1. Introduction
Productivity can be thought of as a gradient property of any linguistic element that captures the likelihood language users will use it in new ways. Barðdal (Reference Barðdal2008) has labelled this property ‘extensibility’. This extensibility is usually understood in formal terms and then applies to how freely forms distribute over other forms by entering into new formal pairings. Clearly, this is also what Bolinger had in mind when he defined productivity as ‘the statistically determinable readiness with which an element enters into new combinations’ (Reference Bolinger1948: 18). For example, English weak past tense formation has high extensibility, because the past tense -ed-suffix will automatically extend to attach to any verb that is newly introduced into the language. All French-origin verbs that entered English in the fourteenth century form their past tense by -ed-suffixation (appealed, cried, retained and so on), just as any new verb stem adopts the weak past tense in PDE (booped, cancanned, windbagged). In contrast, extensibility is low for abstract noun formation through -th (depth, warmth but *thickth, *calmth). We will refer to this dimension of productivity here as combinatory extensibility.
In the current literature on productivity, semantics largely plays second fiddle. Its relevance is recognized but only in as far as it bears on combinatory extensibility. On the one hand, it is generally accepted that combinatory extensibility correlates with semantic transparency. A form will be productive only in as far as its meaning remains predictable whichever combination it enters into. Barðdal (Reference Barðdal2008) argues that semantic coherence is a predictor of a construction’s productivity. Hay (Reference Hay2003) has demonstrated that low productivity correlates with low semantic transparency. On the other hand, it makes sense that combinatory extensibility is constrained by the semantic compatibility between elements (Goldberg Reference Goldberg1995). For example, the prefix un- can combine with verbs to denote a process involving a reversal of the process denoted by the verbal stem. That is why almost all verbal stems un- combines with express accomplishments and achievements (e.g. undress, unload, unlock, untie). These verbal stems prototypically denote a process that involves both a change of state and an inherent endpoint, making their reversal easily conceivable. Attached to verb stems that prototypically express states or activities, un- becomes awkward or ungrammatical (e.g. *undance, *unstand), or else it coerces an accomplishment reading (e.g. unknow).
However, our conception of productivity can embrace the semantic dimension more fully. It is obvious that just as linguistic elements may or may not extend in combinatorial space, they may or may not extend in semantic space. A conceptual unificationFootnote 1 of productivity along the different dimensions of linguistic organization has been proposed by Langacker (Reference Langacker1987) through his concept of ‘sanction’. Langacker defines sanction as a correspondence between a conventional unit and a target item, which leads to the unit sanctioning the target (Reference Langacker1987: 65–71). Full sanction describes the situation where conventional unit and target item are fully compatible and corresponds to the linguistic notion of well-formedness. Partial sanction describes the situation where the correspondence is (as yet) unconventional and involves extension. The following example illustrates the difference between full and partial sanction:
Consider a child in the process of learning the various senses (conventional usages) of the word tree, and suppose the term is first applied, in his early experience, to such ordinary specimens as oaks, maples, and elms. Their perceptual prominence and obvious gross similarities enable the child to extract a conception that embodies their commonality, while excluding the many properties that vary from one instance to the next. … With continued usage, this initial conception becomes more deeply entrenched, and comes to be invoked for the categorization of more divergent experience. … Suppose, then, that our child has mastered the concept [TREE] (the eventual category prototype), as well as the symbolic relationship [[TREE]/[tree]]. When he encounters a tall plant with branches, leaves, and bark he readily sees it as conforming to the specifications of [TREE] and takes it as a straightforward instance of the tree category. What happens, now, when he first encounters a pine, which is [TREE]-like in most respects but has needles instead of leaves? He will quickly learn to call it a tree, either from hearing someone refer to it in this way or because this is the most nearly applicable term at his disposal. This usage implies the symbolic unit [[PINE]/[tree]], derived by extension from the original [[TREE]/[tree]]. (Langacker Reference Langacker1987: 373–4)
But exactly what determines extensibility? On a usage-based perspective, extensibility must arise from the state of speakers’ internalized language system at any given time. In the above quotation, Langacker links both schema formation (‘to extract a conception that embodies their commonality’) and entrenchment (‘continued usage’) to extensibility (‘categorization of more divergent experience’). In this spirit, De Smet suggests that ‘easy availability of a strongly entrenched symbolic relation facilitates its over-application through partial sanction. The result must be semantic change’ (Reference De Smet and Schmid2017: 80). This certainly aligns well with the current understanding of combinatorial extensibility. In the domain of morphology in particular there is broad consensus about the factors promoting or reducing extensibility. Among those are the semantic constraints mentioned above but also, and more importantly, the complex interplay between type and token frequencies that determines how any given combinatory pattern becomes both schematized and entrenched (Baayen & Lieber Reference Baayen and Lieber1991; Plag Reference Plag1999; Hay & Baayen Reference Hay, Baayen, Booij and van Marle2002; Goldberg Reference Goldberg2006; Bybee Reference Bybee2007; Barðdal Reference Barðdal2008; Schmid Reference Schmid2018).
