linguistics.
42 of 1538 NOTE documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| NOTE-029 v2 | Krauska & Lau 2023, moving away from lexicalism | The authors argue that lemma-based production models encode two lexicalist assumptions: a morphology/syntax split, and a stored meaning–syntax–form “triad”. On that basis the models mishandle Inuktitut polysynthesis, separable Vietnamese idiomatic collocations, Hiaki number suppletion and English verb–object idioms. Updating the representations (treelets) is not enough, because the algorithms — separate lexical retrieval and structure building, word-sized increments — also have to change. They sketch an 8-stage non-lexicalist model and derive qualitative aphasia predictions from it. There are no new data and no implementation. | Read |
| NOTE-065 v2 | Bruening 2018, lexicalism: wrong and superfluous | Bruening argues that the lexicalist hypothesis is wrong on three counts. Phrases feed word formation productively (phrasal compounds, zero-derivation from phrases, German resultative nominalizations), ellipsis and focus target both phrases and morphemes, and the claimed principled differences between the word system and the phrase system (locality, adjuncts, headedness, idiosyncrasy) do not hold. It is also superfluous: every restriction it was meant to explain follows from constraints stated over X0 vs XP and from the selectional properties of individual affixes, none of which mention “word”. So one combinatorial component should build words and phrases. | Read |
| NOTE-073 v2 | Haspelmath 2011, the indeterminacy of word segmentation | Haspelmath tests ten morphosyntactic criteria for wordhood and finds that none is necessary and sufficient. No combination picks out what linguists write between spaces, and published test batteries pick criteria opportunistically (Table 1: nine studies, no two with the same battery). Hence (41, p. 65): “Linguists have no good basis for identifying words across languages, and hence no good basis for a general distinction between syntax and morphology”. A fuzzy “word” would rescue the distinction only if units on the boundness continuum were shown to cluster, and that is untested (§6). | Read |
| NOTE-179 | Brandom, inferentialism and meaning entitlement | Boghossian’s objection to substantive, entitlement-based theories of meaning (drawn from pejoratives like “Boche”) works against Dummett but not against Brandom. Brandom’s material inferences are non-monotonic and ceteris paribus, so one can grasp a concept while refusing to endorse the inferences it licenses. | Skimmed |
| NOTE-235 v2 | Dzhafarov, Zhang & Kujala, contextuality in behaviour | Under Contextuality-by-Default, which separates context-dependent marginals (“inconsistent connectedness”) from contextuality, no behavioural or social data set re-analysed is contextual: 73 poll question-order pairs, Schröder staircases, conjoint choices, 23 primed word combinations and psychophysical matching. The quantum question model’s QQ equality implies noncontextuality, so a quantum model here predicts its absence. The worked Rose–Jackson example mislabels two marginals; its conclusion stands. | Read |
| NOTE-243 v2 | Identifying Quantum Structure in AI Language | Two findings are reported. First, prompted as test subjects on Aerts & Sozzo’s “The Animal Acts” concept-combination test, ChatGPT and Gemini give CHSH values above 2: 4 in a guided conversation, and 2.25 and 3.09 (ChatGPT) and 2.12 and 4 (Gemini) over 81 prompted repetitions, against 2.42 for the 81 humans in 2011. Second, the rank–frequency curve of one 2,861-word Gemini story fits a Bose–Einstein form with energy E_i=(i−1)^0.8 far better than a Maxwell–Boltzmann form. The authors take these as showing that LLM “code is structurally quantum” (§5). The CHSH analysis never checks the strongly context-dependent marginals in its own tables. One reported expectation value disagrees with its table. The Bose–Einstein fit is numerically a Zipf–Mandelbrot law. | Read |
