PaperlayerView exact revision on arXiv

Source-linked AI summary

Language Models Model Language

Łukasz Borchmann

arXiv:2510.12766v1cs.CL

TL;DR

The paper challenges critiques that assess LLMs using criteria drawn from contestable linguistic theories. It proposes a Mańczakian reorientation and concludes that LLMs model language, though imperfectly, rather than failing to model it.

  • Problem

    The authors argue that linguistic scholarship has lacked clear validation criteria and that critiques of LLMs apply standards derived from contestable theories.

    Source: §1 · p. 1

  • Method

    The paper reorients discussion of LLMs through Witold Mańczak’s critique of abstract, dualistic linguistics.

    Source: §1 · p. 1

  • Results

    The authors conclude that LLMs are imperfect tools not because they fail to model language, but because they model only language.

    Source: §3 · p. 6

  • Takeaways & Limitations

    The authors recommend designing, evaluating, and deploying systems based on demonstrated mastery of language’s relational logic.

    Source: §3 · p. 6

    Claims that do not meet Mańczakian conditions may be valuable in other fields but do not count as linguistic in that framework.

    Source: §4 · p. 7

Abstract

from arXiv · show

Linguistic commentary on LLMs, heavily influenced by the theoretical frameworks of de Saussure and Chomsky, is often speculative and unproductive. Critics challenge whether LLMs can legitimately model language, citing the need for "deep structure" or "grounding" to achieve an idealized linguistic "competence." We argue for a radical shift in perspective towards the empiricist principles of Witold Ma\'nczak, a prominent general and historical linguist. He defines language not as a "system of signs" or a "computational system of the brain" but as the totality of all that is said and written. Above all, he identifies frequency of use of particular language elements as language's primary governing principle. Using his framework, we challenge prior critiques of LLMs and provide a constructive guide for designing, evaluating, and interpreting language models.

1 Introduction

The authors argue that LLM criticism often applies contestable theoretical criteria rather than objectively assessing models’ capabilities. They propose Mańczak’s empiricist framework as an alternative foundation.

1 Introduction

  • The authors argue that centuries of linguistic scholarship lacked clear validation criteria, leaving the field unprepared for LLMs’ empirical success.They describe current objections as revealing more about linguistic dogma than LLM capabilities.Source: §1 · p. 1
  • Critiques invoking deep structure, correctness versus likelihood, or access to meaning apply criteria derived from specific, contestable theories.The authors characterize these critiques as defenses of a challenged linguistic paradigm, rather than objective assessments of LLM capacities.Source: §1 · p. 1
  • The authors propose turning to Witold Mańczak’s critique of abstract, dualistic linguistics as an alternative foundation for studying LLMs.They contrast their proposed science of language with calls to constrain LLMs as “stochastic parrots.”Source: §1 · p. 1

2 The Ma´nczakian Framework

The section presents Mańczak’s text-first, empiricist framework and uses it to reinterpret debates about grammar, synthesis, acquisition, and meaning. It argues that LLMs’ language modeling reflects frequency patterns and relational structure in text.

2 The Mańczakian Framework

  • Mańczak sought to make linguistics an inductive, quantitative science, requiring claims to be verifiable through statistics or experiments and explained by observable phenomena.His corpus-based work identified frequency of use as a primary force shaping language.Source: §2 · p. 1

2.1 The Map Is Not the Territory

  • Mańczak treated grammar and lexicon, and rules and exceptions, as frequency-based continuums rather than qualitative oppositions.Grammar covers frequent patterns across word classes, while lexicography handles individual words; high-frequency patterns count as rules and low-frequency patterns as exceptions.Source: §2.1 · p. 2

2.2 Talkin’ All About That Syntax When You Can’t Even Generate a Haiku

  • Mańczak argued that analyses should identify only components needed to synthesize texts, and treated synthesis as a test of analysis.For “Sincerity may frighten the boy,” he contrasted a lengthy generative analysis with a reconstruction using five practical rules.Source: §2.2 · pp. 2, 4
  • Examples from Latin-to-Romance change illustrate frequency-linked regularization, word-form divergence, and grammaticalization.Frequent use of habere in a grammatical role accompanied a shift from its lexical meaning, “possess” or “hold,” toward a tense marker.Source: §2.2 · p. 3
  • LLM pretraining fits Mańczak’s thesis that language is text and frequency is its primary organizing force.Pretraining minimizes expected next-token surprisal, and larger training sets sharpen estimates of language frequency structure, especially in the long tail.Source: §2.2 · p. 4

