PaperlayerView exact revision on arXiv

Source-linked AI summary

Translation Indeterminacy and the Distributional Fallacy

Michael Carl

arXiv:2609.07717v1cs.CL

TL;DR

The paper asks whether distributional patterns ground meaning, and whether translation instead can rely on cross-linguistic regularities. It develops an ecological-enactivist account of meaning and argues that stabilized inferential organization can support translation without access to determinate reference.

  • Problem

    The distributional fallacy conflates patterns that help detect or predict linguistic organization with what constitutes meaning, while the relation between meaning indeterminacy and translation remains at issue.

    Source: §1 · p. 1

  • Method

    The paper develops an ecological-enactivist account that grounds meaning in embodied agent-environment interaction and treats linguistic regularities as sedimented traces of those practices.

    Source: §1 · p. 2

  • Results

    The paper argues that stabilized inferential organization in language can support successful translation without reconstructing determinate referential content.

    Source: §1 · p. 2, §5 · p. 17

  • Takeaways & Limitations

    Predictive linguistic organization may suffice for translation even though text-trained LLMs lack the action-grounded basis from which human reference and meaning emerge.

    Source: §5 · p. 17

    A robotically embodied LLM might acquire causal grounding without full human reference, which also requires embodied relevance realization and participation in social practices.

    Source: §3.3 · p. 11

Abstract

from arXiv · show

Large language models (LLMs) are commonly associated with the distributional hypothesis, according to which (1) semantic meaning is grounded in distributional patterns of linguistic context, and (2) knowledge of cross-linguistic distributional correspondences allows for successful translation. This paper rejects the first claim as a causal inversion: linguistic distributions reflect patterns arising from meaning-making practices rather than constituting their source. At the same time, it accepts the second claim, arguing that translation -human or machine - can succeed without requiring access to meaning or reference. Knowledge of interlingual distributional correspondence and their inferential organization may be sufficient for translation. The paper develops an ecological-enactivist perspective, according to which reference and meaning are grounded in agent-environment interaction and stabilized through action-grounded concepts, forms of world-involving cognition that current LLMs do not possess.

1 Introduction

The section rejects the distributional fallacy: linguistic distributions can help predict linguistic organization, but they do not constitute meaning. It argues that translation can succeed through stabilized inferential regularities even when reference remains indeterminate.

1 Introduction

  • Harris treated distributional similarity as a diagnostic for meaning, not as a constitutive theory, and argued that linguistic structure need not parallel meaning one-to-one.Firth’s “You shall know a word by the company it keeps” can likewise function as a tool for predicting linguistic organization [Firth (1957)].Source: §1 · p. 1
  • The distributional fallacy mistakes a useful way to detect or predict linguistic organization for an account of what grounds meaning.The paper distinguishes an epistemological relation from a metaphysical one.Source: §1 · p. 1
  • Quine’s indeterminacy thesis holds that behavioral or distributional evidence cannot uniquely fix semantic content, so translation understood as preserving that content is indeterminate.The section frames the distributional fallacy as directly opposing this thesis.Source: §1 · p. 2
  • The ecological-enactivist view grounds meaning in embodied interaction and treats language as coordinating action and attention within historically stabilized practices.On this account, reference is an achievement of embodied coordination, not the starting point of linguistic practice.Source: §1 · p. 2
  • Distributional traces can preserve inferential organization that LLMs exploit for translation without reconstructing the determinate reference that originally gave rise to those traces.The paper recasts the distributional hypothesis as a prediction hypothesis: distributions provide access to predictive linguistic organization.Source: §1 · p. 2
  • Translation need not uniquely determine reference to reproduce functionally appropriate linguistic behavior.The paper says this weakens the implications of Quinean indeterminacy for translation and accords with observations by translation scholars.Source: §1 · p. 2

2 The Sedimentation of Meaning in Language

Language bears the traces of human meaning-making rather than serving as its source. LLMs can learn those traces, but the account distinguishes this from embodied competence and raises a separate question about whether their names refer rigidly.