| NOTE-245 | Mathematical Foundations for a Compositional Distributional Model of Meaning | The paper builds DisCoCat as a product category FVect × P of vector spaces and a free pregroup. A sentence’s meaning is f(w₁⊗…⊗wₙ), where f is the linear map got by putting vector spaces in place of the pregroup types in the reduction p₁…pₙ ≤ s. The ε maps become inner products and the η maps become Σᵢ eᵢ⊗eᵢ. So “John likes Mary” is Σ_ijk c_ijk⟨v|vᵢ⟩⟨w_k|w⟩ sⱼ ∈ S, and “not” is the 2×2 swap matrix placed on the sentence wire. There is no theorem and no experiment. The sentence space S is shown only as 1- or 2-dimensional truth values, over a toy model with one basis vector per individual, and the paper says how neither S nor the verb tensors are to be built from data. Its headline similarity numbers (3/4, 1/4, 3/8) are unnormalised inner products. Under its own Definition 5.1 they are 0.949, 0.316 and 0.6. | Read |
| NOTE-246 v2 | Reasoning about Meaning in Natural Language with Compact Closed Categories and Frobenius Algebras | The paper sets the sentence space equal to the noun space (F(s)=F(n)=W). It then uses the basis-copying map σ: v_i ↦ v_i⊗v_i of the Frobenius algebra fixed by W’s basis to lift corpus-built verb matrices Σ_i sbj_i⊗obj_i into rank-3 tensors. So a transitive sentence becomes sbj⊙(verb·obj) (“copy-subject”) or obj⊙(verbᵀ·sbj) (“copy-object”), a vector in W. On the Grefenstette–Sadrzadeh verb-disambiguation set, copy-object gets Spearman ρ=0.172, against 0.168 (Kron), 0.163 (Multp), 0.143 (copy-subject), 0.050 (additive) and an inter-annotator bound of 0.620. On a 112-term definition-matching task the best F1 is 0.21 (nouns) and 0.27 (verbs). No significance test is reported. The disambiguation numbers are identical to the authors’ COLING 2012 paper, and the definition numbers are the same data re-reported. | Read |
| NOTE-294 v2 | A theory of concepts and their combinations I: The structure of the sets of contexts and properties | Aerts & Gabora model a concept as an entity with states that change under contexts, a “state context property system” (Σ, M, L, µ, ν). They report an 81-subject e-mail study in which 7 contexts shift the rated frequency of 14 exemplars of “pet” (for example dog 0.50 under “The pet is chewing a bone” against 0.12 in the unit context) and the applicability of 14 properties. They then postulate that contexts and properties each form a complete orthocomplemented lattice, and call the structure quantum-like because a state can be an eigenstate of e ∨ f without being one of e or of f (eq. 19). No theorem is proved, the lattice is not shown to be non-distributive or orthomodular, and the paper contains no Hilbert-space model and no treatment of combination; both are deferred to Part II. | Read |
| NOTE-299 | A theory of concepts and their combinations II: A Hilbert space representation | Aerts & Gabora embed the “pet” data in ℂ¹⁴⁰⁰. Each of 1,400 “basic contexts” is one vector of an orthonormal basis, the ground state is the uniform superposition, and each context is the projector onto the span of its basic contexts. The Born rule then returns the measured frequencies as counting ratios, for example cat under “chewing a bone” = 75/303 = 0.25. “Pet fish” is modelled as the entangled state Σ_{u∈E} |u⟩⊗|u⟩/√100 in ℂ¹⁴⁰⁰⊗ℂ⁴⁰⁸, so that collapsing “pet” to goldfish collapses “fish” to goldfish too. Every projector in the model is diagonal in one basis, so the fitted model is a classical probability space in quantum notation, which the authors half-concede (a “quantization … of this set theoretical model”, §3.3). Two of the chosen counts are infeasible: parrot and guppy. | Read |
| NOTE-337 | A Framework for Representing Knowledge | The frame memo. A frame is a stereotyped situation-structure whose terminals carry conditions and weakly bound defaults. Frames sharing terminals form systems whose transformations encode moves, actions and shifts of viewpoint, and mismatches are routed through a learned similarity network. Applied to vision, imagery, language, memory, search and control, with an appendix arguing that monotonic, consistency-seeking logic cannot carry commonsense knowledge. A programme statement: nothing is built or measured, and the author says what is missing. | Read |
| NOTE-344 | Plain Talk About Neurodevelopmental Epistemology | The first worked mechanism of the society-of-agents view, then “The Society of Minds”. Agents too simple for language exchange no messages: each argument sits at a fixed location on shared channels (c-lines), and the agents themselves are short-term memory. A specificity gradient makes low-level communities local, persistence-memory and a case-shift handle interruptions without a stack, and the infant’s internal “cognitive cases” are conjectured to precede and shape linguistic cases. Speculation throughout; the author says so. | Read |