2.3 An Obituary for the Language Organ

  • The section connects usage-based acquisition to children learning grammar through general cognitive tools applied to heard language patterns.It also links LLMs’ plausible generation to learned frequency patterns and analogical relationships, which it identifies as central to linguistic generalization.Source: §2.3 · pp. 4–5

2.4 This Page is Intentionally Left Ungrounded

  • The authors argue that evaluating LLMs should focus on mastery of the textual world’s internal relational logic, not access to the outside world.They describe most word meanings as relational and use “justice” and its links to fairness, law, equality, and crime as an example.Source: §2.4 · pp. 5–6

3 Conclusion

The conclusion rejects deep-structure and grounded-meaning standards as prerequisites for language, arguing that LLMs are imperfect because they model only language. It recommends designing and evaluating systems for mastery of language’s relational logic, in line with Mańczak’s account of language as texts organized by frequency.

3 Conclusion

  • LLM critics attribute intrinsic flaws to gaps in competence and understanding, treating deep structure or grounded meaning as requirements of language.Users, meanwhile, may be amazed by human-like text and disappointed by confident factual errors.Source: §3 · p. 6
  • The authors argue that LLMs are imperfect not because they fail to model language, but because they model only language.They say the models reveal what language has been all along.Source: §3 · p. 6
  • The proposed practical path is to design, evaluate, and deploy systems that have proven mastery of language’s relational logic, rather than satisfy abstract competence requirements.The authors say those abstract requirements will neither improve linguistic competence nor bring us closer to Artificial General Intelligence.Source: §3 · p. 6
  • Mańczak’s account treats language as the totality of texts, with frequency as its organizing principle; LLMs are built on this foundation and increasingly affect people’s futures.The passage names drafting contracts and structuring arguments as examples of their impact.Source: §3 · p. 6

4 Limitations

The authors describe remaining boundaries in theory and practice, including corpus representativeness, the scope of text-based linguistics, and priorities for evaluating LLMs.

4 Limitations

  • The framework leaves some critical questions unaddressed, and the authors position defining the technology as preparation for—not a substitute for—ethical debate.They state that a text-only definition excludes some subfields and that ethical dilemmas remain to be addressed.Source: §4 · p. 6
  • A text-only definition excludes some subfields, but claims qualify as linguistic when they make testable predictions about text distributions or observable utterances.The authors say pragmatics, sociolinguistics, and ethnolinguistics can qualify when their predictions meet these criteria.Source: §4 · pp. 6–7
  • They do not claim a complete theory of meaning, arguing instead that meaning is usually inferable from textual relations and direct grounding may play a limited role.They acknowledge the concern that this emphasis could appear purely structuralist.Source: §4 · p. 7
  • LLMs’ training corpora are unrepresentative, so the authors characterize them as practically incomplete and propose frequency-weighted text selection reflecting actual language use.Their proposed selection favors texts according to circulation and influence.Source: §4 · p. 7
  • The authors do not reject comparisons with human cognition, but place them after establishing textual-corpus evaluation as the primary benchmark.They describe the Mańczakian framework as a necessary baseline for that comparison.Source: §4 · p. 7
  • The authors connect Mańczak’s text-centered empirical approach to usage-based linguistics and argue that it aligns with how text-trained models function.They also argue that frequency-weighted patterns can serve as templates for novel combinations, making creativity compatible with pattern mastery.Source: §4 · p. 7

A Selected Contributions to Historical and Comparative Linguistics

The passage introduces a French summary of Mańczak’s work and frames his career as a sustained challenge to methodological flaws in modern linguistics.

A Selected Contributions to Historical and Comparative Linguistics

English Translation

The section argues that linguistics has left its criteria for distinguishing truth from falsehood unexamined and often defers to authority. It presents statistics, and occasionally experiments, as ways to test linguistic claims, including claims about frequency and historical change.