Where it comes to semantic extensibility, the literature advances detailed proposals. Specifically, in work on polysemy it has been argued that word senses cluster into radial networks, with the majority of senses deriving from a single core sense or what Tyler & Evans (Reference Tyler and Evans2003) (in an echo of Langacker Reference Langacker1987) call the ‘sanctioning sense’. This view makes predictions about where the potential for semantic extension lies for a given linguistic element, and how semantic extensibility is constrained – more on which below. At the same time, acceptance and consensus in the field have proven elusive. Radial networks raise a range of practical as well as theoretical questions that have so far not been satisfactorily addressed. While these issues should justify a certain reluctance to uncritically embrace Radial Network Theory, it would be a mistake to throw out the child with the bathwater. This article seeks to revisit some of the ideas from Radial Network Theory on what determines the potential for semantic extensibility and explores their applicability to diachronic semantic change. To this end, evidence is drawn from the counterpart to semantic extensibility, namely semantic loss. It is argued that cases of semantic loss support the kinds of sense clusterings and directional sanctioning relations posited by Radial Network Theory, and moreover point to the role of frequency as one of the factors that can elevate specific senses to sanctioning status.
In what follows, we start by discussing Radial Network Theory and its contribution to our understanding of semantic extensibility, but also its current shortcomings (section 2). Next, we discuss a number of specific changes, each of which involves the (near-)loss of a historical sanctioning sense in a linguistic element (section 3). The expressions studied range across the lexical-to-grammatical continuum. They are the adjective awful, the preposition and adverb about, and the multifunctional element so. Two patterns come to the fore. First, in each case the loss of a historical sanctioning sense coincides with a marked frequency increase in one or more other senses, likely reflecting the emergence of a new sanctioning sense. Second, loss of an old sanctioning sense typically extends also to other senses that derive directly from it. When an old sanctioning sense obsolesces, this also implicates the other senses that are derived from it and that receive no sanction from the newly emerging sanctioning sense. These findings support some of the basic intuitions underlying Radial Network Theory. That is, senses stand in directional motivational relations to one another, and frequency is one determinant of the likelihood that a sense can form the basis for semantic extensions (section 4).
2. Radial networks and semantic extensibility
Radial Network Theory (RNT) subsumes a range of closely related proposals that seek to describe and explain polysemy by laying out a number of principles of semantic organization. Its roots lie in the Cognitive Linguistics work of the late twentieth and early twenty-first century (Brugman Reference Brugman1981; Lakoff Reference George1987; Brugman & Lakoff Reference Brugman, Lakoff, Small, Cottrell and Tanenhaus1988; Tyler & Evans Reference Tyler and Evans2001, Reference Tyler and Evans2003; Evans Reference Evans2004, Reference Evans2005). RNT maintains that senses stand in motivational relations to one another, in that they are linked through basic conceptual operations such as metaphor or metonymy. Because senses are connected by these motivational ties, they can be thought of as organizing into networks that reflect synchronic (and diachronic) relations of derivation. RNT also holds that at least one of the senses in such a network will have privileged status as the core or sanctioning sense that gives (or gave) rise to the majority of derived senses. This is why networks have the radial structure that gives them their name.
As a toy example, consider the noun door. A plausible sanctioning sense for door would be ‘movable barrier serving to close or open a passage into a room’ (OED 1a), as in (1a), which gives rise to derived senses including ‘opportunity’ (OED 3) through metaphor as in (1b) or ‘opening or passage into a room’ (OED 2) through part–whole metonymy as in (1c). Figure 1 visualizes these sense relations as a radial network. Finally, motivational relations between senses are mostly directional – for example, in metaphorical extensions it is typically (if not exclusively, see Allan Reference Allan and da Silva2021) the more concrete embodied senses that support extension to more abstract senses, rather than vice versa.
