nucleation

CLAIM-082: A term's value, as distinct from its signification, is fixed by its contrasts with the coexisting terms of its system

5 min read

Properties15
StatusProposed
Version3
Rolethesis
Warrantinterpretive
Tagslinguistics, philosophy-of-language
Linepragmatic-transport
Workswhat-survives-translation
GroundsTHEORY-159 — In a language a term's value is fixed by its relations to the coexisting terms of the same system, so terms that share a signification can differ in value
SupportsCLAIM-115 — What survives translation is the set of communicative distinctions and relations a receiving system can still reconstruct and act upon, not the wording or the proposition, CLAIM-036 — Communicative meaning is constituted through systems of distinctions and relations, and translation transports these systems, selectively preserving or transforming their differential, probabilistic and interactional structure
ComplementsCLAIM-242 — The use-level counterpart of Saussure's value is an utterance's contrast with the alternatives its speaker could have used in context; values do not cross between languages, which is why fidelity is graded transport through an anchored correspondence rather than identity of value, CLAIM-489 — A sign's communicative significance is fixed by its contrasts within a system rather than by correspondence to a referent
Corroborated byTHEORY-175 — Cosine similarities to matched anchor points are unchanged by exactly the maps that preserve each sample's cosines to the anchors, among them every orthogonal map times a positive scalar and every per-sample positive rescaling but not translations; in Moschella et al.'s experiments independently trained encoders agreed in them only approximately, enough for a decoder reading them to transfer untrained between seeds, architectures and languages with loss, while their nearest neighbours mostly differed, THEORY-182 — A rank-limited contrastive word embedding trained from small initialisation learns the top eigenvectors of the co-occurrence matrix's relative deviation from independence, one at a time in order of eigenvalue, not the best low-rank approximation of its unconstrained optimum, THEORY-185 — When each binary attribute of a word affects its co-occurrence independently and multiplicatively, the PMI is affine in the attributes with rank at most d + 1, so a spectral PMI embedding is a linear image of the attribute hypercube and parallelogram analogies hold exactly; over the full hypercube of words, each attribute's two values equally frequent, the raw co-occurrence ratio mixes in products of attributes, and keeps exact analogies only when the strengths are alike enough that every attribute mode outranks every product mode
Illustrated byCASE-019 — "Hug from Doug"
Cited as “concludes” byARG-034 — Saussure's account of value grounds the premise that significance is fixed by contrasts rather than by correspondence to a referent
Cited as “rests on” byARG-033 — Saussure's value, as the record reads it, supports transport of relational identity across languages
Rendered intodocs/arguments-by-conclusion.md

Graph View

The claim

§2: “Saussure’s account of linguistic value makes the significance of a sign depend upon contrasts within a system, rather than on intrinsic correspondence to a referent.” Lévi-Strauss is cited for extending relational and transformational analysis to myth. The manuscript uses the premise to motivate relational identity, with the proviso that not every transformation is a symmetry.

What it does not say

It is a motivating premise, not a result the manuscript relies on formally. Lévi-Strauss is unread here (LIT-775), so his part of it rests on a work cited without a reading.

A qualification from experiment

Bergen, Goodman and Levy (LIT-809), Experiment 1: a novel symbol’s interpretation is fixed by its contrast with the alternative the speaker could have used. But the contrast works only because one alternative is tied to an object by resemblance. Read against this claim, it is contrast together with reference, not contrast instead of it.

A qualification from machine representations

Moschella and colleagues’ relative representations (LIT-849, read in NOTE-652) are a working case of the premise in one respect. Each input is represented only by its cosine similarities to a set of anchor inputs from the same system, and a decoder trained on those similarities transfers, untrained, between independently trained encoders (THEORY-175). Identity by relations within a system is enough to carry a decoder across systems that share no coordinates.

The reading adds two qualifications, and both point the same way as the one from experiment above.