| NOTE-590 | The stance triangle | Stance is one act with three aspects: in taking a stance a speaker evaluates an object, thereby positions a subject (usually the self), and thereby aligns with other subjects. The triangle’s nodes are two subjects and a shared object; its sides are directed vectors of evaluation (two) and alignment, with positioning reflected back on each subject. Stance belongs to utterances in dialogic context, not to sentences, and needs an object even when none is spoken. A stance follow that resonates with a stance lead must mark the relation (too, either). Argued from Santa Barbara Corpus excerpts; no counts are reported. | Read |
| NOTE-594 | Cours de linguistique générale | Linguistics takes as its object langue, the social system of signs, not parole or langage. The sign unites a concept (signified) and an acoustic image (signifier) arbitrarily; the signifier is linear. Arbitrariness makes the sign immune to deliberate change and exposed to blind change. Synchrony and diachrony are irreducible. A term’s value is set only by its oppositions to coexisting terms: “dans la langue il n’y a que des différences sans termes positifs”, though the sign as a whole is positive. Terms relate syntagmatically and associatively, and relative motivation limits arbitrariness. | Read |
| NOTE-596 | Cumulative cultural evolution in the laboratory | Four chains of ten adults each learned a 27-item artificial language from 14 items and passed their test output on. Transmission error fell and string–meaning structure rose. Unfiltered (Experiment 1), the languages collapsed to 2–5 words that systematically underspecified meanings (colour lost in every chain). With homonyms filtered from training (Experiment 2), they became compositional, with substrings for colour, shape and motion. Participants believed they were copying. | Skimmed |
| NOTE-599 | Language Evolution by Iterated Learning With Bayesian Agents | Iterated learning is a Markov chain on hypotheses. If every learner shares a prior and samples from its posterior, the chain is a Gibbs sampler on P(d|h)P(h), so the probability a learner holds language h converges to the prior P(h), whatever the amount of data passed on. If learners take the MAP hypothesis, it is stochastic EM with no observations: the outcome centres on the prior’s mode, depends on data amount and noise, and in a two-language case is the same for every prior strength below a threshold. An unbounded population with equal fitness has the same equilibrium. | Read |
| NOTE-602 | Performative updates | Makes the performative/constative distinction a difference between two operations on a context set: informative update eliminates indices, performative update changes them minimally so that a proposition holds. Declarations are purely performative; an assertion is a performative truth guarantee by the speaker plus a negotiable informative update; the same declarative sentence can be either. Argued from linguistic evidence, not tested. | Read |
| NOTE-606 | Pragmatic language interpretation as probabilistic inference | States the rational speech act framework (a listener inverting a softmax speaker whose utility is the log-probability a literal listener assigns to the intended world) and its uncertain extension (a joint posterior over the world and the speaker’s topic, knowledge or lexicon), and reviews evidence that these predict reference-game judgements, cost and knowledge effects, scalar and embedded implicature, hyperbole, irony, metaphor and vague adjectives. | Read |
| NOTE-608 | The Semantics of Definite and Indefinite Noun Phrases | Argues that indefinites (and definites) are variables without quantificational force, bound by the nearest operator, which yields donkey-sentence readings; then recasts semantics as file change: each sentence updates a structured common ground of numbered cards, truth of a file supplies existential force, and the definite/indefinite contrast reduces to one felicity condition (new card versus familiar card), from which binding, presupposition projection and, with accommodation, narrow-scope definites follow. | Read |
