3.1 The background to the variable

The variable for consideration here is the syntax of cau ‘won’t’ (pronounced /kaɪ/), a new negative modal that has developed in northern dialects. Historically, this item derives from a lexical verb nacáu /naˈkaɪ/ ‘refuse’ (ultimately from nag, the negative particle used in responsives, see Jones 1999, plus causative -ha- and nonfinite (verbnoun) suffix -au). While the lexical verb is attested from the thirteenth century onwards (Thomas et al. 1950–2002: s.v. nacâf), the phonological reduction from nacáu to cau dates only from the nineteenth century. Even more recently it has undergone a semantic extension from ‘refuse’ to a modal meaning ‘be predisposed not to’ and is usually best translated into English as ‘won’t’. As such it has become more integrated into the system of modal auxiliaries (cf. gallu and medru ‘be able’, ffaelu and pallu ‘be unable’) and thus undergone grammaticalization.⁴

In this new meaning it allows non-volitional, non-agentive subjects:⁵

This semantic extension seems to have occurred in all dialects where the phonological reduction occurred. The distribution of cau was tested in the Welsh Dialect Survey (Thomas 2000: item 448). The WDS results are given in Figure 1.⁷ As can be seen there, cau was reported as present (red) across almost the whole of the northern dialects but is entirely absent (blue) in the south. The patchiness of the distribution is also noteworthy. This may be due to difficulties in eliciting it with the WDS methodology (fieldworkers described the item under investigation to encourage informants to say it, but, in general, did not ask about it directly), but, as will been seen below, the same pattern is found in the SAWD data.

Figure 1

Geographical distribution of cau (data from Thomas 2000: item 448).

3.2 Innovation of negative concord

The main focus of interest here is a further change that is currently in progress, integrating this new modal into the negative concord system. To understand this innovation, we need to consider the way in which sentential negation and negative concord are manifested in Welsh. Sentential negation is expressed in two main ways. Verbs in negative sentences undergo morphological or morphophonological changes. This is most evident in the present tense of the verb bod ‘be’, which has a completely distinct negative-concord paradigm. Thus, in (3)–(4), the verb changes from affirmative mae to negative dydi.

Other verbs show more subtle differences involving the system of initial consonant mutation. In (5), in an affirmative clause, the verb may appear either in its unmutated radical form ceith ‘will get, will be allowed’ or in its soft-mutated form geith, while, in a negative clause, in (6), the radical form is impossible, and the possible forms are the soft-mutated form geith or the aspirate-mutated form cheith.⁸

These changes alone are not sufficient to render a clause negative, however. To do this, the sentential negator ddim must also be added in immediately post-subject position, as in (4) above and (6) below. This demonstrates that these changes mark a form of negative concord, and are not an expression of negation in and of themselves.

These changes are also manifested in clauses containing n-words such as neb ‘no one, anyone’:

Borsley & Jones (2005), working within HPSG, posit a restriction on n-words that they must appear in one of a small number of specific contexts (the Negative Context Constraint), the most prominent one being a negative context, a head marked [POL neg], defined widely to include all of the elements discussed here: dydi in (4), geith and cheith in (6), and dydi in (7). Where such an element is not present, n-words are ungrammatical. A further constraint, the Negative Dependent Constraint, requires some verbal elements to have a negative dependent such as ddim in order to be grammatical.

One complication is that, if ddim would appear adjacent to an n-word, it can (and, in many varieties, must) be omitted. In (8), ddim would appear immediately after the subject Aled, which would be directly adjacent to neb. Note that the mutation triggered by negative concord is compulsory on the verb in (8) (radical gwelodd ‘saw’, soft-mutated welodd), indicating that the sentence is not simply opting out of the negative-concord system.

The grammaticality of parallel sentences containing ddim varies depending on dialect and varies among n-words. Use of ddim before byth ‘(n)ever (in the future or generically)’ and erioed ‘(n)ever (in the past)’ is quite widespread. King (1993: 253) notes of the sequence ddim erioed ‘not (n)ever’ that it “is regarded by many speakers as sub-standard”, confirming its widespread existence. Borsley & Jones (2005: 98) consider the sequences ddim byth and ddim erioed to be characteristic of southern dialects, citing sentences such as the following as grammatical, but only for southern speakers:

The sequence (d)dim neb is in fact also acceptable for some speakers, again predominantly, but not exclusively, in southern varieties.⁹ A naturally occurring example is given in (10).

The equivalent sentences with “weak” negative concord are exemplified in (11) and (12). These are grammatical in all dialects.

