Lexical Consensus in Language & AI

• Lexical consensus is a multifaceted phenomenon where shared word meanings emerge through linguistic, psycholinguistic, and computational processes.
• It examines frameworks from morphosyntactic derivation and many-sorted semantics to measurable dialogue entrainment, highlighting convergence patterns.
• Computational approaches validate consensus via crowd-sourced lexical resources, model agreement metrics, and grounded perceptual label learning.

Lexical consensus denotes a family of convergence phenomena concerning lexical units, lexical relations, and word–meaning mappings across linguistics, psycholinguistics, computational linguistics, and artificial-agent research. In one line of work, it concerns what counts as a word and what a lexicon stores; in another, it concerns the stabilization of shared referring expressions in dialogue; in another, it is operationalized by repeated agreement among independent speakers building a lexical network; and in recent AI work it concerns whether agents can acquire and stabilize grounded labels over a fixed perceptual substrate (Murphy, 2024, Joubert et al., 2012, Shi et al., 2023, Vera, 20 Jun 2026). Taken together, these usages suggest that lexical consensus is not a single doctrine but a recurring problem of how lexical structure becomes shared, measurable, and theoretically constrained.

1. Wordhood, roots, and the lexicon as process

A central theoretical use of lexical consensus arises in debates over wordhood. Murphy explicitly denies that there is any single, primitive mental object corresponding to the common-sense word. On this view, a word is the read-out of a derivational process that selects a conceptual root, merges it with syntactic or categorial features, and routes the resulting structure to conceptual and sensorimotor systems. The stored lexicon is therefore not “a dictionary in your head” or a list of monolithic word entries, but a set of processes and stored feature inventories enabling mappings among roots, syntax, semantics, and form (Murphy, 2024).

Within this framework, a root becomes a word only when it is paired with a syntactic category and relevant semantic or formal features. A “word” is thus a morphosyntactic object: a root plus a bundle of syntactic and semantic features, together with its mapping to modality-specific realization. Murphy’s formulation makes wordhood dependent on derivational configuration rather than on whitespace, orthographic segmentation, or a lemma-like atomic entry. The architecture is explicitly procedural: select root; merge root with syntactic features such as $N, V, A, P, D, T$, person, number, gender, or case; send the structure to conceptual and sensorimotor systems; assemble semantic features such as argument structure, event structure, and Pustejovsky-style Formal, Constitutive, Telic, and Agentive properties; and realize the result phonologically or orthographically (Murphy, 2024).

This position rejects several older assumptions that have often underwritten informal appeals to lexical consensus. Orthographic and phonological “lexicons” are treated as sets of instructions for sensorimotor transformations rather than as the lexicon proper. Dictionary-like inventories of words are rejected in favor of separable stores of conceptual roots, syntactic features, semantic instructions, and mapping rules. Lemma-based models are criticized as incompatible with a non-lexicalist morphosyntactic architecture, especially in the presence of polysynthesis, suppletion, idioms, and context-sensitive form–meaning relations. A direct implication is methodological: lexical decision, naming, priming, production, and aphasia studies that assume atomic word representations risk conflating concept selection, morphosyntax, and phonology. Murphy accordingly argues that experiments should factor argument structure, event type, Telic content, and category flexibility, and should target distinct stages such as message generation, message-to-syntax recoding, syntax-to-form realization, and articulation (Murphy, 2024).

2. Lexical sorts, transformations, and the search for shared formal inventories

In formal lexical semantics, lexical consensus concerns the status of base types and the formal organization of lexical meaning. Retoré and colleagues argue that standard Montague Grammar, with primitive types $e$ and $t$, is too coarse to model selectional restrictions, polysemy, and copredication. Their many-sorted alternative replaces a single entity type with lexical sorts $T_1, T_2, \dots, T_n$, so that lexical entries become typed resources in a richer compositional system. Yet they also emphasize that, although researchers widely agree that richer types are needed, there is no consensus on what the base types actually are. Their proposal is that classifier systems in classifier languages and sign languages supply a cognitively and linguistically motivated basis for lexical sorts, with inventories on the order of 100–400 classes, fine-grained for physical entities and structured around dimensions such as animacy, shape, and function (Mery et al., 2013).