  • Across systems, contrast needs a correspondence. Between languages or modalities, the anchors must be matched from outside: translated reviews, sentences aligned in WikiMatrix. Within one system relations suffice; for transport between systems they work only once something ties the anchor in one to the anchor in the other. That is contrast together with correspondence, as in Bergen, Goodman and Levy.
  • The relational encoding is lossy and approximate. Independently trained encoders agree in their relative vectors only approximately: the vectors correlate highly, but most nearest neighbours differ. Where a shared absolute space already exists (a multilingual encoder), the relational encoding transfers worse than the absolute one in every language pair tested.

The analogy has a limit of its own. A cosine to an anchor is a metric relation in a vector space, invariant only to cosine-preserving maps. Saussure’s value is opposition among coexisting terms, and THEORY-159 does not make it a similarity. The case shows that relational identity can be engineered and carried between systems. It does not show that this is Saussure’s value.

A model whose meanings are purely relational

Karkada, Simon, Bahri and DeWeese (LIT-855, read in NOTE-660) solve a close proxy of word2vec in closed form (THEORY-182). The learned embeddings are the top eigenvectors of one matrix: the relative deviation of each word pair’s co-occurrence from independence. Nothing else enters, and the loss sees only inner products between embeddings, so each embedding is fixed only up to a common rotation. A word’s learned representation is therefore nothing but its pattern of departures from chance co-occurrence with every other word, and the analogy structure the model shows comes out of that.

That is the premise realised in a model, in one respect: identity by relations within a system, with no reference anywhere in the training signal. Its limits match the qualifications above. Co-occurrence is not Saussure’s opposition either (THEORY-159). The result is proved for a proxy of word2vec, not word2vec. And it is a fact about what a model learns from text, not evidence about how significance is fixed in a language.

Its sequel (Karkada et al. 2026, LIT-860, read in NOTE-664) adds a sharper case and a qualification. The months are still placed on a circle from their co-occurrence with other words after their co-occurrences with each other are deleted (Fig. 4), so each one’s position is fixed by its relations within the vocabulary. But those relations are organised by an extralinguistic variable, the time of year, and the geometry learned is isomorphic to it. Relations within the system carry the structure of what the words are about. That supports “fixed within a system” without setting relation against reference, the same qualification as Bergen, Goodman and Levy’s.

The precursor, Korchinski, Karkada, Bahri and Wyart (LIT-863, read in NOTE-665), is the closest formal case so far. Each word is modelled as nothing but a set of binary contrasts, and when each contrast multiplies co-occurrence independently, the PMI embedding is an affine image of the word’s position among them, so the analogy parallelograms come out exactly (THEORY-185). That is opposition, not similarity, which is nearer Saussure’s value than the cosine and co-occurrence cases above. A contrast also survives deleting every word pair that differs only in it, so each word’s relations to the rest of the vocabulary carry it. The same qualification holds: the contrasts are posited in advance, and many are properties of what the words refer to.

Note of 2026-10-28: narrowed; what holds it back

The sixth status review narrowed this claim to what Saussure’s value carries, recording CLAIM-072’s concession, which ARG-034 judged. The first title is CLAIM-489. Kept Proposed: its reading of Saussure, THEORY-159, is Proposed, and nothing in the record meets defeated_if.

Note of 2026-10-29: what holds it back

Held by THEORY-159’s promote_when: no study in the record adds, loses or merges a term in a language’s inventory and measures whether its neighbours’ use shifts, against terms outside the field, and that is also the second clause of this claim’s defeat condition.

Note of 2026-10-30: what holds it back

THEORY-159 is a premise here, not corroboration or an attribution (ADR-058). The title asserts Saussure’s account, and the second clause of defeated_if is THEORY-159’s promote_when. A “Saussure proposes” split of THEORY-159 would not free it, since the claim needs the account true, not only proposed. It waits on a study that adds, loses or merges a term in a language’s inventory and measures the neighbouring terms’ use against terms outside the field.

Backlinks