2 The Sedimentation of Meaning in Language

  • LLMs learn linguistic regularities by predicting tokens from context, without the perceptual, embodied, or social interactions through which human concepts are acquired.A text-only model does not encounter the objects, events, or practices its language concerns.Source: §2 · pp. 2–3
  • LLMs recover the statistical imprint of human communication: linguistic structures resemble conceptual organization because generations of embodied agents produced those regularities.The paper characterizes a model as shadowing meaning’s imprint on language rather than meaning itself.Source: §2 · p. 3
  • Human conceptual competence develops through perception, social practices, and coordination of attention and action, making language a historical product of embodied interaction.The inferential organization LLMs capture is described as culturally sedimented rather than the origin of meaning.Source: §2 · p. 3

2.1 Rigid Designation

  • Whether an LLM’s name-use is a link in a causal communication chain or parasitic on reference already secured by training data remains contested, and the author takes no stand on that issue.The author locates the substantive disagreement elsewhere, in whether the name tracks an individual rigidly or a descriptive cluster.Source: §2.1 · p. 3
  • The paper argues that an LLM’s use of “Aristotle” behaves descriptively rather than rigidly when it tracks associated descriptions instead of the individual across counterfactual variation.The author distinguishes this test of rigidity from whether the token has causal ancestry.Source: §2.1 · p. 3

2.2 Descriptive Designation

The section argues that LLMs use names through associated descriptions rather than rigid designation. On this view, corpus history may connect tokens causally to referents without transmitting speakers’ referring intentions.

2.2 Descriptive Designation

  • Wheeler argues that language models extract word-sequence probabilities without processing texts as sentences with intended referents, and that design-intention secures meaning and truth-value for outputs.This account treats an LLM’s function as derived from its designers, deployers, and interpreters.Source: §2.2 · p. 4
  • On the paper’s conditional account, LLMs treat names as predicates—bundles of associated descriptions—rather than as rigid designators.The account is presented as close to Searle’s cluster-of-descriptions theory.Source: §2.2 · p. 4
  • LLM tokens may be causally connected to referents through training-corpus history without being connected through a chain of referring intentions.The paper describes any reference secured by a descriptive cluster as contingent and derivative via design-intention.Source: §2.2 · p. 4

2.3 Action-grounded Concepts

The section distinguishes embodied, action-grounded concepts from LLMs’ linguistic competence. LLMs can learn inferential relations useful for tasks such as translation, while lacking the grounding that calibrates concepts through physical success and failure.

2.3 Action-grounded Concepts

  • Action-grounded concepts are practical knowledge developed through perception and action, with predictions that can be corrected by experience.The squirrel’s branch-related behavior illustrates how failed action can falsify an action concept.Source: §2.3 · p. 4

2.4 Inferential Concepts

  • LLMs can articulate linguistic concepts without practical grasp; their concept of “branch” is a collocational profile learned from text produced by people with action-grounded concepts.The paper contrasts linguistic form of knowledge with its practical substance.Source: §2.4 · p. 4
  • LLMs can make valid inferences within collocational structures, a competence the paper suggests may count as inferential concepts rather than embodied knowledge.The example infers “Socrates is mortal” from “Socrates is a man” and “All men are mortal.”Source: §2.4 · p. 4
  • Inferential concepts support formal reasoning, linguistic manipulation, text pattern recognition, and translation, but are limited for novel physical reasoning, world modeling, and perception-action tasks.Unlike squirrel concepts, they are calibrated through statistical regularities of linguistic practice rather than physical success and failure.Source: §2.4 · p. 5
  • Transformer attention computes context-sensitive relations in an embedding space shaped by corpus regularities, supporting next-token prediction and linguistic inference without action-grounding.The latent space is described as encoding entailment and exclusion relations among expressions.Source: §2.4 · p. 5

2.5 Metaphoric Extension

Human inferential organization is rooted in bodily action, and metaphorical language carries traces of those mappings. LLMs can reproduce these traces without participating in the embodied interactions that shaped them.

2.5 Metaphoric Extension

  • Human inferential structure draws on bodily engagement, with abstract domains inheriting entailments from sensorimotor schemas through metaphorical extension.Temporal reasoning draws on spatial traversal, quantity on vertical schemas, and causal reasoning on physical force.Source: §2.5 · p. 5
  • Human action-grounded competence can function without inferential abstractions, while LLMs depend on the embodied host layer from which they derive structure and epistemic grip.The paper describes this dependence as inferential competence being parasitic on action-grounded competence.Source: §2.5 · p. 5
  • LLMs fluently reproduce metaphorical expressions inherited from grounded agents, but reproducing their linguistic traces is not participating in the embodied interactions that structured them.Examples include “rising prices,” “approaching deadlines,” and “forces driving change.”Source: §2.5 · p. 5

2.6 Referential Profiles

Rhee’s referential profiles describe public, norm-governed patterns through which expressions converge on entities. Humans realize these profiles through embodied and social capacities, whereas LLMs manipulate their linguistic traces without participating in the grounding loops.