Toy radial network for door
This model can in principle explain why some semantic extensions may gain or lose in likelihood over time. When a new extension gains in prominence in a network, it can assume core status and begin sanctioning further extensions. Conversely, because sanctioning relations are mostly one-directional, it will not support the old sanctioning sense, nor any of the senses derived from the old sanctioning sense. Thus, over time, a peripheral extension in the network may become more salient, to the point that the prototypical centre shifts from the old core to the extension. Györi (Reference Györi2002) gives the example of PIE *(s)keu ‘to cover’ to Germanic huson ‘covering for the legs’ and eventually English hose, where the extension ‘covering the legs’ becomes culturally salient and prototypicalizes at the expense of the older centre. At this point, the new prototype can start sanctioning new extensions like the metaphorical ‘flexible tube’. According to Györi, the process is supported by the extension’s original low degree of prototypicality, which leaves it less strongly integrated in the network and makes it easier for it to detach as a new prototype. Baumann et al. (Reference Baumann, Hofmann, Marakasova, Neidhardt and Wissik2023) call attention to the role of frequency in the shift.
While RNT could potentially explain part of the dynamics of semantic loss, there is no denying that RNT faces major challenges. First, with its emphasis on a core and periphery, the model clearly draws inspiration from Prototype Theory (based on Rosch Reference Rosch, Rosch and Lloyd1978; for influential work see i.a. Geeraerts Reference Geeraerts1997 or Taylor Reference Taylor1995), but in doing so it more or less tacitly extends the idea of prototypical organization from purely conceptual categories to the larger constellations of conceptual categories that make up the semantic pole of linguistic expressions. As far as conceptual categories go, the evidence for prototypical organization is strong, but to assume that the same organizational principle can be scaled up to describe polysemy is arguably a leap of faith (however, see Cai & De Smet Reference Cai and De Smet2024a: 4).
Second, RNT struggles to disentangle synchrony from diachrony. Patterns of diachronic extension offer some of the best evidence of motivational relations between senses. But semantic extensions that were once motivated may no longer be so in synchrony. For example, the extension in film from ‘thin skin or layer’ to ‘movie’ only made sense as long as movie production typically involved the use of thin sheets with a layer of photosensitive chemicals. In other words, synchronically, not all senses of an expression will be equally motivated, but little effort has been made to further theorize the role of connection strength in radial networks. Recent psycholinguistic studies have additionally shown that metaphorical and metonymic extensions are processed differently (Lopukhina et al. Reference Lopukhina, Laurinavichyute, Lopukhin and Dragoy2018), underscoring the idea that not all senses and sense connections are equal.
Third, and most problematically, RNT lacks both a unified representational format and a well-circumscribed method to identify senses and sense relations (Sandra & Rice Reference Sandra and Rice1995). Whereas Lakoff (Reference George1987) adopts a fine-grained full-specification approach whereby every sense in the network has its own image schema, Tyler & Evans (Reference Tyler and Evans2003) argue for a principled polysemy model, which attempts to constrain the number of senses and place an increased importance on the role of real-world knowledge and context by distinguishing between full senses and vagueness. The tests they propose do not, however, fully remedy the problems of classical polysemy tests (Geeraerts Reference Geeraerts and Riemer2015). Lumping/splitting controversies inevitably persist.
In the same way, RNT faces problems determining the sanctioning sense in a radial network (Gilquin & McMichael Reference Gilquin and McMichael2018). A variety of criteria have been suggested to solve this problem. For instance, the sanctioning sense in a radial network is said to be experientially more basic, more frequent, first listed under elicitation, older, the basis for morphological derivations and (perhaps somewhat circularly) the most plausible basis from which to derive other senses. But these various criteria may be at odds or simply hard to apply in practice. Moreover, even when one sense supports a majority of extensions, other senses will often support further extensions, at least suggesting the possibility that a single network may sustain multiple cores (Langacker Reference Langacker1987: 396).
