Revisiting bridging contexts in language change: core properties and the role of frequency vs. salience

2.1 An amended definition of bridging contexts

I propose the following amended definition of bridging contexts, which builds on Hansen (2021: §5.2), but takes into account insights from Hansen and Terkourafi (2023):

A bridging context is a context that underspecifies the meaning of a given linguistic expression used in an utterance, such that it allows for both a conventional and an innovative interpretation as being plausibly perceived by hearers as communicated. In a bridging context, the innovative interpretation will typically – perhaps always – be facilitated via some kind of metonymical (foreground/background) relationship to the conventional sense of the expression. The inference to the innovative interpretation need not be invited by the speaker, although, in some cases, it may be. Bridging contexts may trigger change at the levels of both lexicon and grammar.

Where bridging contexts are attested prior to change – i.e., prior to the appearance of switch contexts – the innovative interpretation arises as a context-based inference and does by definition not yet exist as a conventionalized part of the grammar or lexicon of the language in question. It follows that bridging contexts that precede change cannot, as such, be characterized by ambiguity (either semantic or structural), cf. De Smet (2009: 1729), given that ambiguity can obtain only if there is a choice between two already conventionalized expressions. Instead, a bridging context is one that fails to sufficiently constrain the interpretation of a particular element of a given utterance.^[4] It may therefore lead hearers to attribute certain salient inferences that are compatible with, but go beyond, the conventional meaning of the expression to the expression itself, or even – as we shall see below – to completely misinterpret the speaker’s communicative intention.

My notion of an underspecifying context is thus very close to Denison’s (2017: 293) use of the notion of vagueness, which applies “where a linguistic analysis is in some relevant respect underdetermined at least for [the hearer …], but no further information is needed for interpretation”. I differ from Denison (2017) in two ways, however: firstly, I include “wrong” readings (i.e. ones that are not compatible with the speaker’s intentions) as a possible outcome of contextual underspecification (pace Denison 2017: 294). Secondly, I treat underspecification specifically as a feature of certain contexts, rather than as a feature of certain linguistic expressions. That means that it becomes possible for a given expression to have a quite specific meaning, but for the derivation of that meaning to nevertheless be insufficiently supported by a given context.

Consider a couple of comparatively simple examples of lexical expressions that have relatively recently become polysemous, indeed auto-antonymical, in contemporary Danish:

Let us look first at the figurative expression X verb op ad bakke (‘X verb uphill’).^[5] The older sense of this expression, which is becoming obsolete, is one where something or someone is evolving in a positive direction (for instance, towards increased wealth, improved well-being, etc.), as in (1) below. The newer sense illustrated in (2) is one where some activity is described as difficult, effortful, and not necessarily tending towards a particularly positive result.

Clearly, embodied experience suggests that walking up a hill is typically harder than walking on level ground. At the same time, the idea of progress towards a positive outcome is fully compatible with having to expend substantial effort in order to achieve that outcome. Hence, it is not difficult to imagine contexts like the one in (3), which displays a metonymical relationship between the older and the newer sense of X verb op ad bakke, and where at least some hearers may well perceive the effort involved as more important than the progress towards promotion as such, even if the speaker’s focus is on the latter:^[6]

The fact that the old and the new interpretation stand in such a metonymical relationship does not mean, however, that the two must necessarily be mutually compatible as in the case of (3). Hansen (2021: 7) adduces the example of another auto-antonymical expression in contemporary Danish, namely the transitive verb forfordele, which to older speakers means ‘to give someone less than their fair share of something’, but which tends to mean the exact opposite (‘to give someone more than their fair share of something’) to speakers born from the mid-1960s onwards.^[7] These two interpretations do stand in a metonymical relationship to one another in so far as the idea that one person is given less than their fair share of something will normally imply that one or more others are receiving more than their fair share. At the same time, the two are evidently not communicatively equivalent. It is therefore essentially inconceivable that any speaker for whom the older interpretation is the conventional one could plausibly ever have intentionally invited the newer interpretation. As Hansen (2021: 7) argues, the rise of this polysemy is likely to have been triggered by contexts of use in which the notion of unfairness towards a subset of a group of people was evidenced, but where the identity of the particular subset involved was underspecified by the context, as when a speaker – assuming that the conventionalized sense of the verb is familiar to the hearer – simply utters (4):