2.6 Referential Profiles

  • Rhee’s profile account replaces naive world-hooking with a distributed, public pattern of use intended to supply reference stability without causal metaphysics.The passage contrasts descriptive-cluster accounts with Kripkean causal hookups and attributes the public, norm-governed proposal to Rhee (2026).Source: §2.6 · p. 5
  • Rhee’s referential profiles are structured patterns of descriptions, memories, perceptions, inferences, affective orientations, social uses, and correction practices through which expressions converge on entities or situations.Humans realize these profiles through perception, episodic memory, embodiment, affect, and social responsibility; LLMs do not.Source: §2.6 · p. 6
  • On Rhee’s account, profiles capture community histories of correct and incorrect continuations, with context-sensitive computation activating a situational profile from recurrent relations sedimented in language.The passage attributes this account to Rhee and describes LLMs as inheriting profiles from human practice in linguistic form.Source: §2.6 · p. 6
  • LLMs reproduce inferential and metaphorical patterns but do not participate in the perception–action loops that give those patterns normative grip on the world.The account describes human interlocutors as supplying grounding conditions while models manipulate linguistic traces.Source: §2.6 · p. 6

2.7 Affordances

The section treats affordances as action possibilities and control structures, then explains how historically stabilized linguistic practices can carry reference. Reference requires both occasion-bound anchoring and inferential identity-tracking.

2.7 Affordances

  • Affordances are possibilities for action arising from relations between the material environment and abilities available in a form of life; landscapes and fields distinguish available from situationally relevant possibilities.The passage attributes these distinctions to Rietveld & Kiverstein (2014) and Bruineberg & Rietveld (2014).Source: §2.7 · p. 6
  • A translator coordinates nested affordance landscapes across timescales, including typing, syntactic expectations, discourse anticipation, and terminology, rather than evaluating against stored representations.These action-readiness states are assessed as situated appreciation expressed in normative behavior.Source: §2.7 · p. 6
  • Affordances function as situated control constraints that channel behavior toward some trajectories and suppress others, rather than encoding detached descriptions of the world.Their normativity is gradual and context-sensitive, distinguishing better from worse or appropriate from inappropriate action.Source: §2.7 · p. 6
  • LLMs model linguistic traces of human affordance-responsiveness and re-enact historically sedimented linguistic behavior despite lacking the embodied abilities to engage with environments.The passage lists inferential expectations, discourse conventions, genre regularities, and pragmatic patterns among what their latent structure learns.Source: §2.7 · p. 7
  • Enlanguaged affordances are historically stabilized constraints on appropriate linguistic continuations, interpretations, and corrections within a discourse community.Kiverstein & Rietveld frame language use without internal, content-carrying representational states.Source: §2.7 · p. 7
  • Enlanguaged affordances carry reference when community practices let a use be identified across speakers, phrasings, and occasions, and allow correction and reaffirmation.The passage equates this pattern with Rhee’s norm-governed referential profile.Source: §2.7 · p. 7
  • On the control-structure account, reference consists in a community’s historically stabilized pattern of agreement, correction, and apt continuation, not a separate representation consulted about an object.The passage illustrates this with the convergent use of “Aristotle.”Source: §2.7 · p. 7
  • Reference requires both occasion-bound affordance engagement to anchor a pick and inferential identity-tracking to recognize sameness across paraphrases; neither layer suffices alone.The combined practice is stabilized and corrigible over time.Source: §2.7 · p. 7

Emergence of Representation

The account distinguishes action-grounded anticipatory stabilization from full representation. Full representation emerges when communities develop propositional practices that make commitments detachable and reusable.

Emergence of Representation

  • Recurrent interaction stabilizes anticipatory, action-relevant patterns that can be re-presented in perception, memory, or imagery beyond the immediate situation, but remain tied to individual engagement.Action-grounded concepts supply content that can later be abstracted and articulated; without them, there is no independently stabilized content for representation to redeploy.Source: Emergence of Representation · p. 8
  • Full representation requires propositions that create assertible and accountable slots, allowing their contents to be detached from an occasion and reused, negated, embedded, or quoted.What is detached is an assertoric commitment already discretized as true or false, not a raw affordance pattern.Source: Emergence of Representation · p. 8
  • This account reverses the conventional view: propositionality is constituted through enlanguaged affordances, and representational content emerges as socially stabilized patterns become assertible commitments.Source: Emergence of Representation · p. 8

2.8 Relevance Realization

The ecological-enactive account grounds meaning in embodied relevance realization and socially coordinated action. It distinguishes this grounding from the inferential organization that may support translation and from the linguistic regularities LLMs reproduce.