New approaches to semantic representation have primarily addressed the methodological issue of sense discrimination by leveraging distributional evidence to analyse polysemy (Gries Reference Gries2010; Heylen et al. Reference Heylen, Wielfaert, Speelman and Geeraerts2015; Hilpert & Correia Saavedra Reference Hilpert and Saavedra2017; Fonteyn 2018; Budts Reference Budts2020). The family of distributional methods has certainly infused a healthy dose of objectivity into the debate, and has proven very apt at NLP tasks. On a theoretical level, its successes have encouraged a view of meaning as a contextually determined point in a continuous and multidimensional semantic space, challenging lexicographic orthodoxy and the very notion of senses (Li Reference Li2024). Unlike RNT, however, this type of approach says little about the constraints on semantic extensibility.
For this reason – and despite the very real problems with the model – some of the basic intuitions underlying RNT continue to be worth exploring, as they offer potential insight into the dynamics of semantic change. Specifically, this article builds on the concepts of sense clustering, extensibility and directionality. First, senses cluster into networks structured by the motivational ties that connect them. In such networks some senses are more closely tied to each other than others, leading to groupings of related senses. Second, some senses in these networks have privileged status as sanctioning senses in that they provide the motivation that both supports new semantic extensions and that sustains existing polysemy. For example, Petré frames the core use of a multifunctional expression as ‘a type of glue by which all uses can stick together’ (Reference Petré2014: 74). Similarly, Huang et al. (Reference Huang, Geeraerts and Zhang2021) show that the more prototypical senses of Chinese 火 huǒ ‘fire’ give rise to more extensions. They cast this in terms of productivity:
the more typical and salient senses of the FIRE character seem to be more productive in terms of the number of the extensions which develop from them. For instance, there are more extensions that develop from LIGHT (three extensions) than from FLAMES (only one extension). (Reference Huang, Geeraerts and Zhang2021: 38)
The idea that polysemous networks consist of more central (core) senses which are productive and sanction peripheral uses is also invoked by Cai & De Smet (Reference Cai and De Smet2024b: 57), who state that:
Sanctioning can be thought of as a form of semantic productivity: the more salient and core-like the sanctioning sense, the stronger the support for any extensions based on it and the more likely the use of any of the resultant secondary senses becomes.
This ties in with the idea that innovations to an expression are likely only to the extent that they are supported by well-entrenched uses (De Smet Reference De Smet2016: 86). The frequency of a sense, then, should be one of the determinants of its extensibility. Third, sanctioning is typically directional, in that the core support its extensions but not vice versa. This is well-supported by the observation that, in diachrony, recurrently attested sense extensions are not reversible (e.g. ‘door’ will give rise to ‘opportunity’ but not vice versa). Therefore, it is plausible that in synchrony, too, a sanctioning sense supports its extensions, but the extensions do not necessarily support the sanctioning sense. When sanctioning status shifts to a new sense, the new sanctioning sense may no longer sustain the old sanctioning sense from which it once derived, resulting in the loss of the old sanctioning sense.
3. Case studies
In what follows, we examine both causes and consequences of diachronic shifts in what constitutes the likely sanctioning sense in the radial network of a polysemous expression. The discussion draws examples across the lexicon-grammar continuum with awful (section 3.1), about (section 3.2) and so (section 3.3).
As a methodological note, the semantic analyses presented below are based on the semantic analyses in the Oxford English Dictionary (OED). Although this article revisits concepts from RNT, it does not seek to adopt all its ideas, let alone solve its shortcomings. These include the difficulties of deciding whether senses should be split or lumped together. By drawing on the semantic descriptions in existing OED entries we have relied on a properly independent source to ensure the best level of replicability for our analyses, and as a guardrail against confirmation bias. It should be noted that by its nature as a reference work, the OED leans towards a splitting approach to semantic description, often drawing maximally subtle sense distinctions. Where we deviate from its analyses it has typically been to lump senses together, so as to keep corpus annotation doable.Footnote 2
3.1 Awful
The dominant sense of awful today, ‘unpleasant, bad’, originated in colloquial English at the end of the eighteenth century. This first case study evaluates the impact of the new extension on the older senses and further development of the adjective in terms of extensibility. It does so by establishing the radial network of the adjective on the basis of the relevant OED entry and corpus data, and by mapping the frequency changes of its different senses from Early Modern English to Present-Day English.