This proposal treats classifiers not as peripheral morphosyntactic ornaments but as candidate base sorts for a many-sorted logic. Japanese classifiers such as $Nin$ for persons, $Hon$ for long thin objects, $Mai$ for flat objects, and $Dai$ for vehicles, as well as classifier handshapes in sign languages, are taken to reflect recurrent lexical partitions already grammaticalized in natural language. The underlying claim is that lexical consensus in formal semantics should not rest on ad hoc ontologies or a single undifferentiated $e$, but on an explicit and testable inventory of base sorts with both linguistic evidence and cognitive plausibility (Mery et al., 2013).

A complementary formalization appears in the $\Lambda TY^n$ framework, where lexical items are many-facetted objects with a main typed term plus optional typed transformations. Meaning assembly uses System $e$0, while semantic representations are stated in many-sorted higher-order logic. Optional transformations such as $e$1, $e$2, or $e$3 capture regulated facet shifts, and their combinatorics are constrained by flexibility markings such as rigid and flexible. This makes copredication calculable rather than intuitive: a sentence like “Liverpool is a large city and voted Labour” is licensed because the relevant transformations are jointly available, whereas “Liverpool is a large city and won the cup” is blocked when the club-reading transformation is rigid. The same system introduces polymorphic quantifiers such as

$e$4

and a polymorphic Hilbert operator

$e$5

so that quantification and determination remain uniform across lexical sorts. On this view, lexical consensus consists not in fixing a single denotation for each word, but in sharing a constrained inventory of types, transformations, and admissible composition patterns (Mery et al., 2013).

3. Dialogue-level consensus: entrainment, referring expressions, and expectation alignment

In dialogue research, lexical consensus is most directly formalized as lexical entrainment. The relevant unit is not an abstract word type in the lexicon, but an established referring expression used by multiple interlocutors for the same object. The definition adopted for conversational systems restricts lexical entrainment to noun-phrase referring expressions and defines two expressions as equivalent when, after removing modifiers, their head nouns are the same regardless of noun form. A lexical entrainment expression is considered established iff equivalent referring expressions have been produced by at least two speakers and at least one instance is free rather than merely constrained inside a larger noun phrase. The degree of lexical entrainment for speaker $e$6 is then

$e$7

where $e$8 counts instances of established lexical entrainment expressions in the $e$9-th utterance from $t$0 (Shi et al., 2023).

This definition turns lexical consensus into an explicitly measurable dialogue phenomenon. In MULTIWOZ-ENTR, a dataset derived from MultiWOZ 2.1, 62,961 lexical entrainment expressions are identified; the expression lexicon size per dialogue has mean around $t$1; approximately $t$2 of established expressions are initiated by the user and $t$3 by the agent; $t$4 and $t$5; and priming distance has median and mode $t$6, with $t$7 of expressions adopted within two turns. Human agent utterances reach $t$8, whereas AUGPT, MinTL, HDSA, and MARCO score $t$9, $T_1, T_2, \dots, T_n$0, $T_1, T_2, \dots, T_n$1, and $T_1, T_2, \dots, T_n$2, respectively, all significantly below the human value. The practical conclusion is that current response generators under-entrain: they often fail to maintain shared terminology even when task completion remains adequate (Shi et al., 2023).

