J.R. Firth

The Problem of Meaning Before Distribution

For much of the early twentieth century, linguistics found itself caught between two unsatisfying poles when it came to meaning. On one side stood the mentalists — scholars who located meaning inside the head, treating words as labels affixed to pre-existing concepts, and semantic analysis as an exercise in introspection. On the other stood the behaviorists, led by Bloomfield, who were so suspicious of anything unobservable that they essentially declared meaning off-limits to serious linguistics, a problem too messy and subjective to handle scientifically. Between the Scylla of fuzzy mentalism and the Charybdis of principled avoidance, there wasn’t much room to actually do semantics.

John Rupert Firth — Professor of General Linguistics at the University of London from 1944 to 1956, the first person to hold such a chair in Britain — walked directly into this gap. His project was to make meaning tractable without reducing it to either private mental images or bare stimulus-response chains. The solution he proposed was deceptively simple in its formulation and genuinely radical in its implications: meaning is function in context. Or, in the line that would outlive everything else he wrote: “You shall know a word by the company it keeps.”

Collocation, Context, and the Levels of Meaning

Firth’s approach was polysystemic — he insisted that linguistic analysis should operate across multiple, semi-independent levels (phonology, grammar, lexis, semantics, situation) rather than trying to derive a single unified description. This already put him at odds with the dominant structuralist paradigms of his time, both Saussurean and Bloomfieldian, which tended to seek economy in their formalisms. Firth was comfortable with a messier architecture because he believed language itself was messy in precisely this way.

His most enduring technical contribution was the theory of collocation — the idea that part of a word’s meaning is constituted by the habitual company it keeps with other words. “Dark” means something slightly different in “dark night” than in “dark horse” or “dark secret,” and these differences aren’t random or merely pragmatic; they’re systematic patterns that can be studied empirically. The collocational profile of a word is a dimension of its meaning, not just a clue to some deeper semantic essence hiding underneath.

This was paired with his broader framework of context of situation, borrowed and reworked from his teacher Bronisław Malinowski. For Malinowski, studying the Trobriand Islanders, language could not be understood apart from the activities, social relations, and material circumstances in which it was embedded. Firth generalized this insight: every utterance occurs within a context of situation that includes the participants, the relevant objects and events, the effect of the verbal action, and so on. Meaning at the situational level is about what the language is doing — its function in the lived world. This makes Firth a precursor to pragmatics, speech act theory, and systemic functional linguistics, even though he articulated his ideas before Austin published How to Do Things with Words and well before Halliday, his most famous student, built his own comprehensive framework.

What I find most striking about Firth’s architecture is its refusal of reductionism. He didn’t think collocational meaning could be collapsed into situational meaning, or that phonological patterns could be fully derived from grammatical rules. Each level had its own patterns, its own systems. This pluralism was intellectually honest — it reflected the genuine complexity of how language works — but it also made Firthian linguistics harder to formalize and teach, which partly explains why it never achieved the institutional dominance of Chomskyan generativism or even Bloomfieldian structuralism.

The Line That Built an Industry

The afterlife of Firth’s ideas in computational linguistics and natural language processing is extraordinary and deeply ironic. When word2vec landed in 2013, and the broader family of distributional semantic models exploded into mainstream AI research, the foundational assumption was exactly Firthian: you can learn what a word means by examining the statistical patterns of its co-occurrence with other words. The entire architecture of modern word embeddings — from GloVe to BERT to the transformer-based language models that now generate fluent text — rests on a distributional hypothesis that Firth articulated decades before anyone had the computational power to operationalize it at scale.

Firth himself would likely have found this both gratifying and reductive. His conception of meaning was never only distributional in the narrow word-co-occurrence sense. He layered collocation within a much richer theory involving phonological prosody, grammatical patterning, and the full context of situation. Modern embedding models capture co-occurrence statistics with breathtaking thoroughness, but they flatten everything onto a single representational plane — a dense vector — that erases precisely the polysystemic architecture Firth thought essential. The company a word keeps in a 300-dimensional vector space is a powerful abstraction, but it’s not the same thing as knowing the social situation, the participants’ roles, the prosodic contour of the utterance. We’ve industrialized one stratum of Firth’s theory and largely ignored the rest.

This is why I think the “grounding problem” in contemporary AI — the question of whether language models actually understand anything, or merely manipulate distributional patterns — is, at its core, a Firthian question. Firth would say that collocation alone is insufficient. Meaning also lives in the context of situation, in the relationship between language and the non-linguistic world. A model trained only on text has access to only one of his levels. Whether that level is sufficient for something we’d want to call understanding is an open and genuinely unresolved question.

The Neo-Firthian Tradition and What Remains

Firth’s direct intellectual legacy runs through Halliday’s systemic functional linguistics, through John Sinclair’s corpus linguistics at Birmingham (Sinclair took collocation analysis and ran with it, building COBUILD and demonstrating that vast tracts of language are formulaic and phraseological in ways traditional grammar couldn’t capture), and through the broader British tradition of applied linguistics. It’s a quieter legacy than Chomsky’s — no revolution, no dramatic paradigm shift — but it’s arguably more empirically productive, especially in lexicography, language teaching, and computational text analysis.

What remains unresolved is whether Firth’s polysystemic vision can be made formally rigorous enough to compete with more unified frameworks. His writings are notoriously allusive, aphoristic, and sometimes frustratingly imprecise. He had the instincts of a great empiricist and the prose style of an essayist, which left his followers considerable interpretive work. The question of how exactly the multiple levels of meaning interact — whether they can be given a coherent formal architecture or are irreducibly pluralistic — is still open.

Why This Matters

Firth matters because he identified, with remarkable clarity, a principle that turned out to be computationally explosive: that distributional patterns carry semantic information. But he also matters because he saw the limits of that principle before anyone had the chance to test it. The current moment in AI — where we have extraordinarily powerful distributional models that nonetheless fail in ways that seem to involve exactly the situational grounding Firth insisted on — is a moment that vindicates both his insight and his caution. He gave us the engine, and then warned us it wouldn’t be enough. We built the engine. The warning still stands.