The OED entry for awful lists nine senses. Three of those senses reflect the use of awful as an intensifier (emphasizing extent, something pleasant, and something enjoyable, respectively I.3.a, I.3.b and I.3.C in the OED 3 entry) and have been collapsed into one category in what follows. The sense ‘frightened or apprehensive of doing something’ is not further discussed in this article as it does not occur in the corpus data that were analysed (see below). In this way, eight senses can be distinguished for discussion.
The historically earliest sense of awful is ‘inspiring fear’ (2a), straightforwardly reflecting the meaning of its stem awe ‘reverential fear, admiration’. This oldest sense sanctions the use of secondary, related senses, specifically ‘arousing respect’ (2b) and ‘filled with fear’ (2c). The former differs from ‘inspiring fear’ in its positive connotation, which is often signalled by other coordinated adjectives. It next forms the basis for a further extension, ‘demonstrating respect’ (2d). These senses make up the semantic profile of awful until the beginning of the nineteenth century, when a new sense extends from ‘inspiring fear’.
The new extension, ‘unpleasant’ (3a), retains the negative connotation of ‘inspiring fear’ but is weaker and more general in use. The situations or persons it refers to do not evoke an emotion as strong as fear, but a general disdain. As this sense gains in prominence, it starts sanctioning the new uses ‘of poor quality’ (3b) and ‘unwell, troubled’ (3c). Second, an intensifier use (3d) develops from ‘inspiring fear’ and ‘unpleasant’, emphasizing an unpleasant or conversely enjoyable experience, or a large amount. This use is most apparent in the collocation an awful lot.
The eight senses can be regrouped into three clusters on the basis of their diachronic connections on the one hand and their semantic relatedness on the other. A first group consists of the oldest senses, (2a) to (2d). Later extensions (3a) to (3c) with a weaker expression of emotion make up the second cluster. Last, (3d) differs from all other senses in activating the scalarity in the element it modifies. Figure 2 visualizes the semantic development of awful. The successive clusters are enclosed in dashed lines and labelled chronologically (A representing the oldest cluster, C the most recent). The arrows between senses show the direction of the extensions.
Radial network for awful
To map the frequency changes of the (clusters of) senses of awful, the Reference De Smet, Flach, Tyrkkö and DillerCorpus of Late Modern English Texts (CLMET, version 3.1) and the written sections of the British National Corpus (BNC Consortium 2007) were queried for all spelling variants of the adjective. CLMET is divided into three subsequent periods of 70 years, namely 1710–80, 1780–1850 and 1850–1920. The BNC covers the last two decades of the twentieth century.Footnote 3 Per period, a random sample of 200 hits was annotated for sense and sense cluster. Ambiguous instances (most often between an intensifier use and another sense) were labelled as such. Though somewhat more frequent in the third period, they are generally few and are not further discussed in what follows.
Figure 3 shows that all attested senses in the first period belong to the A-cluster. Awful is used in the sense ‘unpleasant’ in almost a third of all cases, but the other senses, ‘arousing respect’, ‘filled with fear’ and ‘demonstrating respect’, all have a substantial share in the semantic make-up of awful. The second period is likewise characterized by the dominance of the A-cluster. The first sense of the newer B-cluster, ‘unpleasant’, and the intensifier use occur but remain infrequent, while‘of poor quality’ occurs only once and ‘unwell’ is not yet attested. The second half of the nineteenth century marks a notable shift in the semantic profile of awful. The extension furthest removed from ‘inspiring fear’ in the A-cluster, ‘demonstrating respect’, is only attested once and ‘filled with fear’ has become rare, while the share of ‘inspiring fear’ has also declined. The share of ‘unpleasant’ conversely increases, as does the intensifier use. The sense ‘unwell’ is attested for the first time. By the end of the twentieth century, all senses of A but the oldest have disappeared (save for a single occurrence of ‘arousing respect’) and ‘inspiring fear’ appears to be on its way out. The new dominant sense of awful is ‘unpleasant’ and the other B senses have increased in frequency. The intensifier use likewise continues to increase.