Consequently, changes triggered by bridging contexts must be conceived as fundamentally hearer-based. Expressions that have acquired new auto-antonymical senses constitute clear evidence of this, as speakers need not (and probably typically will not) be at all aware that the context makes an alternative interpretation available. This type of change therefore cannot plausibly be seen as driven by speakers. And even in (non-auto-antonymical) cases where a speaker may in fact be using an expression in the hope of communicating something over and above its conventional sense, hearers must still pick up on the invited inference in order for it to trigger change.^[8] That said, it is of course important to point out that any reanalysis triggered by a given bridging context will, in the first instance, effect change only to the grammar(s) of the hearer(s) making the reanalysis. Hearers who have made the reanalysis are then in a position to propagate it in their role as speakers by using it in switch contexts. For that reason, we would expect that, by default, a given expression will have to occur in more than a single instance qualifying as a bridging context, such that the reanalysis can be made by multiple hearers, in order for the change to take hold across the speech community. This does not imply that a very large number of individual reanalyses need be involved, however, as language change can be triggered by a relatively small number of individuals who happen to enjoy significant social prestige (Labov 2001).

As implicitly suggested by examples (3) and (4), “context” is not limited to linguistic “co-text” on the present understanding of bridging contexts. The literature has naturally tended to focus on co-text, as that is typically all that is available for the purpose of diachronic analysis, and this focus may go some way towards explaining why semantic/structural ambiguity has been thought to be a necessary feature of bridging contexts. Examples (3) and (4), as well as several other examples below, show that this is not so, however. Hearers who reinterpret linguistic expressions in actual communication situations may in principle do so on the basis of any aspect of context in its broadest possible sense, including:

1. the (para)linguistic behavior by co-interactants and third parties who are known to be present;

2. the physical context, comprising the environment at large, as well as non-verbal behaviors by co-interactants and third parties;

3. the knowledge context, comprising knowledge and assumptions about co-interactants, about ways of doing things (including – but not limited to – linguistic and interactional norms, text types, and speech events), and about the world more generally (i.e., encyclopedic knowledge).

Now, objectively speaking, the elements of (ii) and (iii), in particular, are infinite in number. Language users therefore cannot, and do not, pay equal attention to all objectively identifiable aspects of the context of communication. As argued by Sperber and Wilson (1986: 15–16), in order to be useful in accounting for the interpretation of utterances, context must be understood as a cognitive phenomenon. Utterance context, on such an understanding, consists of the subset of assumptions about (i)–(iii) that individual language users bring to bear in interpreting a given utterance. Context, in other words, is not “out there”, but is something that is subjectively constructed in the process of interpretation.

Hence, there is no necessary identity between the individual cognitive contexts thus constructed by different participants in any communicative event. Given that bridging contexts are a subtype of utterance contexts, this means that they need not actually be perceived as bridging contexts, i.e., as involving two different possible interpretations of any given expression, by all – or even any – of the participants in the communicative event. As shown by (4), what counts as a bridging context for the purpose of explaining language change may be one where both participants most plausibly perceive one and only one specific interpretation of the expression in question, but the hearer’s interpretation happens to differ from that of the speaker. For this reason, bridging contexts are best seen as a second-order, or analyst’s, concept. In other words, while linguists can empirically identify instances in corpora where two possible interpretations of a given expression, which is known to be either synchronically or diachronically polysemous, are contextually possible, we cannot know whether the actual discourse participants perceived things in that same way in all cases.^[9]

Despite what has been suggested in the previous literature (cf. Section 2 above), bridging contexts are not a clearly demarcated stage of change. For one thing, such contexts can continue to be found after change has taken place, provided the original use of the expression continues to be available. Indeed, it is the normal case for the original use to remain current for at least some time (in some cases, indefinitely) following any given change. To take a stock example, the French compound future tense is formed by the present indicative of the motion verb aller (‘go’) used as an auxiliary, followed by the infinitive of the lexical verb, as in (5):

This construction arose towards the end of the 15th century, in bridging contexts where aller functioned as a lexical verb describing actual movement on the part of the subject and was followed by an infinitival clause indicating the purpose of the movement (Togeby 1974: §231). As in the preceding lexical examples, this change involves metonymy, in as much as the purpose adjunct necessarily designates a future state-of-affairs whose realization is made plausible by the fact that, at reference time, the subject is moving to a place where the purpose can be fulfilled. In other words, in the source construction, purposeful movement in the present is foregrounded, whereas the (plausible) future realization of the purpose is backgrounded. The target construction, where aller has been grammaticalized as a future-tense auxiliary, on the other hand, foregrounds the futurity of the state-of-affairs described, while any sense of physical movement is backgrounded, indeed very frequently altogether absent.

