge in a second that the hunter's dog is one gang and bit the hunter to death is not, yet you cannot state the grounds. Perhaps being able to state it was never the test of reality — you recognise a face instantly without being able to say which features you compared. If so, what do the substitution and movement tests measure: the ability itself, or the traces it leaves?
One answer is that islands hold, and the Swedish structures that look like relative clauses simply are not relative clauses. That sounds a lot like redrawing a boundary to save a conclusion. What evidence could separate it from the claim that islands genuinely vary by language? If none exists, is the disagreement still an empirical one?
If communicative efficiency were pressure enough, structures like the hunter's dog should have worn away. They are everywhere instead. Perhaps ambiguity is barely a cost, because context nearly always removes it and the listener does the removing for free. If that is the story, in what settings does the cost become visible again?
English moves the question word to the front, which publicly announces that an empty slot is coming and lets the listener prepare. Mandarin does not move it, so what tells a listener this is a question, and which part is being asked about? If the work is done by intonation and a sentence-final particle, are the two routes doing the same job with the marker placed at opposite ends of the sentence?