Semantic development of awful from Late Modern English (CLMET) to Present-Day English (BNC)
Based on the frequency data and the potential for licensing new extensions, ‘inspiring fear’ and ‘unpleasant’ are the clear successive cores of the network. The other senses in A and B form the periphery, with ‘demonstrating respect’ as the most peripheral sense. When awful extends to ‘unpleasant’ and this extension takes off, the share of the older senses begins to decline. The first sense to be lost is the furthest extension into the periphery, followed by the other peripheral senses. At this point, the old sanctioning sense begins to lose its potential to sustain its various extensions. For its part, the new sanctioning sense begins to support new extensions, forming the emerging B-cluster, but it fails to support the old sanctioning sense, which also goes into decline. By the beginning of the twenty-first century, the old sanctioning sense is the only sense of the A-cluster to survive, albeit at a very low frequency
3.2 About
Just as for awful, the OED was used to establish a semantic network for about. The OED distinguishes four major clusters of senses. The first cluster, A, henceforth called position, contains all static position senses of about. This includes the oldest sense of about, ‘encircling’ as in (4a). This sense gives rise to ‘accompanied by’, ‘in every direction’ and ‘in the vicinity of’. The latter, exemplified in (4b) later extends to ‘near, within reach’ as in have your wallet about you, and its figurative counterpart as in have your wits about you. A new dynamic sense, ‘moving in a circuit around’, also arises from the static ‘encircling’. Together with ‘moving in various directions’ (4c), it makes up the B-cluster, motion. Second, from the A-cluster (and specifically the ‘near’ sense), an approximator use of about extends. This includes approximation in time as in (4d), approximation in number and the general approximator ‘approximately, more or less’. Note that about is an adverbial preposition in (4c) and a degree adverb in (4d).
The motion-cluster gives rise to a mental connection use of about, ‘concerning, regarding’ (5a), with a transfer from motion on a surface to mental motion. This sense later extends to a mental state sense, which describes emotions (such as happy about) rather than thought processes. Several senses extend from both ‘concerning’ and a second sense. For example, ‘consisting of’ (5b) connects to both ‘concerning’ and the figurative reach use described above. Example (5c) contains the sense ‘occupied with’, which is attached to ‘concerning’ and ‘moving in various directions’. It later develops a more grammatical extension ‘on the point of’ as in I’m about to leave. All these senses make up the last cluster of senses, labelled mental connection. Figure 4 summarizes the description of the senses of about in the form of a radial network.
Radial network for about
Next we traced the frequencies of the senses using data from Reference PetréEarly English Books Online (EEBO), CLMET and BNC. The first and third period from CLMET were selected, and data from EEBO were selected for two seventy-year periods, 1470–1540 and 1570–1640, based on general frequency changes of about in these periods. Spoken data from the BNC were excluded, to improve diachronic comparability. The corpora were queried for all spellings of about and random samples were drawn of 200 hits per period. Annotation for sense and sense cluster was done according to the classification outlined above. Ambiguous or unclear instances were again labelled separately but are not further discussed below as they are very infrequent.
The frequency of the individual senses shows that ‘encircling’ (the source sense of about), approximation in time and number, and ‘concerning, regarding’ are the most common senses of about, respectively in the first, second and third and the last two periods. Because ‘encircling’ and ‘concerning, regarding’ are additionally central to the network as hubs for further extensions, they were identified as the core senses. For further analysis, clusters A and D were therefore split up into A1 and D1 for core and A2 and D2 for periphery (i.e. all other senses in the cluster). Figure 5 visualizes the normalized frequencies of the (sub-)clusters A1, A2, B, C, D1 and D2 throughout time.
Semantic development of about throughout Early Modern English (EEBO), Late Modern English (CLMET) and Present-Day English (BNC) in terms of normalised frequency (per one million words)
Before the second half of the sixteenth century, the oldest cluster is the most frequent. All peripheral senses together outnumber the core, but individually none of them are ever nearly as frequent as ‘encircling’. After this first period the frequency of the core and its periphery begins to steadily decline to the point of near-disappearance. By the end of the twentieth century, position does not feature in the data except for two occurrences of ‘in the vicinity of’. The motion cluster first slightly increases and later declines in frequency again. The approximation cluster likewise first increases in frequency, although the increase is steeper, and then stabilizes .
The core of the mental connection cluster steadily increases from the second period onwards. Its periphery behaves unexpectedly in the first period, when it actually outnumbers the core. There is an abrupt rise in frequency of the core after 1850, which is when the periphery really takes off in frequency as well. This rise of the mental connection senses causes an overall jump in the frequency of about. In the last period, cluster D, and especially its core, is dominant.