Adjacent work on speech perception broadens the notion from shared lexical labels to shared expectations about upcoming dialogue. Speech is modeled as carrying a lexical channel, consisting of the words themselves, and a non-lexical channel, consisting primarily of prosodic realization. In a generalized turn-discrimination paradigm, non-lexical information systematically reduces ordinal and permutation entropy across participants, even when it lowers discriminative accuracy. For ambiguous stimuli, lexical-only accuracy is $T_1, T_2, \dots, T_n$3 while acoustic presentation yields $T_1, T_2, \dots, T_n$4; $T_1, T_2, \dots, T_n$5 falls from $T_1, T_2, \dots, T_n$6 to $T_1, T_2, \dots, T_n$7, and $T_1, T_2, \dots, T_n$8 from $T_1, T_2, \dots, T_n$9 to $Nin$0. For remaining stimuli, lexical-only accuracy is $Nin$1 and acoustic accuracy $Nin$2, yet consensus still increases, with $Nin$3 falling from $Nin$4 to $Nin$5 and $Nin$6 from $Nin$7 to $Nin$8. A plausible implication is that lexical consensus in interaction is not exhausted by exact surface-form agreement: interlocutors also converge on discourse-level expectations shaped by prosody and other non-lexical channels (Wallbridge et al., 2023).

4. Crowd agreement as a lexical resource construction principle

A different operationalization appears in large-scale lexical resource construction. In the JeuxDeMots project, lexical consensus is the foundational acquisition mechanism. Two players, working double blind and asynchronously on the same target term $Nin$9 under the same consigne $Hon$0, independently produce sets of associated terms $Hon$1 and $Hon$2. Only the intersection is retained: $Hon$3 Each agreed response creates or reinforces a typed relation, yielding relation quadruples of the form $Hon$4, with weight operationalizing the number of agreeing player pairs. Consensus is thus both binary, because at least two players must converge, and quantitative, because repeated convergence increases the relation weight (Joubert et al., 2012).

The same principle governs evaluation and consolidation through the tip-of-the-tongue tool AKI. Given cue words $Hon$5, AKI computes lexical signatures $Hon$6, proposes candidates with highest activation, and normally intersects signatures across cues: $Hon$7 When intersection empties the candidate set, a recovery procedure uses sum or union instead. This makes retrieval a test of whether a term’s network neighborhood is sufficiently structured and redundant to support identification from its lexical environment. The empirical results are strong: on 500 “Tabou inversé” cards AKI succeeds in $Hon$8 of cases, whereas humans reach approximately $Hon$9; on 6,522 real sessions, overall success is around $Mai$0–$Mai$1, about $Mai$2 for frequent words and about $Mai$3 for unrestricted open vocabulary. By March 2011 the network contained 229,000 terms, more than 1,100,000 semantic relations, approximately 900,000 games, more than 2,500 players, and about 5,000 sense refinements, while expert intervention affected less than $Mai$4 of relations (Joubert et al., 2012).

In this tradition, lexical consensus means that lexical relations are validated by repeated convergent judgments of non-expert speakers. It is therefore closer to common lexical knowledge than to prescriptive lexicography. The resulting network is not merely a crowdsourced thesaurus; it is an explicit graph whose edges exist and strengthen only when independent speakers repeatedly agree that they should.

5. Community convergence, polarization, and collective identity

In social-media analysis, lexical consensus appears as lexical convergence within polarized communities. Users are represented by normalized bag-of-words vectors over lemmatized comment vocabularies, and lexical convergence between users $Mai$5 and $Mai$6 is measured by cosine similarity: $Mai$7 Interaction level is defined from co-commenting behavior by

$Mai$8

On Italian Facebook data from January 2010 to December 2014, covering 73 public pages, 271,296 posts, 9,164,781 likes, and 1,017,509 comments, science- and conspiracy-polarized users share 38,258 lemmas, accounting for $Mai$9 of all word tokens used in both communities, yet their interaction networks are highly segregated (Brugnoli et al., 2019).