As seen in (6), the source construction is still in use more than 500 years later:

Conversely, as we shall see in Sections 4.1 and 4.2, contexts that are compatible with two different interpretations of a given expression may be instantiated across very long time periods, even centuries, without triggering change. Language users are never compelled to avail themselves of changes made possible by bridging contexts, so the attestation of such contexts cannot, in and of itself, be taken as evidence that change is in progress (cf. Heine 2002: 85).

In other words, bridging contexts may or may not be followed by switch contexts, in which only the innovative interpretation is possible. The attestation of switch contexts is clear evidence that, for at least some language users, change has indeed taken place. Unlike bridging contexts, the first appearance of switch contexts is therefore a more or less clearly delimitable stage in the process of change.^[10]

4.1 Case study 1: the rise of auto-antonymy in the French verb sanctionner

My first case study relates to the lexicon. It concerns the French transitive verb sanctionner (‘to sanction’), which is synchronically polysemous between two more or less auto-antonymous senses.

According to the Trésor de la langue française informatisé , sanctionner is first attested in 1777, with a sense that may be paraphrased as ‘to ratify a legal arrangement’.^[21] In other contexts, this original sense, which is still in use, paraphrases variably as ‘(formally or officially) approve of/attest to/confirm the status of/provide support for’ etc. Very broadly speaking, sanctionner used with this sense thus describes an action that is beneficial to the entity filling the role of theme or patient. An example from a Modern French text is found in (8) below, where the meaning is that the giving of presents confirms the cultural status of the holidays in question:

In the early 20th century, sanctionner gains a new and very different sense, namely ‘to penalize/take disciplinary action against’. According to the Trésor de la langue française informatisé, the first attestation of this sense dates from 1930, and this is also what I find in my corpus data.^[23] A more recent example is given in (9), where the meaning is that the owner of the bicycle was penalized for their errors by having use of the bicycle withdrawn:

Compared to the older sense, this newer sense thus roughly describes an action that is detrimental to the theme/patient. While these two senses are not logically opposed in the same way as the two senses of the Danish verb forfordele adduced in Section 2.1 above, they are to some extent auto-antonymical. Both senses of sanctionner are found in contemporary French and, unlike the Danish verb, the use of sanctionner is subject not just to inter-speaker variation, but also to intra-speaker variation. Thus, the two different senses may occasionally be found within one and the same text.

For both this and the following case study (cf. Section 4.2 below), I made use of the electronic Frantex database. Frantex is the largest extant database of synchronic and diachronic written French. It spans the entire history of the language, from 950 to the present day. At the time of writing, it contains a total of 5,679 texts (272,389,277 words), representing a variety of different written genres, although with an emphasis on literary text.

For the study of sanctionner, I confined my analysis to the Modern French subcorpus, which comprises texts from the period 1800–1979. The total number of words in this subcorpus is 151,948,368. I performed an automated search for the lemma sanctionner, yielding a total of 658 tokens, all of which were analyzed manually. Dividing the corpus into 25-year periods (with the exception of a final 30-year period), the 658 tokens were distributed diachronically as shown in Table 1.

I classified each example in the database as representing (i) the source meaning, (ii) the target meaning, or (iii) a bridging context, where both interpretations were potentially available. Clear bridging contexts were found both before and after the first attestation of switch contexts in 1930, from the earliest period and up until the present day. Example (10) is taken from approximately a quarter-century before the change took place, (11) follows the change by four decades, thus illustrating the point made in Section 2.1 above that bridging contexts are not a strictly delimited stage of change:

In identifying bridging contexts, I deliberately chose to be relatively liberal, so as to give the proposals made by Larrivée and Kallel (2020) and Winter-Froemel (2021) the benefit of the doubt, as it were. Thus, I included examples like (12) below, where it is clear to a careful reader that only the source meaning could plausibly have been intended (insofar as rules are not entities that can be subject to penalization), but where the general tenor of the passage and the syntactic structure of the host clause is such that I could imagine that a not-so-careful reader, or a reader who was unfamiliar with the genre, might understand the target meaning to be the most relevant one:

The three categories of examples were quantified across the seven time periods in Table 1, and the results are shown in Figure 1. As the figure shows, there is no subperiod where the proportion of bridging contexts comes anywhere near the 50 % of all occurrences suggested by Larrivée and Kallel (2020) and by Winter-Froemel (2021).^[24] Not only that, but we see no marked increase in the proportion of bridging contexts in the 25-year period (1900–1924) preceding that in which the change takes place. On the contrary, a fairly substantial decrease in bridging contexts can be observed, compared to the period from 1875 to 1899.