Cluster A starts out as the dominant one. Both this core and its periphery then begin to lose ground. Only ‘in the vicinity of’ outlasts the core, but with only two occurrences in the last period. By the second half of the nineteenth century, ‘concerning’ becomes much more frequent. At the same time, its periphery starts growing. At this point, the new core seems to be strong enough to start sanctioning new uses, but it does not support the old senses. The overall frequency of about is also impacted by this shift in core, as it more than doubles. The anomaly of the periphery outnumbering the core in the first period is the result of ‘occupied with’ being frequent. Arguably, this sense was already sanctioned by ‘moving in various directions’. The motion and approximation clusters behave independently from the other clusters. They both first rise in frequency and then respectively slightly decline and stabilize by the last period. They are not affected in the same way peripheral senses are by the shift in core. It should be noted, however, that the frequency of the approximator use might be distorted by the vastly increased use of numbers in texts. This increase leads to more opportunities for using adnumeral markers and perhaps also greater need. In addition, the relative independence of the motion and approximation clusters may reflect their distinct syntactic status as adverb(-like) uses of about. This suggests that syntactic divides may to some extent insulate against the effects of semantic reorganization in the radial network.
3.3 So
The semantic development of so, rather than showcasing the dynamics of core and periphery, supports the idea of ‘semantic glue’ holding senses together. So appears in many different uses, ranging from manner adverb to discourse marker. Across this range of uses, its syntactic status is so diverse that connections across the whole network are probably weak at best, and a single synchronic core cannot be easily identified. However, uses that are plausibly related in meaning are seen to undergo the same fate over time, testifying to the role of local clusters in radial networks.
The manner adverb use is the oldest use of so, and has a deictic function (6a). It can be paraphrased as ‘like this, in this way’ and refers to the extralinguistic context. Similarly used deictically, but referring to a stretch of preceding (or occasionally following) discourse is place holder so. In (6b), so anaphorically stands for really very fond of her.
The second function of so is intensification. Intensifier so (7a) plausibly derives from manner and place holder so. The former is likely where so would have accompanied extraverbal comparison (e.g. through gesture), the latter where it would have involved comparison to the same property in a previously mentioned referent. Correlative so, which anticipates a following subordinate clause, is closely related to the intensifier use and often also functions as a booster (7b).
Correlative so can also be used deictically (8a). Often, however, the meaning of correlative so is neither deictic nor solely intensifying. In most cases, the meaning of the following clause is resultative, working in tandem with intensifier so: the result functions as a precise measure of the intended degree of intensification. Example (8b) illustrates this mix of intensification and causality.
The causal meaning is the sole meaning of conjunct so (9a). This form of so can be paraphrased as ‘therefore’ and establishes a causal connection to the previous clause. Subordinator so (9b) also has a causal meaning, introducing a subordinate clause of purpose or result. The subordinate clause is introduced by that, or occasionally as. The causal element is also important for discourse marker so (9c). It mimics a link to previous discourse, as if the utterance somehow logically follows from what was said before, but in fact preceding discourse may be lacking, and so essentially signals that the utterance is the speaker’s logical response to the (extralinguistic) situation at hand. In (9c), a TV presenter comments on a climbing competition. As the competition has come to a close, the presenter signals a change in topic (which logically results from the competition having ended, but this is never made explicit).
In summary, so can be used as a manner adverb or place holder with a deictic meaning, as an intensifier which functions as a booster, or as a conjunct, subordinator or discourse marker with a sense of causality. Correlative so appears with all three senses. To chart the development of so, CLMET and BNC were queried for so and random samples of 200 hits per period were annotated for form and function. The results are shown in Figure 6.