| NOTE-611 | Referential intentions in word learning | A Bayesian learner that infers speakers’ intended referents and the lexicon jointly, allowing words to be used non-referentially and intentions to be empty, learns a much more precise lexicon from annotated child-directed video than association or translation models and reproduces mutual exclusivity, word-driven object individuation and intention-guided mapping as consequences of its likelihood. | Read |
| NOTE-613 | That’s what she (could have) said | Recursive Bayesian speaker–listener reasoning yields specificity implicatures but cannot derive Horn’s principle from utterance costs alone; letting the agents be uncertain about the lexicon breaks the symmetry, so costlier signals go to less likely meanings in one-shot games without equilibrium refinements, and two online games with novel symbols show people drawing both implicatures without prior conventions. | Read |
| NOTE-614 | On Linguistic Aspects of Translation | Names the three kinds of translation (intralingual, interlingual, intersemiotic) and argues that interlingual translation renders whole messages, since code-units rarely have full equivalents; that every cognitive content can be conveyed in any language; and that languages differ in what their grammar obliges a speaker to say, so translation adds and drops information and a chain of translations can erode a message. Poetry, where grammar and sound carry meaning, allows only creative transposition. | Read |
| NOTE-616 | Iterated learning (Kalish, Griffiths & Lewandowsky) | Puts Bayesian iterated learning to a laboratory test with a bias known beforehand: human function-learning chains converged on a positive linear function in 28 of 32 families, from positive linear, negative linear, U-shaped and random starting functions, as convergence to a prior favouring positive linear functions predicts. The prior was taken from earlier studies, three families settled on the negative linear function, and the authors do not claim the chains reached stationarity. | Read |
| NOTE-617 | Defaults in Update Semantics | Sets out update semantics (meaning as change of information state), proves a dynamic system reduces to static propositions exactly when updates are total, idempotent, persistent, monotone and strengthening, and shows epistemic “might” and “presumably” are tests that fail persistence. Gives a decidable non-monotonic logic of default rules in which specificity and other priorities follow from coherence and applicability conditions, with predictions that differ from Reiter’s, Delgrande’s, Asher and Morreau’s and inheritance-net theories. | Read |
| NOTE-618 | Structuralism | From a one-page-in-four preview of chs. I–IV: a structure is defined by wholeness, transformation and self-regulation. Its elements are subordinated to laws of composition; “all known structures… are, without exception, systems of transformation”; and in logic and mathematics regulation is perfect because every operation has an inverse. Structures whose transformations unfold in time, linguistic or social, are regulated but not fully reversible. Structures are real only if governed from within, by a subject defined as “the center of functional activity”. | Skimmed |
| NOTE-627 | Informative communication in word production and word learning | Writes the informative speaker down: utility is −D_KL(intended meaning ‖ literal word meaning), so a speaker referring to one object picks a true word with probability ∝ |extension|^(−α), the size principle from communicative premises. Learners who know the referent and invert this speaker bet on novel-word meanings as predicted (r = .93); speakers’ free descriptions follow informativeness only weakly (r = .19), and colour terms not at all. | Read |
| NOTE-628 v2 | Larger communities create more systematic languages | Varies community size alone in a laboratory language-creation game: groups of eight built more systematic, compositional languages than groups of four, faster and more uniformly across groups, while reaching the same communicative success and convergence. Input variability, greater in larger groups, predicted gains in structure; small groups differed more from each other on every measure. | Read |