Figure 1:

Classification and evolution of the uses of sanctionner across time periods.

4.2 Case study 2: the reanalysis of the French pas plus (X) (que)

My second case study is grammatical, rather than lexical, in nature. It concerns the change undergone by the French negative analytic comparative construction pas plus (X) (que) (‘not more (X) (than)’), which has been reanalyzed as a negative connective (‘neither/nor/not either’). As in the case of sanctionner in Section 4.1, the source construction is still in (very common) use in contemporary French, so together the two uses constitute a case of what Hopper (1991: 22) calls “layering”.

The original comparative meaning of the construction is illustrated in (13):

Here, what is negated in the source construction is simply the existence of a difference between the focus of comparison (= the entity being compared to something, in casu the Holy Virgin) and the baseline (= the entity that serves as a benchmark for comparison, in casu all other women). Importantly, (13) thus does not negate the possibility that either entity has value, but remains neutral on that point. Whether or not the speaker believes that either or both have value in and of themselves is something that the hearer will have to work out from the wider context in which the construction is used. In other words, the use of pas plus (X) que is fully compatible with both the Virgin Mary and women in general being seen as having great value.

The newer negative connective meaning is illustrated in (14):

Here, the speaker is not denying that there’s a difference between Perec writing text for the machine and him trying to make the machine write text for him. Instead, each of the two clauses is negated independently, and in addition to negating the second clause, pas plus que functions as a connective, instructing the hearer to process the contents of two negative clauses, both of which represent new information in the context, as related in terms of some relevant umbrella category (such as “things that Perec will not do in relation to the machine”). Crucially, a “literal” comparative interpretation would not make sense in this instance: once one has asserted that a state-of-affairs X (here, ‘Perec writing text for the machine’) does not obtain, it is communicatively pointless to go on to add that X does not obtain to a greater degree than another state-of-affairs Y (here, ‘Perec making the machine write text for him’), particularly when the degree to which Y obtains has not been stated elsewhere in the context. The negative connective processing instruction, on the other hand, is part and parcel of what is communicated by pas plus que here.

An example of a plausible bridging context is seen in (15):

Notice the difference between the clear use of the target meaning in (14) and the bridging context in (15): in (14), the idea that Perec will not himself compose text for the machine is explicitly coded in the negative clause preceding pas plus que. Furthermore, the clause marked by pas plus que introduces new information, viz. the idea that Perec will also not attempt to make the machine in question produce text. In (15), on the other hand, the idea that Fatmé may not have liked the sentiment at all is not explicitly expressed, but is merely a plausible inference: it is possible to like a sentiment at least a little bit and still ultimately be bored by it. Moreover, in this case, the constituent following pas plus que expresses old information, namely the idea that Fatmé liked the sentiment to a low degree (or perhaps not at all).

As this case study grew out of a broader ongoing study of negative connectives in French, I included all of the Frantex database, from Old French up to the present day, in order to study the rise of that use of pas plus (X) (que). The periodizations used follow those of the database itself, and they are seen in Table 2.

All instances occurring in Old and Middle French texts were analyzed. As the use of pas plus (X) (que) is highly frequent across all subsequent time periods, I had to sample the data from the Pre-Classical period onwards to make qualitative analysis feasible. Depending on the total number of examples in each period, the data from that period were randomly sampled at regular intervals ranging from every 5th instance in Pre-Classical French, through every 25th instance in both Classical and Contemporary French, to every 100th instance in Modern French. The distribution of the sample across the six time periods is shown in Table 3.

As in the case of sanctionner above, each example was then classified as instantiating the source construction, a bridging context, or the target construction. The same principle of generosity in attributing bridging-context status was followed as in the case study involving sanctionner. As shown in Figure 2, bridging contexts are found already in Middle French (albeit represented by only a single instance), whereas we have to wait until the Classical French period for the target meaning to appear.

Figure 2:

Classification and evolution of the uses of pas plus (X) (que) across time periods.