2.8 Relevance Realization

  • Relevance realization selects environmental features relevant to ongoing action, avoiding the computationally intractable task of representing every possibility.It dynamically selects affordances through embodied, goal-directed engagement and shapes the agent-environment landscape.Source: §2.8 · p. 8
  • Action-grounded concepts need not be inferential, so translation may depend on stable cross-context linguistic regularities without requiring propositional thought.A dog may recognize and interact with a caregiver without articulating the human inferential network around caregiving.Source: §2.8 · p. 8
  • Meaning is stabilized through embodied relevance realization and socially coordinated linguistic practice, a grounding process in which LLMs do not participate.LLMs can reproduce aspects of linguistic inferential organization from recorded traces without engaging in the activities that give those practices meaning.Source: §2.8 · p. 9
  • LLMs can reproduce entailments and exploit linguistic regularities, but propositions do not function for them as embodied, action-guiding commitments.Their language may approximate aspects of truth through patterns sedimented in training data.Source: §2.8 · p. 9

3 Possible Objections

The paper argues that conversation, human-feedback training, and robotic embodiment do not by themselves establish reference or intentionality. It distinguishes linguistic or causal coupling from ongoing, norm-governed participation in the world and shared practices.

3.1 Conversational Interaction

  • Conversational grounding remains the human's work: the model's fixed parameters process dialogue without acquiring grounded concepts or retaining learning across turns.The paper notes that some systems may have persistent memory, retrieval, tools, or multimodal input.Source: §3.1 · p. 9
  • LLM fluency as an interlocutor reflects reproduction of grounded thinkers' linguistic patterns, not sharing the experiences that generated them.A human may learn from the exchange while the model merely processes it.Source: §3.1 · p. 9

3.2 RLHF and the Illusion of Grounding

  • Direction-locked answers in the Reversal Curse challenge the claim that accurate responses show a model represents relations compositionally.The paper suggests transformer feedforward layers may support key-cued retrieval that produces direction-locked results.Source: §3.2 · p. 10
  • RLHF rewards human judgments of text rather than independently verified worldly outcomes, transmitting scalar preferences instead of the raters' grounding.Factuality, helpfulness, and harmlessness are evaluated through raters' judgments or imagined outcomes.Source: §3.2 · p. 10
  • Closed-loop signals from verifiable mathematics, executable code, or tool errors may support world-answerable content, but do not give LLMs action-grounded concepts.The paper contrasts these real-outcome signals with feedback loops that close through human judgment.Source: §3.2 · p. 11

3.3 Grounding LLMs in a Robot

  • Robotic embodiment can provide direct sensorimotor coupling and causal grounding, yet that alone does not establish reference.Reference also requires functional and normative organization that closes the loop with the world itself.Source: §3.3 · p. 11
  • A robot's sensorimotor coupling can create causal links to worldly events, but full reference requires ongoing world-answerability, functional norms, and social practices.The account requires online or continual learning from real-world consequences; adding perception and action alone is insufficient.Source: §3.3 · p. 11

4 How do Humans and LLMs translate

LLMs can translate without perceptual grounding or uniquely determined reference by preserving cross-linguistic inferential organization. The paper explains this through linguistic regularities sedimented from embodied human practices, while distinguishing that inheritance from human grounding.

4 How do Humans and LLMs translate

  • Quine’s thought experiment shows that behavioral evidence permits incompatible translation manuals, yet LLMs translate despite having only text and no perceptual access.This raises the question of how translation can succeed if evidence does not uniquely fix reference.Source: §4 · pp. 11–12

4.1 Geometry of Embedding Spaces

  • Multilingual alignment can arise without explicit sentence-paired training, through incidental cross-language document co-occurrence and shared token, syntactic, and structural patterns.The resulting embedding alignment can support translation under some conditions but fails especially for typologically distant and low-resource languages.Source: §4.1 · p. 12

4.2 Where implicit Training breaks down

  • Implicit alignment is weaker for low-resource languages and may miss language-specific concepts, grammatical features, and pragmatic distinctions.Translation may drop or approximate features without an anchor in the target language; register choices also require social knowledge statistical models only partly capture.Source: §4.2 · pp. 12–13

