Concept

Dependency Length Minimisation

conceptlinguisticslanguagepsychology
Teach me this →

Dependency Length Minimisation

Dependency length minimisation is psycholinguist Edward Gibson‘s account of why the world’s languages have the grammatical shape they do: speakers and listeners keep grammatically connected words as close together as possible, because long-distance connections are harder to produce and harder to understand. It is a single, simple mechanism that explains two things that otherwise look unrelated — cross-linguistic word order — and the universal, extreme difficulty of a specific sentence type called center embedding.

Gibson works within dependency grammar, a framework (running from the ancient Indian grammarian Panini through the French linguist Lucien Tesnière) in which every word in a sentence connects to exactly one other word, producing a tree rooted at the main verb. This makes distance between connected words explicit in a way that Noam Chomsky’s rival phrase-structure grammar — mathematically close to equivalent — buries inside nested brackets.

Word order and Greenberg’s harmonic universals

Stanford typologist Joseph Greenberg first catalogued a striking cross-linguistic pattern: languages that place the verb before its object (‘the dog chased the cat’, as in English) consistently use prepositions (‘to John’), while languages that place the verb after its object (Japanese, Hindi) consistently use postpositions (‘John to’). Of the roughly 1,000 of the world’s ~7,000 languages with adequate documentation, about 95% fit one of these two ‘harmonic’ patterns.

Gibson’s explanation, formalised in later work by his former student Richard Futrell: both patterns independently minimise the total length of dependencies between words. Futrell tested this directly across roughly 60 languages with parsed dependency corpora, comparing each real sentence’s dependency lengths against many random reorderings of the same sentence (constrained to preserve non-crossing dependencies). Real sentences were reliably shorter than the scrambled controls, in every language tested, regardless of whether that language is verb-initial, verb-final, or one of the rare verb-first (VSO) languages.

Center embedding: the universal failure case

The clearest demonstration of dependency cost is center embedding (also called nesting): grammatically valid recursive structures formed by nesting one subject-verb dependency inside another, as in ‘the boy who the cat, which the dog chased, scratched, cried.’ Every human language permits this construction; every human language also makes it close to impossible to produce, understand, or even complete once nested more than once. In Gibson’s crowdsourced completion experiments, people given a partial doubly-nested sentence supply the wrong number of verbs roughly 60% of the time, even reading at their own pace with unlimited time.

Legalese is the one systematic exception Gibson has found to languages generally avoiding long dependencies — and it does so by embracing center embedding rather than minimising it. Working with the lawyer-linguist Eric Martinez, Gibson found roughly 70% of sentences in ordinary contracts contain a center-embedded clause, against 20–30% in other technical prose — usually from defining a term inline, mid-sentence, between a subject and its verb. Passive voice and rare vocabulary are also more common in legal writing but have no measurable effect on comprehension once tested directly; center embedding alone degrades recall and understanding, equally for laypeople and for a sample of 100 practising lawyers. Lawyers, asked directly, preferred plain-language, non-center-embedded rewrites just as strongly as laypeople — evidence against the idea that the convention serves anyone’s actual reading. Gibson’s best guess for why it persists anyway is a ‘magic spell hypothesis’: center embedding functions as a stylistic marker that a text is binding law, similar to how rhyme marks an incantation, rather than a device anyone consciously chooses for communicative value.

Contrast with Chomsky’s movement and innateness

Gibson’s account sits in direct opposition to Noam Chomsky’s theory of English syntax. Chomsky explains a form like ‘will’ shifting to the front of a question (‘Two dogs will enter the room’ → ‘Will two dogs enter the room?’) as literal movement from an underlying deep structure to a surface structure, and argued in 1971 that such movement is unlearnable from ordinary child input — hence must be innate, part of a universal grammar. Gibson favours the rival account defended by linguist Ivan Sag: lexical copying, in which ‘will’ simply has two independently listed forms (declarative and interrogative) with no movement involved. The lexical account correctly predicts that some auxiliaries only work in one form — ‘aren’t I invited?’ is valid with no corresponding declarative (‘I aren’t invited’ does not exist), while ‘ought’ works declaratively but sounds archaic as a question (‘Ought I?’). Removing movement, on Gibson’s telling, removes the specific learnability problem that motivated Chomsky’s innateness claim in the first place.

Where mainstream views differ

Gibson’s dependency-based, learning-heavy account and Chomsky’s phrase-structure-and-movement account, with its innate universal grammar, remain a live and unresolved dispute within linguistics — this concept page presents Gibson’s side, developed from decades of experimental and corpus work, without a direct Chomskyan rebuttal in the room. Gibson’s own methodological framing (experiment and corpus analysis against introspective intuition) is itself part of the disagreement, not a neutral arbiter of it. The broader claim that human languages are optimised for communicative and cognitive efficiency is influential in modern psycholinguistics and computational linguistics but is not universally accepted as the full explanation for word order; Gibson himself flags a related claim — that certain word-order features are shaped by robustness to a noisy communication channel — as speculative, a ‘just-so story’ rather than an established result.

In the wiki