Moreover, just as in the case of sanctionner, we see that occurrences of the pas plus (X) (que) construction in bridging contexts never reach an overall frequency even approaching the 50 % benchmark suggested by Larrivée and Kallel (2020) and by Winter-Froemel (2021) in any time period.^[25]

That said, Classical French is a 150-year period, and in principle much may happen during that length of time. In a second step, I therefore divided the Classical French period into 25-year intervals and quantified the uses of pas plus (X) (que) separately within each interval. The results are shown in Figure 3. The figure reveals that the target meaning is attested only in the last third of the overall period, and that the proportion of bridging contexts does not reach as much as 20 % in any interval. While a slight rise in bridging contexts can be observed in the interval immediately preceding the change (i.e., 1725–1749), they still constitute a mere 12.8 % of uses in that interval, and the slight rise does not appear to be part of an overall rise in bridging contexts across Classical French. Indeed, the proportion of bridging contexts is at its highest (17.6 %), not in the later intervals, but in the earliest one from 1650 to 1674. If the frequency of bridging contexts were the main driver of reanalysis, the change should thus have been more likely to take place in the period from 1675 to 1699, rather than as late as 1750–1774.

Figure 3:

Classification and evolution of the uses of pas plus (X) (que) across sub-periods of classical French.

This second case study, from grammar, thus reinforces the conclusion drawn from the lexical case study in Section 4.1, that high frequency of bridging contexts does not appear to be a pre-requisite for change.

4.3 Salience of the innovative interpretation in bridging contexts

The alternative hypothesis that will be defended here is that the frequency of bridging contexts as such is, in fact, not centrally important to whether or not such contexts end up triggering change. Instead, what is more likely to be important is the contextual salience, to hearers, of the innovative interpretation they make available.

There are two related aspects to the salience of an innovative interpretation of a given linguistic expression: (i) how accessible is that interpretation given the context of utterance?, and (ii) how much attention do hearers pay to it as a possible alternative to the conventional interpretation, once they have accessed it?^[26]

In the first instance, the salience of an innovative interpretation seems like a strong candidate for an alternative triggering mechanism because reanalysis implies the acquisition, on the part of hearers, of a new meaning for the reanalyzed expression. Intuitively, it is plausible that more salient (a fortiori highly salient) phenomena would be more easily acquired than less salient or non-salient ones.

Secondly, not only is salience an already widely used concept within cognitive, usage-based approaches to linguistics (e.g., Evans 2019: passim), but it has for a long time been acknowledged as an important factor in language users’ acquisition and evaluations of sociolinguistic variants, and of so-called social meanings more generally (Labov 2001: 196–197). Sociolinguists have pointed out that not only may degrees of salience vary across different social meanings/variants, but the salience of a given meaning/variant may also vary across different language users (Johnstone 2016: 638–639). I argued in Section 2.1 above that language users’ awareness of bridging contexts may similarly vary across time and across individuals. This difference in awareness can be described in terms of how salient hearers perceive the availability of more than one interpretation of a given linguistic expression to be in such contexts.

The existing literature (e.g., Auer 2014; Divjak 2019; Ellis 2017; Giora 2003; Kerswill and Williams 2002; Rácz 2013) has proposed a variety of different characterizations of salience, and of what makes a given (socio)linguistic phenomenon salient. Some of these involve frequency considerations. However, when examined together, the different approaches suggest that the relationship between salience and frequency is far from straightforward. Schmid and Günther (2016: 1), in a short, but very helpful paper, systematize the available characterizations into four types, based on the interaction between sources of hearer expectation (i.e., long-term memory vs. the current context) and two mechanisms of salience, viz. the confirmation vs. the violation of expectations. This interaction and the four types of salience it gives rise to are represented in Table 4, which should be read as follows:

1. Where the linguistic input matches expectations derived from knowledge stored in long-term memory about language usage in general, we get what Schmid and Günther (2016: 2) call “salience by context-free entrenchment”. Linguistic expressions that trigger this type of salience are usually highly frequent in general usage.

2. Where the linguistic input matches expectations based on knowledge of the particular type of context in which an utterance is produced, the result is “salience by contextual entrenchment”. Expressions that are salient in this way are not necessarily highly frequent in general usage, but will typically occur with substantial frequency in this particular context type.

3. If the input fails to match expectations derived from knowledge stored in long-term memory of usage patterns across context types, we have “salience by novelty”. Only expressions that have overall low frequency will normally trigger salience by novelty.

4. Finally, if the input fails to match expectations associated with the context type, the type of salience that may be triggered is “salience by surprisal”. Here, the triggering expression will normally occur with only low frequency in the specific kind of context in question (but may occur with higher frequency elsewhere).