The option in (9) and (10) with ddim (or dim) adjacent to an n-word may be thought of as “strong” negative concord, as compared to the regular “weak” negative concord system exemplified in (8), (11) and (12). A similar difference will arise in our discussion of cau, see example (26) below and discussion in section 3.5.4.

Returning to the syntax of cau, consider again example (2) above. Here, cau does not participate in this negative concord system and triggers only the affirmative form of the tense-aspect auxiliary mae. This option is still widely attested and is clearly the historically older form. However, it appears to be losing ground to a relatively recent innovative variant, in which the negative form of the tense-aspect auxiliary, dydi, is used:¹⁰

The central research questions here are: what is the geographic distribution of this variant; why did it arise; and why has it been successful?

Given what we saw in (9) and (10) above, another conceivable possibility is that cau might additionally be associated with the presence of the sentential negative marker ddim, that is, sentences of the following type:

We will also test whether sentences of this type are found and, if so, answer the same questions with reference to them.

3.3 Innovation of VP-ellipsis (auxiliary stranding)

Finally, we want to test how cau relates to ellipsis phenomena. The availability of VP-ellipsis provides a way to test the categorial status of cau and thus its degree of grammaticalization along the pathway from lexical (control) verb to modal auxiliary. Ellipsis after modal auxiliaries is possible in Welsh, as in (15), leaving the modal auxiliary stranded after the aspect particle:

After control verbs, however, as in (16), and raising verbs, in (17), we do not find direct ellipsis. Instead, a complement containing gwneud ‘do’ is possible, and its complement may undergo ellipsis. In these examples, omission of gwneud leads to partial or complete degradation of grammaticality:

Availability of ellipsis therefore provides us with a test for auxiliary status in Welsh.¹¹

In these cases, we can take gwneud to be the light verb that is found elsewhere in periphrastic constructions. An example is given in (18), where gwnaeth ‘did’ is followed by a nonfinite VP. This is synonymous with an inflected verb of the same tense, as given in (19).

Given this, we can interpret the VP-ellipsis in a sentence like (16) as involving elision of a VP-complement of light verb gwneud in v:

The observation that ellipsis of the VP-complement of v is possible in Welsh leads us to an account of the development of cau in which the emergence of ellipsis after cau is linked to its ongoing grammaticalization as a modal auxiliary.

Rouveret (2012), building on such work as Merchant (2001), Gallego (2009) and van Craenenbroeck (2010) develops a theory of ellipsis in which ellipsis is licensed for complements of phasal heads, notably v. Before its grammaticalization as an auxiliary, cau is a lexical, eventive control verb. In Rouveret’s terms it is the spellout of vn+Root in the structure [_vnP vn [_RootP Root [_CP …]]], where vn is the functional projection that labels the root as a verb and introduces nonfinite (“verbnoun”) morphology. Applying this account to the case at hand, the historical (pre-grammaticalization) cau is a full verb, and therefore unable to license ellipsis of its entire complement. Ellipsis is possible only of the complement of the lowest v head in the derivation. As a lexical verb, this cau takes a full clausal complement, and it is only the complement of the v head within this structure that can undergo ellipsis. Since this v head is spelled out as gwneud ‘do’, ellipsis cannot occur directly after cau, and instead speakers will require the sequence cau gwneud (or some other paraphrase).

After grammaticalization has occurred, cau is an auxiliary, and is therefore inserted into v. We would therefore expect it to allow full deletion of its complement. We thus predict that, when cau grammaticalizes from lexical verb to modal auxiliary, it will gain the ability to license bare VP-ellipsis of its complement. That is, we expect conservative speakers to require insertion of gwneud in sentences like the following, while we expect innovative speakers to permit its omission (modulo also variation in negative concord patterns as outlined above):

In this example, before grammaticalization, cau spells out the eventive lexical verbal root (RootP, effectively VP). The structure is:

Ellipsis must target the complement of a vP. Although the highest AspP is complement to a vP, it cannot target this, because doing so would delete semantic content (i.e. the semantics of cau itself) that cannot be recovered from the first conjunct. Ellipsis can, however, target the complement of the second vP. If it does, this vP will itself spell out as light verb gwneud.

After grammaticalization, cau spells out a functional rather than a lexical head (in a much reduced, monoclausal structure), treated as v in the following:

Ellipsis again must target the complement of vP, but now this can be the entire complement of cau itself, with the result that the structure does not spell out as gwneud, and cau is left alone in clause-final position.

Given this prediction, we would like to be able to test the extent to which speakers are in fact moving in their judgments along this pathway.