Development of the forms of so from Late Modern English (CLMET) to Present-Day English (BNC) in terms of normalised frequency (per one million words)
Manner adverb so occurs only once, in the second period, and it is therefore not included in the chart. The picture of so in the first two periods, until 1850, is one of stability. The intensifier and correlative are by far the most frequent. Place holder so is the third most frequent use and gains modestly in frequency in the third period. By the 1980s, however, it loses ground again, alongside the intensifiers and correlatives, which show a marked dip in frequency. This dip goes hand in hand with a general decline in frequency. On the other hand, the subordinator, conjunct and discourse marker use are on the rise. The frequency of the latter two respectively more than doubles and quintuples. So most frequently occurs as a conjunct in this last period, and the three categories expressing causality together far outnumber all other uses. A reversal of the situation at the start of the eighteenth century presents itself in terms of frequency: the uses of so with a deictic and intensifier sense have faded while the causal senses have become central.
A major reshuffling takes place in the use of so between the second half of the nineteenth century and the end of the twentieth Figure 6 repeats the pattern whereby decreases in one usage area of a polysemous expression are offset by increases in another. In addition, the changes directly reflect the major usage clusters of so. On the one hand, uses tied to the deictic and intensifier meanings decline whereas uses tied to causal meanings are stable or increase. What this shows is that the fates of different uses of so are tied together by their degree of semantic commonality. Supporting evidence comes from the break-down of change in the correlative uses in Figure 7 below. As discussed earlier, correlatives can express deictic meaning (see (8a) above), intensifying meaning (see (7b) above), or a combination of intensifying and causal meaning (see (8b) above). In our PDE data, excepting one instance of deixis, it is only the latter type that survives reflecting both the general weakening of deictic and intensifying uses of so, and the relative strength of uses that link to causal so.
Semantic development of correlative so from Late Modern English (CLMET) to Present-Day English (BNC) in terms of normalised frequency (per one million words)
4. Concluding remarks
The case studies presented above support two related aspects of Radial Network Theory (RNT). First, they point to the role of frequency in determining which senses are more or less likely to support semantic extensibility. This does not mean that frequency is the only determinant of sanctioning status, but it is a plausible contributor. Second, shifts in what makes up the productive core of a radial network are found to coincide with shifts in the frequency profiles of different senses, and where an old core is seen to go into decline this is typically related to a parallel frequency increase for another sense or sense cluster. As sanctioning relations are directional, a new core will start sanctioning its own periphery, but it will not support the older senses. Third, patterns of semantic loss largely reflect the kinds of sense clusterings posited by RNT in that senses that are closely related tend to go into decline jointly. This could be taken to vindicate a lumping approach to semantic description, whereby whatever goes into joint decline makes up one major sense. That, however, would fail to capture the accompanying shifts in extensibility, as well as more subtle differences in the development of senses on a finer-grained analysis. For instance, the results from the case studies indicate that the periphery is lost before the core, starting from the sense furthest removed from the core.
These different points are illustrated by awful, where loss of the original core sense (‘inspiring fear’) coincides with the emergence of a new and highly frequent evaluative use (‘unpleasant’) that generates its own extensions (such as ‘of poor quality’). As the old sanctioning sense loses its extensibility, senses derived from it (e.g. ‘filled with fear’) become obsolescent as well. The same pattern is seen in about, which has come close to losing its old sanctioning sense (‘encircling’) along with any immediately derived spatial senses (e.g. ‘in the vicinity of’), at the same time as its semantic profile has come to be dominated by the cluster of mental connection senses organized around a new core sense ‘concerning’. Interestingly, some senses – particularly the approximator use of about – appear relatively insulated from the effects of this reorganisation, possibly reflecting their different syntactic status. Finally, so offers the most complex picture. Its uses are so semantically and syntactically diverse that over the period investigated there is no point when its use could be said to be organized around a single sanctioning sense. Yet, semantic clustering in its uses is in evidence all the same during a major overhaul of its use in the twentieth century, when deictic and intensifying uses consistently decline, while causal uses are stable or increase.
These case studies do not solve the problems that beset RNT. They offer no ready answers where it comes to describing polysemy, deciding on the most likely sanctioning sense (or senses) in a radial network, or the exact nature of its motivational ties, or the number of separate senses it should include. Even so, they show that RNT makes plausible predictions about semantic change, in line with the assumptions that sense relations are motivational, that such motivation is mostly directional, and that it depends on the sanctioning status of a core sense. As such, these assumptions hold out the promise of a better understanding of the dynamics of semantic change.
Acknowledgements
We thank the reviewers for their constructive comments on an earlier version of the article. This research was supported by the Research Foundation Flanders (FWO), grant number 11PF524N. The first author is the lead author.
Open access