Within same-community pairs, lexical convergence rises strongly with interaction. Weighted least squares regression of average convergence on $Dai$0 yields $Dai$1 with $Dai$2 for science–science pairs and $Dai$3 with $Dai$4 for conspiracy–conspiracy pairs. Dynamic analysis shows that the increment in lexical convergence across an interaction history also grows with interaction period length, with $Dai$5 and $Dai$6 for science pairs and $Dai$7 and $Dai$8 for conspiracy pairs. A randomization test based on permuting users’ bag-of-words vectors yields $Dai$9, excluding a trivial frequency-based explanation (Brugnoli et al., 2019).

The interpretive claim is that lexical convergence can function as a proxy for the emergence of collective identities and polarization in opinion dynamics. This does not mean that communities use wholly disjoint lexicons; they do not. Rather, a large shared lexical core coexists with distinctive topic-specific and affectively loaded subsets, and repeated homophilic interaction drives increasingly similar lexical profiles within each echo chamber. In this usage, lexical consensus is not a settled lexicon-wide ontology but an interaction-driven alignment of vocabulary associated with group formation.

6. Computational and artificial-agent formulations

In LLM reliability research, lexical consensus is operationalized much more narrowly as surface-form overlap among model outputs. The Multi-Model Consensus Reasoning Engine includes lexical and structural cues such as token overlap, Jaccard similarity, ROUGE-L, length, numeric patterns, and identical final answer strings, but explicitly contrasts them with richer semantic and clustering features. Majority vote is treated as a nearly pure lexical-consensus baseline. Empirically, lexical agreement is useful but weak: macro-average accuracy is $e$0 for majority vote, $e$1 for the best single model, and $e$2 for the graph-attention consensus model; the latter improves by $e$3 points over the strongest single model and $e$4 over majority vote. In ablation, removing semantic similarity and clustering reduces macro-average accuracy from $e$5 to $e$6, whereas removing lexical and structural features reduces it only to $e$7. On TruthfulQA myth-style hallucinations, majority vote selects the false-but-plausible option in $e$8 of relevant cases, the best single model in $e$9, and the consensus GAT in $\Lambda TY^n$0. In this setting, lexical consensus is explicitly a secondary signal, subordinate to semantic agreement, reasoning quality, and model priors (Kallem, 12 Jan 2026).

A more literal computational use appears in grounded word learning. The Lexical Consensus framework studies whether artificial agents can acquire and stabilize nonce labels over a fixed perceptual substrate using frozen DINOv2 embeddings, Carroll-style labels, and simple lexical learners. The main result is a perceptual-coherence gradient. For centroid learners in 3-way, 5-shot episodes, naming accuracy is $\Lambda TY^n$1 for native concepts, $\Lambda TY^n$2 for near-disjunctive coherent overextensions, $\Lambda TY^n$3 for mid-disjunctive concepts, and $\Lambda TY^n$4 for far-disjunctive concepts, approaching chance in the last case. A pre-registered CIFAR-100 dissociation experiment shows that perceptual distance predicts acquisition accuracy with partial $\Lambda TY^n$5 and $\Lambda TY^n$6, whereas semantic distance adds no significant explanatory power, with partial $\Lambda TY^n$7 and $\Lambda TY^n$8. Bidirectional evaluation separates naming from retrieval: under hard candidate pools, exemplar retrieval outperforms centroid prototypes in all tiers, for example $\Lambda TY^n$9 versus $e$00 for native concepts and $e$01 versus $e$02 for near-disjunctive concepts. The paper’s conclusion is that frozen perceptual geometry both enables lexical grounding and limits what can be acquired without representational adaptation (Vera, 20 Jun 2026).

These computational uses show two distinct extensions of the concept. One treats lexical consensus as literal agreement over strings or answers, then demonstrates that such agreement is less informative than semantic clustering. The other treats it as stabilization of shared label usage over perceptual space, then shows that consensus depends on the geometry of the substrate rather than on arbitrary symbolic association. Across both, lexical consensus is measurable only when the relevant unit of agreement—surface form, semantic cluster, or grounded label extension—is precisely specified.