3.4 Cau in the project questionnaire

Given this context, it was decided to include five test sentences containing cau in the pilot SAWD questionnaire.¹² Three of them present speakers with different options for negative concord: one without negative concord, in (24); one with negative concord of the auxiliary alone (weak negative concord), in (25); and one with both negative concord of the auxiliary and insertion of a concording negative marker ddim (strong negative concord), in (26).¹³

One question, (27), has cau accompanying an n-word, namely neb ‘no one’, in subject position, offering a clause which takes the option of including negative concord. We have seen above that neb requires negative concord. If a speaker has a grammar that does not allow negative concord with cau, sentences of this type should either be ungrammatical for them (because of the conflicting requirement of having both negative concord and its absence) or should yield a double negative interpretation (‘No one won’t come out with me.’), which is obviously pragmatically infelicitous given the context sentence.

Consequently such speakers would be expected to have no grammatical way to combine cau with an n-word (without producing a double negation reading). If they use the affirmative form of the auxiliary, as in (28), then neb fails to occur within a negative context. If they use the negative form, then cau introduces an additional, lexical negation, resulting in a double negation reading (‘No one won’t come out…’). We expect them to paraphrase, either keeping the form of (28), but replacing neb with pawb ‘everyone’, eliminating the licensing problem, or by omitting the modality completely and replacing cau with some item in the same general semantic area, such as isio ‘want’, as in (29).

The final question, given in (30), tests for the grammaticality of bare VP-ellipsis after cau, as discussed in section 3.3 above. From personal observation, it was expected that a few (younger) speakers would permit VP-ellipsis after cau in a sentence such as this.

3.5 Results: Patterns of diffusion

3.5.1 Weak negative concord

Consider first the overall patterns for the presence or absence of weak negative concord, that is, concord reflected in the form of the finite verb only. Figure 2 shows the raw data for all speakers.¹⁴ Speakers are divided into four types according to their responses across the 5 questionnaire items containing cau: (i) those who retained cau, but removed negative concord in at least one item where it was offered (that is, items 17 = example (25) above, 34 = (30), and 41 = (26)) (blue); (ii) those who retained negative concord with cau at least once where it was offered, but did not add it to another item (pink); (iii) those who added negative concord to at least one cau-sentence that lacked it (that is, either item 16 = (24) or item 42 = (27)) (red); and (iv) those who rejected use of cau entirely by rephrasing all of the test sentences (black). This gives a general overall sense of the data, showing that negative concord with cau is found in a diffuse way over much of the area under consideration. However, Figure 2 does not represent the ages of speakers, so gives us no access to age differences or apparent-time change; nor does it allow us to see other more fine-grained differences in the use of negative concord with cau.

Figure 2

Overall distribution of negative concord with cau in the raw SAWD and Siarad data.

To see the geospatial patterns more clearly, we can run a geographically weighted regression (GWR) (for further discussion of the application of this technique in linguistics, see Willis 2017). In a GWR model, regressions are performed at different points across a geographic area, with data weighted according to distance from that point in question. Using year of birth as a global factor, and allowing the intercept to vary across geographical space, this allows us to create a model in which change diffuses from one place to another.

A global logistic regression, that is, one that takes no account of geography, was performed using R to establish a baseline against which to compare the GWR model.

In order to produce more easily interpretable results, year of birth was coded relative to 1950; that is, 1950 is coded as 0, 1972 as 22 etc. (cf. Kroch 1989b: 225). This is an arbitrary choice of date, but the move allows us to interpret the intercept of the model as representing the extent of change in 1950 and the age coefficient as representing the subsequent progression of the change for each additional year since then.

This model produced an intercept value of –1.47199 logits (standard error = 0.40352, z-score = –3.648, one-tailed p-value = 0.000132). This can be interpreted as meaning that a speaker born in 1950 (selected randomly from anywhere for which interviews were conducted) has a probability of accepting negative concord of 0.187 (the link function is p = exp(k)/1+exp(k), where k is the intercept value).¹⁵ The age coefficient is 0.05182 logits per year (standard error = 0.01674, z-value = 3.096, one-tailed p-value = 0.000981).¹⁶ This shows year of birth to be a significant factor at the p < 0.001 level, with younger speakers more likely to use negative concord.¹⁷ The global probability of a speaker accepting negative concord thus rises to 0.393 for a speaker born in 1970 (–1.47199 + (20 * 0.05182) = –0.4356, followed by application of the link function) and to 0.646 for a speaker born in 1990. The AICc of this model (a measure of goodness of fit that takes into account model complexity) is 80.467, and 13.2% of deviance is accounted for.

