Notes — Edward Gibson on Human Language, Psycholinguistics, and LLMs

Notes on Edward Gibson in conversation with Lex Fridman — Lex Fridman Podcast (https://lexfridman.com/edward-gibson), 17 April 2024.

Four questions [Adler frame]

Q1 — What is it about as a whole? Gibson, an MIT psycholinguist, presents a single unifying account of why human languages have the grammatical shape they do: speakers and listeners minimise the cognitive cost of long-distance dependencies between words. From that one pressure he derives cross-linguistic word-order universals, the specific unintelligibility of center-embedded sentences, and the peculiar difficulty of legal writing. He then pivots to a separate but connected claim — that language and thought are distinct brain systems, evidenced by fMRI localisation and by stroke patients who lose language but keep full reasoning — and uses that separation to frame what large language models have and have not learned: excellent form, unreliable meaning. A final third of the conversation covers fieldwork among the Piraha and Tsimane of the Amazon, where the absence of exact-counting vocabulary shows that even something as basic as number words is a cultural invention rather than a cognitive universal.

Q2 — How is it argued? Gibson argues from controlled experiment and corpus analysis, explicitly contrasting his method with Noam Chomsky’s reliance on introspective thought experiments. Key evidence types: cross-linguistic corpus statistics (Richard Futrell’s ~60-language dependency-length study); psycholinguistic experiments (sentence-completion tasks, acceptability ratings, reading times); fMRI studies (Evelina Fedorenko’s language-network localisation); a contract-analysis study with Eric Martinez comparing lawyers and laypeople on center-embedded legal text; and field elicitation among the Piraha (counting and matching tasks) and comparative colour-term counts across cultures. The dependency-length claims are presented as well-supported; the noisy-channel explanation for some word-order features is explicitly flagged by Gibson himself as speculative (‘just-so stories’).

Q3 — Is it true, in whole or part? The dependency-length-minimisation account of word order is empirically well grounded — Futrell’s cross-linguistic corpus result is a real, replicated finding, and the center- embedding difficulty is uncontroversial and easy to demonstrate. The Fedorenko language- network findings are an active, mainstream research programme, though Gibson’s inference from them (language is not necessary for thought) is stronger than some cognitive scientists would accept — the global-aphasia evidence supports ‘not sufficient/not identical’ more securely than it rules out language playing some supportive role in ordinary thinking [?]. Gibson’s characterisation of Chomsky’s movement theory and the 1971 innateness argument is one side of a live, decades-long dispute in linguistics, presented here without a Chomskyan rebuttal in the room [?source] — a fair discipline note, not a flaw in Gibson’s own reasoning. The Piraha counting findings (Gibson, Everett, and Frank’s published work) are well documented; the ‘magic spell hypothesis’ for legalese’s center embedding is offered as Gibson’s own best guess, not an established result.

Q4 — What of it? This is the wiki’s first dedicated psycholinguistics source and gives it two durable, homeless mechanisms worth their own concept pages: dependency length minimisation (the cognitive-cost theory explaining word order and center embedding) and the language-thought dissociation (Fedorenko’s fMRI case that language is a communication system separate from reasoning). Both bear directly on how the wiki should think about LLM capability claims elsewhere — Gibson’s form/meaning distinction is a sharper, more falsifiable version of a debate that recurs across the wiki’s AI-focused episodes. The episode also surfaces Noam Chomsky as a load-bearing subject with no existing person page despite being referenced across several other wiki pages, which this ingest corrects.

Glossary

Dependency grammar — a grammatical framework in which every word in a sentence connects to exactly one other word, producing a tree rooted at the main verb; contrasted with Chomsky’s phrase-structure grammar, which is mathematically near-equivalent but makes distances between connected words harder to see. [§ Dependency grammar]

Dependency length — the number of words (or, more precisely, an unresolved unit — Gibson suspects an exponential cost function) separating two words that are grammatically connected; the quantity Gibson argues all languages minimise. [§ Generalizations in language]

Harmonic word order — Joseph Greenberg’s observation that verb-initial languages (like English) consistently pair with prepositions, while verb-final languages (like Japanese) consistently pair with postpositions — both patterns independently minimise dependency length. [§ Generalizations in language]

Center embedding (= nesting) — grammatically valid but cognitively near-impossible recursive structures formed by nesting one subject-verb dependency inside another; the one construction type that is difficult in every human language without exception. [§ Center embedding]

Movement (Chomsky) — Chomsky’s theory that syntactic forms like questions are derived by literally moving a word (e.g. an auxiliary verb) from an underlying ‘deep structure’ to a ‘surface structure’. [§ Dependency grammar]