| NOTE-631 | Efficient compression in color naming | Casts a colour lexicon as an information-bottleneck encoder from perceptual meanings to words and shows the World Color Survey languages and English lie near the IB bound at β ≈ 1.03, beating hue-rotated variants of themselves in 93% of cases and fitting full naming distributions much better than a deterministic efficiency model; IB optima are soft, and the path of optima through β has phase transitions that roughly follow Berlin and Kay’s sequence. | Read |
| NOTE-634 | Predicting pragmatic reasoning in language games | A listener who inverts, by Bayes’ rule, a speaker choosing words by informativeness to a literal listener, combined with an empirically measured salience prior, predicts mean listener bets in simple three-object reference games at r = .99 with no fitted parameters, and the speaker model predicts speaker bets at r = .98. | Read |
| NOTE-644 | Semantic Unification | Puts basic DRT into sheaf form: a DRS is a section of a presheaf of consistent literals, an anaphoric resolution is a cover by variable maps, and the discourse’s meaning is the gluing, unique if it exists (Proposition 1). All the interpretive work is in choosing the cover; the only obstruction to gluing shown is inconsistency, so the paper models context dependence and ambiguity, not contextuality in the Abramsky–Brandenburger sense. Its probabilistic section ranks covers by summed corpus counts, not by gluing distributions. | Read |
| NOTE-645 v2 | On the Quantum-like Contextuality of Ambiguous Phrases | Models meaning selection in two-word ambiguous phrases as a Bell-type measurement scenario. Hand-set supports give one possibilistically contextual (Hardy-type) model; every corpus-estimated model is signalling, so only Contextuality-by-Default applies, under which two noun–verb pairs read in both grammatical orders are contextual (1/30 and 7/30). The paper’s own bootstrap leaves both plausibly noncontextual (probabilities above .56 and .08), on very small counts. | Read |
| NOTE-648 | Quantum-Like Contextuality in Large Language Models | Builds a three-sentence anaphora schema with the support of the PR prism, instantiates it 51,966,480 times from Simple English Wikipedia, and takes one referent probability per sentence from BERT: 0.148% of the models pass a signalling-corrected sheaf test and 71.1% the CbD test. Because the schema fixes the support, both verdicts depend only on how uncertain BERT is in the three sentences; the tie to embedding distance is a softmax identity with weak correlations. | Read |
| NOTE-654 | Analysing Ambiguous Nouns and Verbs with Quantum Contextuality Tools | Proves that in a binary cyclic system CbD’s signalling quantity Δ is twice the sum of each content’s minimal direct influence over Jones’s canonical causal models, which is the total-variation distance between its marginals. In 90 rank-2 noun–verb systems from British corpora, Δ is about 1.35 for every ambiguity class, and homonymous verbs carry about 70% of it against about 50% for polysemous verbs. That difference rests on 14 systems from six verbs and is marginal by a recomputed test (p ≈ 0.06 two-sided). No system is tested for contextuality. | Read |
| NOTE-687 | How the structure of language is acquired | Extends the Random Hierarchy Model to predicting a masked last token. The correlation between two tokens shrinks by about a factor m per level of their common ancestor, a sample of P sequences resolves it only above a (v²P)^(−1/2) noise floor, and deep transformers and CNNs show loss steps at P_ℓ ≈ v m^(2ℓ−1), where their layers come to encode the level-ℓ hidden symbol. On character-level text, the correlation decay exponent predicts where the loss for a given context length levels off. The route from correlations to learned hidden variables is argued, not proved. | Read |
| NOTE-689 | Deriving neural scaling laws from language statistics | Predicts the data-limited scaling exponent of language-model test loss as γ/(2β), from two measured corpus statistics: the decay with lag of the top singular value of the token–token covariance (β), which sets the usable context n*(P) ~ P^(1/(2β)), and the decay of next-token conditional entropy with context (γ). The derivation rests on assumed scaling forms and a fast-learning condition; on TinyStories and WikiText-103 transformers match the predicted exponents up to about 10^8 tokens and horizons of a few tens of tokens. | Read |