English Translation

  • Linguistics has not addressed how to distinguish truth from falsehood, while linguists are said to treat opinions as true or false according to the speaker’s authority.The author calls the question taboo and describes authority rather than verification as the operative criterion.Source: English Translation · p. 8
  • The author proposes statistics and, exceptionally, experiments as criteria for testing linguistic opinions, many of which statistical data invalidate.The examples are presented as cases where faith in authorities gives way to statistical evidence.Source: English Translation · p. 8
  • Mańczak attributes word forms to regular phonetic change, analogy, and irregular frequency-driven change; one cited assessment puts the chance his theory is erroneous below 1 in 10 million.The probability estimate is attributed to a professor of applied mathematics.Source: English Translation · p. 8
  • Statistical data challenge Bartoli’s norm and laryngeal theory, while Mańczak argues that language is characterized by asymmetry and greater differentiation of more-used elements.The frequency claim contrasts more-used with less-used linguistic elements.Source: English Translation · p. 8
  • Statistical data support revising Verner’s law, treating Old Church Slavonic as a compromise, and identifying the Indo-European homeland with that of the Slavs.These are presented as separate conclusions from Mańczak’s work.Source: English Translation · pp. 8–9
  • Probability calculus favors a single origin for Romance verbs meaning “to go,” while statistical data support derivation of Romance languages from Classical rather than Vulgar Latin.The “go” forms had about sixty proposed etymologies; the Latin-origin claim contrasts with the longstanding Vulgar Latin account.Source: English Translation · p. 9
  • Parallel Germanic texts place the Goths’ original homeland in southernmost ancient Germania, and language classification is justified by lexical rather than inflectional or phonetic convergences.The classification claim concerns divisions such as Germanic, Slavic, Baltic, and Romance.Source: English Translation · p. 9
  • Homonymy challenges the stressed/unstressed word distinction, while the proposed difference between common and proper nouns is that common nouns are usually translated and proper nouns almost never are.The examples contrast ville with Italian città and English town, and Paris with Italian Parigi and English Paris.Source: English Translation · p. 9

French Original

The French passage frames the criteria for distinguishing truth from falsehood as an unaddressed problem in linguistics. It reports that authority, rather than verification, functions as a criterion and points to statistics and experiments as alternatives.

French Original

  • The passage calls criteria for distinguishing truth from falsehood a fundamental but taboo question in linguistics.It notes that, despite linguistics’ long history and extensive publication record, no works are devoted to those criteria.Source: French Original · p. 10
  • The author describes linguists as judging opinions by whether their proponents are authorities, rather than by verifying the opinions themselves.Opinions attributed to authorities are treated as true, while those from non-authorities appear false.Source: French Original · p. 10
  • The author proposes statistics and, exceptionally, experiments as ways to assess linguistic opinions, noting that statistical data invalidate some authority-backed views.Source: French Original · p. 10

1. Dans toutes les langues, la forme des mots

This section presents Mańczak’s frequency-centered account of language change alongside statistical challenges to established linguistic claims. Its examples range from phonetic development and word forms to language origins, classification, and noun distinctions.

1. Dans toutes les langues, la forme des mots

  • Word forms depend on regular phonetic change, analogy, and irregular frequency-driven change; one cited assessment places the chance that the frequency theory is erroneous below 1 in 10 million.The probability estimate is attributed to a professor of applied mathematics.Source: §1 · p. 10
  • Statistical data invalidate Bartoli’s norm and laryngeal theory, while Mańczak argues that language is characterized by asymmetry and more-used elements being more differentiated.The frequency claim contrasts more-used with less-used linguistic elements.Source: §1 · p. 10
  • Mańczak argues that Verner’s law needs revision, Old Church Slavonic is a compromise between two speech varieties, and the Indo-European homeland matches that of the Slavs.Source: §1 · pp. 10–11
  • Indo-European language divisions are justified by lexical convergences, not inflectional or phonetic ones.The passage names Germanic, Slavic, Baltic, and Romance as examples of language divisions.Source: §1 · p. 11
  • Statistical data support deriving the Romance languages from Classical Latin rather than Vulgar Latin.Source: §1 · p. 11
  • Comparison of parallel Germanic texts places the Goths’ original homeland in southernmost ancient Germania, not Scandinavia.The comparison includes Gothic, several German varieties, Danish, and Swedish.Source: §1 · p. 11
  • The stressed/unstressed word distinction is challenged by homonymy, while common nouns are usually translated between languages and proper nouns almost never are.The passage notes that about ten definitions of proper nouns had been proposed, none applicable to all such nouns.Source: §1 · p. 11

Related papers

Loading 2510.12766v1…