4.3 Inferential Concepts as the Basis of Translation

  • Successful translation can preserve inferential relations across languages without recovering source reference or determinate semantic content.LLMs model patterns of entailment, association, and substitutability while lacking the concepts' referential content.Source: §4.3 · pp. 13–14

4.4 The representational Bias in Translation Studies

  • Translation theories often posit an intermediate representation, but practical adequacy can rely on situated correspondences rather than one uniquely privileged semantic representation.Theories also assess translation through functional profiles and communicative purposes.Source: §4.4 · p. 14

4.5 Representationalism in Cognitive Translation Studies

  • Cognitive Translation Studies often explain translational choices through mental representations, while enactivist accounts recast translation as coordination in situated activity.The ABC framework treats cognition as emerging from action, relevance realization, and environmental interaction.Source: §4.5 · p. 15

4.6 The Ecological-enactive Perspective

  • An ecological-enactive account treats translational norms and functions as enlanguaged affordances that solicit skilled action within socio-material and sociolinguistic settings.Translation reproduces appropriate patterns of engagement under new cultural and linguistic constraints rather than transferring representations.Source: §4.6 · pp. 15–16

4.7 LLM Translation beyond Representationalism

  • LLMs can translate by identifying functionally corresponding positions in historically evolved affordance structures, without recovering a unique semantic representation.Their translation depends on inferential and pragmatic roles, not on referentially anchored content or fully articulated semantic representations.Source: §4.7 · pp. 16–17

5 Conclusion

The conclusion argues that predictive linguistic organization can support translation without constituting meaning or fixing reference. It grounds that organization in historically sedimented embodied practices, while noting that language can also reshape human experience and action.

5 Conclusion

  • Predictive linguistic organization may suffice for translation because corpora preserve public consequences of meaning-making without giving LLMs its embodied, action-grounded basis.The paper argues that translation success is not evidence that linguistic distribution constitutes meaning.Source: §5 · p. 17
  • The weaker distributional hypothesis reconciles Quinean indeterminacy with translation by allowing linguistic regularities to constrain translations without fixing reference or constituting semantic content.The regularities capture sedimented consequences of meaning-making rather than its ground.Source: §5 · p. 17
  • The ecological-enactive account grounds meaning in embodied interaction and explains translation as operating within historically accumulated inferential organization rather than reconstructing presumed mental representations.Human communities sediment embodied regularities into inferential relations; organisms develop action-grounded concepts through relevance realization.Source: §5 · p. 17
  • LLMs exploit inferential organization encoded in language despite lacking action-grounded concepts, perceptual experience, and referentially anchored propositional content.They inherit products of human relevance realization without reproducing the processes that generated them.Source: §5 · p. 17
  • Table 3 contrasts human translators, who may traverse cognition from embodied interaction to linguistic production, with LLMs, which enter after organization is publicly available in language.The table describes humans as inheriting enlanguaged affordances through enculturation and LLMs as inheriting statistical traces of historically sedimented linguistic practice.Source: §5 · p. 18, Table 3 · p. 18
  • Although Table 3 presents linguistic and inferential structure as downstream from embodied interaction, human concepts may also develop in reverse, with words and normative uses reorganizing experience and action.The passage gives rights, obligations, contracts, and social statuses as cases where linguistic practices may constitute institutional reality.Source: §5 · p. 18

HUMAN TRANSLATION LLM TRANSLATION

The account resolves translation indeterminacy and the distributional fallacy by treating linguistic regularities as products of embodied relevance realization, not as the source of meaning. Translation can succeed by preserving sufficiently similar inferential and pragmatic organization across languages without reconstructing the embodied processes that produced it.

HUMAN TRANSLATION LLM TRANSLATION

  • Translation can preserve sufficiently similar inferential and pragmatic organization across languages without uniquely determining reference or semantic ontology.This is the account's proposed resolution of the tension between Quinean indeterminacy and practical translation success.Source: HUMAN TRANSLATION LLM TRANSLATION · p. 18
  • Linguistic structures preserve public consequences of grounded cognition, supplying enough information for many translation and other language tasks without reconstructing their embodied origins.The account does not claim that corpora reproduce the embodied processes from which those structures emerged.Source: HUMAN TRANSLATION LLM TRANSLATION · p. 19
  • Enlanguaged affordances are publicly stabilized possibilities for thought, discourse, perception, and action, rather than networks of internal descriptions.This reframes Searle's descriptive clusters as resources embedded in linguistic practices.Source: HUMAN TRANSLATION LLM TRANSLATION · p. 19
Loading 2609.07717v1…