| NOTE-718 | Gallego & Orús — Language Design as Information Renormalization | Reads Chomsky’s MERGE as a coarse-graining and builds sentence probabilities as mostly loop-free tensor networks of 3-index MERGE tensors, one per node of a given parse. Because MERGE’s output is fixed by its input, a parsed sentence’s probability is a product of local coefficients. Tree-shaped parses make correlations decay with tree distance, about log |j − i|, so polynomially in word separation. The averaging step that turns this into a claim about languages is asserted, and nothing is tested on data. Syntax only: semantics and context are set aside by the authors. | Read |
| NOTE-734 | Grice — Logic and Conversation | pp. 45–47 only: conversational implicatures are to be tied to a Cooperative Principle (contribute what the accepted purpose of the talk exchange requires) and to maxims of Quantity, Quality, Relation and Manner. Grice flags the second Quantity maxim as disputable, defers Relation, and calls the information-exchange framing too narrow. | Skimmed |
| NOTE-781 | Torii, Maeda and Hidaka — Distributional hypothesis as isomorphism | An SVD of log raw co-occurrence counts solves 61% of Google analogies over a 1,487-word vocabulary, against 75% for full word2vec. In a 24-sentence toy corpus whose eight subjects carry every combination of three binary aspects, the subjects’ raw co-occurrence rows form a parallelepiped exactly when, for each object, its four sentences are equally probable. The attributes are independent, so the concept lattice is Boolean. | Read |
| NOTE-1186 | Bayazit, Mueller and Bosselut — Crosscoding through time | Trains one sparse crosscoder jointly on the mid-layer activations of two or three pretraining checkpoints, so features share an index across them, and scores each feature’s causal importance for subject–verb agreement at each checkpoint (Relative Indirect Effect). In Pythia-1B, OLMo-1B and BLOOM-1B the important features move, by the authors’ annotation, from token-specific detectors to grammatical-category ones, and in BLOOM from language-specific to crosslingual ones for Latin-script languages; importance keeps shifting after accuracy plateaus. A method for describing which features emerge, persist and lose importance, not for predicting it. | Read |
| NOTE-1339 | Srinivasan & Choi — TyDiP | TyDiP gives graded politeness ratings, three crowd annotators per item, for 500 Wikipedia talk-page requests in each of nine languages. Mean pairwise Spearman agreement is 0.46 for target-language items and 0.36 for English items rated by the same bilinguals. Leave-one-out human accuracy on the binary top-and-bottom-quartile set hovers around 90%, about 10 points above an English-trained XLM-R in six languages. But annotators were screened for agreement with that classifier, and the paper’s “across all nine languages” is false for Tamil. | Read |
| NOTE-1435 | Gutt — Translation and Relevance (thesis, 1989) | Gutt’s thesis puts translation under Sperber and Wilson’s interpretive use. Utterances resemble each other to the degree that the assumptions each communicates in its own cognitive environment overlap, so identical propositional form does not guarantee identical interpretation. A translation is faithful when it resembles the original closely enough in the respects that make it relevant to its audience, which is a context-dependent standard. Direct translation is complete interpretive resemblance in the context the original envisaged; its communicative clues, from word order to connectives and greetings, are fixed by their effect on interpretation, not by their form. | Skimmed |
| NOTE-1451 | Nida — Principles of Correspondence | Nida sets out formal and dynamic equivalence. Formal equivalence matches the message’s form and content and suits gloss translations with notes. Dynamic equivalence follows the principle of equivalent effect: the receptor’s relation to the message should be substantially what the original receptors’ was, as in Phillips’s handshake for the holy kiss. It is “the closest natural equivalent” to the source message, natural to the receptor language, the message’s context and the audience. It must still reflect the source’s meaning and intent, with meaning before style. There is no single right answer. | Read |