In other words, salience does not result directly from frequency, nor does it result from frequency alone. Indeed, salience can correlate with both high and low frequency, depending on what causes the salience.

If we apply this model to bridging contexts, it seems to follow that if bridging contexts are as frequent prior to change as Larrivée and Kallel (2020) and Winter-Froemel (2021) claim – i.e., if they constitute 50 % or more of all the contexts in which a given expression is used – then they should be highly familiar to hearers, and any salience achieved by the innovative meaning that is derivable in bridging contexts would therefore have to be of either Type (1) or Type (2).

This, however, raises the question why hearers should feel the need to attribute a new coded meaning to the expression in langue, if that meaning can quite standardly be inferred from the use of the expression in parole. The pragmatics literature contains a wide range of examples of expression types with highly frequent additional “non-literal” interpretations, such as expressions triggering Generalized Conversational Implicatures (or GCIs, for short) (cf. Levinson 2000). Now, GCIs are precisely not usually reanalyzed as coding the implicated interpretation; in fact, as pointed out by Hansen and Waltereit (2006: 240), the very frequency with which such implicatures are triggered may act to impede reanalysis.

An analogous argument might be made about very frequent bridging contexts, not least because the status of an innovative interpretation that is derivable in such a bridging context prior to the attestation of switch contexts seems in fact to be very similar to that of a GCI, given that it is not coded, but is rather the result of a common, but defeasible, inference.

Conversely, infrequent bridging contexts may in at least some cases be what pushes some hearers into attributing a new meaning to a given expression. This latter scenario would be consistent with both classic and more recent work in learning theory (e.g. Rescorla and Wagner 1972; Stahl and Feigenson 2015), whereby organisms (presumably including language users) learn new things principally when events violate their expectations. When events conform to expectations, on the other hand, there is reduced, if any, impetus to learn from them.

With respect to language change, it is evidently not possible to accurately annotate historical corpora for the presumed salience to hearers of any innovative interpretations made possible by bridging contexts. That said, corpus data may in some cases provide at least speculative support for the role of salience. Thus, it is interesting to observe, in the data on sanctionner adduced in Section 4.1 above, that when looking at the overall frequency of the lemma, what we find is that, in both absolute and relative terms, it is at its lowest in the period from 1900 to 1924 – i.e., the period immediately preceding the one in which the innovative ‘penalize’ sense of the verb is first attested in switch context. The normalized frequencies of the lemma are shown in Table 5, and raw frequencies of the three categories of uses (and, by implication, of the lemma) are shown in Figure 4.

Figure 4:

Absolute frequencies of sanctionner and its different uses across time periods.

It is evidently possible that this observation is a mere artifact of the composition of the particular corpus used: it may simply be the case that the texts from 1900 to 1924 happen to treat topics where sanctionner is a less relevant verb to use than in the texts from other periods. However, as shown in Table 1, Section 4.1, sanctionner is a low-frequency verb in any period, suggesting that it may be relatively unfamiliar to many language users even at the best of times. The apparent further drop in frequency in the period preceding the change may have contributed to salience by novelty, particularly in bridging contexts, and thus helped facilitate its reanalysis.

Due to the inevitably speculative nature of considerations like this, however, empirical substantiation of my hypothesis about the role of salience in triggering change will have to be obtained principally through experimental work, rather than by looking at historical corpus data. Importantly, work of this nature has already been done on the acquisition of sociolinguistic variants, and the results point precisely in the direction of salience by novelty/surprisal being a key factor.

For instance, Sumner et al. (2014) find that atypical, i.e., less frequent, but socially salient pronunciations of words are remembered better than typical, i.e., more frequent, but socially less salient ones. The experiments carried out by Rácz et al. (2017) demonstrate that, at equal frequencies of exposure, people are better able to acquire and to generalize socially salient linguistic patterns in an invented language than otherwise comparable patterns that are not socially salient.

The results reported in Lai et al. (2020) seem particularly germane to the hypothesis defended in the present paper: in a nutshell, these authors use experimental data from an invented language to show that salience by both novelty and surprisal facilitates the acquisition of a (socially conditioned) innovative form for an already established linguistic meaning. The hypothesis I am proposing here is that salience by novelty/surprisal may be central to facilitating the acquisition of an innovative meaning for an already established linguistic form. It seems not unreasonable to expect that the two types of processes may have something (indeed, possibly much) in common.