In the GWR model, we make the following assumptions. These are in conformity with much existing work on language change, and are also the simplest assumptions consistent with the need to allow both change and dialect variation:

1. change proceeds via an S-curve (Greenberg, Osgood & Saporta 1954: 155), which we can model using the logistic function (Altmann et al. 1983; Kroch 1989a; b; Denison 2003; Blythe & Croft 2012);

2. for a given change, this S-curve has the same slope everywhere; that is, the speed at which a given change diffuses is constant, presumably because the fitness bias that favours the innovation (whether prestige, functional utility or ease of acquisition) is constant across some given language area (Kroch 1989a; b);

3. a given change may reach different geographic locations at different points in time.

The first two of these are shared with a global logistic model; however, the third is specific to a geographically weighted model. The model chosen is semiparametric; that is, not all factors vary geospatially: the effect of age is set globally, while the intercept (effectively an estimate of how early or late an innovation reaches a particular location) varies from place to place.

A semiparametric geographically weighted logistic regression was thus performed using the GWR4 software package (version 4.08) (Nakaya 2014). A fixed Gaussian kernel was used, with bandwidth selected so as to produce the optimal model in terms of AICc minimization via the Golden Search function of GWR4. This procedure yielded a bandwidth of 12.36 km. The age coefficient of this model is 0.05825 logits per year (standard error = 0.02020, z-score = 2.88382, one-tailed p-value = 0.001965).

The results of this model are most easily visualized in map form, as given in Figure 3. Here sample locations are coloured according to the earliest date at which a majority of speakers born in that year (who use cau) find it acceptable with negative concord. Henceforth, speakers who reject the cau-construction entirely are excluded from the analysis, since they are neither conservative nor innovative with respect to the innovation at hand. A local intercept of 0.000 indicates that this situation was reached in 1950 (given the recoding of dates discussed above). In fact, all locations had negative intercept scores, indicating that this situation was reached only later. Given the model’s age coefficient of 0.05825, a local intercept of –0.5825 would indicate that negative concord had become acceptable to the majority ten years later, in 1960, and so on.

Figure 3

Plot of intercept values for a GWR model of the diffusion of negative concord with cau.

The AICc of this model is 75.316. Lower AICc values indicate improved model fit even allowing for the greater complexity of a GWR model. Since this is lower than the AICc of the global model (80.467), we conclude that the geographic model performs substantially better that the comparitor global one set out above, confirming the geographic basis of the variation. Consistent with this, the model also accounts for a substantially higher proportion of the deviance in the data than the global model, 33.6% compared to 13.2% for the global model.

Looking at Figure 3, we see a diffuse, non-wave-like pattern of geospatial diffusion. Early-adopting locations, with majority negative concord by 1970, are found sporadically in both the west (around Caernarfon) and the east (around Bala and Denbigh). Subsequently, a more focused dialect distribution rapidly emerges, so that, by 1980, we find a coherent central zone with majority negative concord. The far west (the Llŷn Peninsula around Pwllheli) and the far east (around Wrexham) remain untouched by the innovation.

These patterns are rather unexpected, fitting neither a wave model (Schmidt 1872: 27; Bailey 1973: 64–109; Britain 2013: 478) nor a gravity model (Hägerstrand 1967; Trudgill 1974) of diffusion. Indeed, one significant problem is that it is difficult to identify a single locus of innovation. While the highest intercept values are in the west, on the island of Anglesey and in Caernarfon, there are also high intercept values further east. It is hardly plausible that the innovation began in the west and jumped to more easterly rural locations before spreading via contagious diffusion to the intervening areas.

Related to this is another feature of the change, namely its speed. An age coefficient of 0.058 indicates a relatively rapid increase over time. It lies at the upper end of the range of values typically found (for instance, Willis 2017: 58 finds values for age coefficients between 0.017 and 0.049 for those syntactic contexts currently undergoing change in the form of the second-person singular pronoun in Welsh). The result of this is that the change achieves a coherent dialect distribution rather quickly after the initial innovation(s).

3.5.2 Interaction between negative concord and n-words

Consider next the results for item 42 of the questionnaire, given as (27) above, testing for the availability of n-words in the subject position of cau-clauses. The raw data are given in Figure 4 in the same format as Figure 2 above. We see limited acceptance of the structure, concentrated above all around Caernarfon with a few scattered outliers elsewhere. A global logistic regression shows age to be a significant factor at the p < 0.01 level, with younger speakers more likely to accept n-words in the subject position of cau-clauses. The global model produces an intercept value of –2.06441 logits (standard error = 0.448981, z-score = –4.598061, two-tailed p-value = 4.26 × 10^–6). This means that an arbitrary speaker born in 1950 has a probability of accepting the construction of 0.113. The age coefficient is 0.043785 logits per year (standard error = 0.016870, z-score = 2.595396, one-tailed p-value = 0.004724). The AICc of this model is 75.726, and 9.6% of deviance is accounted for.

