Robyn Carston
Affiliation: University College London
Category: Linguistics
Keywords: polysemy, schematic meaning, context-sensitivity, lexical pragmatics
Date: Wednesday 3rd of September
Time: 15:30
Location: Gen. Henryk Dąbrowski Hall (006)
View the full session: Polysemy & Communication
It’s widely recognised that the words comprising our substantive vocabulary (nouns, verbs, adjectives) are polysemous, i.e. they have multiple interrelated senses. For instance, the noun/verb ‘mouth’ has a different sense in each of the following: ‘the whale’s mouth’, ‘the mouth of a cave’, ‘the mouth of a river’, ‘five mouths to feed’, ‘to mouth obscenities’, ‘he’s a loud mouth’, among others (Falkum & Vicente 2015; Vicente 2018; Carston 2021). As with most cases of polysemy, ‘mouth’ is cross-categorial, that is, the interrelatedness of senses spreads across words of different syntactic categories, so it may be that the basic unit of polysemy is not a word but a categoryless root (Carston 2022).
There is a strong intuition that underlying these families or networks of senses there is a common meaning, that is, a single unifying core meaning (perhaps a set of features) which all the different senses of a given word/root share. This is typically assumed to be an abstract, schematic, semantically-underspecified (i.e. pre-truth-conditional) entity which must be given a fully semantic value in specific contexts of use, whether via pragmatic selection of one of the established senses or via a process of pragmatic specification resulting in an ad hoc sense; that is, the core meaning is in the ‘wrong format’ (Recanati 2004, 2017) and a pragmatic operation is necessary to recover a fully semantic contribution. The pull of this ‘core meaning’ idea is widespread. The linguist Ruhl (1989), perhaps its strongest advocate, maintains that all content words are linguistically monosemous, their lexical meaning comprising an abstract, unitary schema (inaccessible to consciousness), which is common to all actual uses and underspecified along multiple dimensions (e.g. concrete/abstract, static/dynamic, causal/noncausal, etc.), which become specified in contexts of use. Experimental psychologists investigating the online processing of polysemous words often express a similar position, e.g. ‘By core representation, we mean specifically a memory structure encompassing all semantic features that are common across multiple senses of a polysemous word’ (Klepousniotou et al. 2008, p. 1535); also Frisson (2009) and Brocher et al. (2016, 2018). Among philosophers of language, Pritchard (2017) argues for a ‘something in common’ across all uses of a word, while Vicente (2024) and Liu (2025) view a single underspecified meaning of polysemous words as at least a live option.
In this talk, I present a range of considerations that lead me to conclude that, seductive though the idea is, a common core meaning is a myth. First, it is noticeable that despite intensive efforts to provide accounts of these core meanings, they have proved elusive; Ruhl (1989) maintains that it is impossible to articulate them because ‘these general abstract meanings elude consciousness’ (p. 51); Vicente (2018) says the main problem for the underspecification approach is that ‘there is no summary or abstract representation that encompasses all the senses of a regular polysemous expression.’ (p. 13), here confining himself to regular cases such as the metonymically related book-information/book-tome, church-institution/church-building, lamb-animal/lamb-meat, etc. alternations. In fact, the same goes for irregular polysemies, such as those that arise from various kinds of pragmatic modulation in context, including metaphorical cases (Carston 2021).
The main claim of this talk is that the failure to delineate core shared meanings is inevitable, due to the very nature of the pragmatic processes by which one sense is derived from another. There may be shared features between a given pair of senses, specifically between the sense/concept that is the input to the pragmatic modulation process and the sense/concept that is its output, e.g. the two senses of ‘run’ in ‘run a mile’ and ‘a home run’. However, it’s not long before chains of highly context-sensitive pragmatic inference take off in myriad directions, with new sense modulations, metaphorical uses and metonymic uses building one upon another, so that any sharing of features across whole polysemy families soon disappears (cf. ‘run a business’, ‘run in a stocking’, ‘run for president’, ‘the school run’, ‘run an argument’, etc.), a situation well captured by Langacker’s (1991) networks of dozens of pragmatically interrelated but, typically, non-overlapping senses.
Another consideration mediating against semantically underspecified meanings for words concerns children’s acquisition of such abstract non-semantic meanings. What the child encounters in communication and learning is fully semantic (truth-conditional) word meanings (senses); how would she, then, induce some other schematic core meaning, and to what end? Consider what function or purpose a core meaning might have and whether in fact there is evidence of anything performing such a function. Two potentially useful purposes come to mind: (a) a unifying function, something binding the myriad polysemes together into a single representation and so distinguishing polysemy from homonymy (e.g. bank/bank); (b) a constraining purpose, such that ad hoc concepts/senses, although responsive to the specific contexts in which they occur, cannot be too inventive or fanciful because the core meaning imposes a restriction on communicators’ intentions regarding word senses, thereby aiding hearers’ comprehension processes. Regarding (b), I show that there just is no such constraint and that, as some advocates of core meanings concede, any underspecified meaning is an abstraction over features of established senses, so ‘a novel interpretation of a word cannot be captured by the underspecified meaning.’ (Frisson & Pickering 2001: 159). Regarding (a), I argue that the representational unity of a polysemous word is captured by assuming that the related senses are stored together as a cluster or network (a single, albeit structured, unit); this also adequately explains the online processing evidence that distinguishes polysemy and homonymy.
Finally, I will consider whether the way in which word meaning is represented in the currently burgeoning field of LLMs might run counter to my ‘anti-core-meaning’ claim. Employing a vector-based system of ‘meaning’ in a multi-dimensional space, these models appear to assume that, for any word, there is a ‘key’ which is context-independent and neutral across the word’s myriad senses (Buckner 2024, 2025; Lettieri 2024). I intend to explore the extent to which this bears any significant relation to the idea of a shared core meaning for all senses and uses of a word.