Lexical copying — the rival account (associated with linguist Ivan Sag) in which a word like ‘will’ simply has two independently listed forms — declarative and interrogative — with no movement involved; Gibson’s preferred alternative to Chomsky’s movement theory, in part because it is more learnable. [§ Dependency grammar]

Language network — Evelina Fedorenko’s fMRI-defined, left-lateralised brain network, localisable in about fifteen minutes per person, that activates for comprehending or producing language (spoken or written, in any language known) and stays silent during chess, music, arithmetic, or computer programming. [§ Thinking and language]

Global aphasia — a condition, typically from a large left-hemisphere stroke, in which the language network is destroyed while other cognitive faculties (chess, driving, arithmetic) remain intact — evidence that language is not required for thought. [§ Thinking and language]

Construction grammar — a usage-based linguistic theory in which the basic unit of language is a form-meaning pairing (a word or a combination of words), rather than an abstract generative rule system; Gibson locates large language models closest to this tradition. [§ LLMs]

Noisy channel — Claude Shannon’s information-theoretic framing of communication as a signal degraded by background noise, speaker error, and listener inattention; Gibson speculates some word-order features are shaped by robustness to this noise, distinct from the dependency-length pressure. [§ LLMs]

Key claims by section

Generalizations in language [§ Generalizations in language]

About 95% of the roughly 1,000 well-documented languages (of ~7,000 worldwide) are either consistently verb-initial-with-prepositions or verb-final-with-postpositions.
Richard Futrell’s cross-linguistic study (~60 languages with parsed corpora) shows real sentences have shorter dependencies than random or even word-order-constrained scramblings of the same sentence, regardless of the language’s underlying word order.

Dependency grammar [§ Dependency grammar]

Dependency grammar and Chomsky’s phrase-structure grammar are near mathematically equivalent (each dependency grammar has a phrase-structure grammar generating the same sentences, and vice versa), but dependency grammar makes distance between connected words explicit.
Gibson’s central methodological difference from Chomsky: he uses experiments and corpus data; Chomsky, by Gibson’s account, has never used either, relying on introspective intuition alone.
Gibson rejects Chomsky’s movement theory in favour of lexical copying, partly because lexical copying correctly predicts that some auxiliaries work only in one form (‘aren’t I’ has no declarative counterpart; ‘ought’ has an archaic-sounding interrogative).

Center embedding [§ Center embedding]

Legal contracts contain a center-embedded clause in roughly 70% of sentences, versus 20–30% in other technical prose — the highest rate Gibson’s team found in any text type.
Passive voice and low-frequency vocabulary appear more often in legalese but have no measurable effect on comprehension; center embedding alone degrades recall and understanding, equally in laypeople and in 100 tested practising lawyers.
Lawyers, when asked directly, preferred non-center-embedded rewrites just as strongly as laypeople did — undercutting the idea that dense legal prose serves lawyers’ own reading.

Thinking and language [§ Thinking and language]

Fedorenko’s language network is domain-specific: active for language comprehension and production (any modality, any known language, including constructed languages like Klingon) but silent for chess, spatial memory, music, arithmetic, and programming.
Patients with global aphasia (language network destroyed by stroke) retain the ability to play chess, drive, and do arithmetic — evidence for Gibson’s claim that language is a communication system for meaning, not the substrate of thought.

LLMs [§ LLMs]

Gibson calls LLMs arguably the best current theory of a language’s form, while denying they are a good theory in the explanatory sense, because their ‘theory’ is an opaque set of billions of parameters rather than a compact, checkable grammar.
LLMs reliably fail meaning-dependent reasoning (a rephrased Monty Hall problem with a stated 100%-certain answer) by pattern-matching surface form, an error humans essentially never make.
LLMs struggle to complete center-embedded sentences in the same pattern as humans, without having been trained on labelled bad sentences — suggesting they have absorbed something about human sentence-processing difficulty as a side effect of form-only training.

Culture and language [§ Culture and language]

The Piraha (Amazonian Brazil) have no words for exact counting, not even ‘one’; what looked like a one/two/many system turned out, on controlled testing, to be three context-dependent approximate quantifiers.
Piraha speakers match visible sets of objects perfectly (no counting required) but default to rough estimation once the objects are hidden and exact counting would be needed.
Colour vocabulary tracks communicative need, not perceptual capacity: English settles on about eleven basic colour terms; the Dani of Papua New Guinea use only two, despite normal colour vision, because their community rarely needs to distinguish manufactured objects by arbitrary colour.