Figure 4

Distribution of the availability of n-word neb ‘no one’ in subject position with cau in the SAWD data.

Again, a GWR model is more successful at accounting for variation and allows us to see a pattern of diffusion. This was done in the same way as above. The bandwidth was kept at 12.36 km as before to ensure comparability with the model for negative concord overall, even though a somewhat higher bandwidth would have produced a slightly better fit. In practice, varying the bandwidth did not substantially alter the geospatial patterns of diffusion produced by the model.¹⁸ The age coefficient of this model is 0.041808 logits per year (standard error = 0.019316, z-score = 2.164408, one-tailed p-value = 0.015217). The AICc of this model is 74.504638, a modest improvement over the global model, and 26.2% of deviance is accounted for.

The age coefficient is very similar to that found above for the overall diffusion of negative concord (0.042 logits per year compared to 0.058), suggesting that both innovations are progressing at more or less the same rate, the age coefficient being the slope of the S-curve. Such close S-curve slopes are normally considered (as a result of the Constant Rate Hypothesis) to be evidence that the changes in question reflect a single underlying innovation (Kroch 1989a: 153–163; 1989b; 2001: 720–721 etc.).

However, the rise of this construction is rather later in time than negative concord overall. This can be seen clearly from the plot of intercept values across geographic space given in Figure 5. Compared to Figure 3, we typically see a 20-year time lag in the model in the emergence of acceptability of n-words with cau. Furthermore, the geospatial pattern of diffusion is more wave-like. A clear origin is observable around Caernarfon, and the extent to which the innovation has progressed decreases broadly in proportion to the distance from this area.

Figure 5

Plot of intercept values for a GWR model of the diffusion of the acceptability of n-words in the subject position of cau-clauses.

3.5.3 Licensing of VP-ellipsis (cau-stranding)

Recall that item 34 of the questionnaire, given in (30) above, tests for the innovation of licensing of VP-ellipsis after cau (cau-stranding). The distribution of acceptance of this structure, given in Figure 6, is even more limited, with only 10 informants repeating it faithfully in their interview. The majority of informants either added gwneud ‘do’ as a pro-VP (‘…but he’s refusing to do.’) or rephrased without the modality (‘…but he hasn’t.’). This limited acceptance makes it difficult to be confident of the age and dialect distribution of the feature and the following comments should be treated with some caution. Dialectally, acceptance is spread fairly widely, albeit at fairly low levels, across the core central area for cau.

Figure 6

Distribution of the availability of VP-ellipsis (cau-stranding) in cau-clauses in the SAWD data.

A global logistic regression shows age to be a significant factor at the p < 0.05 level, with younger speakers more likely to accept the construction. The global model produces an intercept value of –2.57803 logits (standard error = 0.54912, z-score = –4.695, two-tailed p-value = 2.67 × 10^–6). This means that an arbitrary speaker born in 1950 has a probability of accepting the construction of 0.071. The age coefficient is 0.039156 logits per year (standard error = 0.01974, z-score = 1.983, one-tailed p-value = 0.0237). The AICc of this model is 59.592, and 7.2% of deviance is accounted for.

A GWR model was performed as before with the same bandwidth of 12.36 km. This is plotted in Figure 7. For this structure, the GWR model does not perform better than the global model: the distribution of informants that accept the construction does not show sufficient spatial autocorrelation. While the GWR model performs somewhat better in terms of overall data fit, its AICc is worse (64.382223 compared to 59.592214 above) due to its increased complexity compared to the global model.¹⁹ The age coefficient of the model is 0.036776 logits per year (standard error = 0.021800, z-score = 1.686956, one-tailed p-value = 0.045811), and 19.4% of deviance is accounted for.

Figure 7

Plot of intercept values for a GWR model of the diffusion of acceptability of cau-stranding.

Note that the age coefficient for this innovation (0.037 logits per year) is somewhat lower than that for the diffusion of acceptability of n-words in subject position of cau-clauses (0.042 logits per year) and much lower than that for the diffusion of negative concord itself (0.058 logits per year). Indeed, it lies just outside the standard error of the age coefficient of the diffusion of negative concord, which suggests that this is not a manifestation of the same change. We will see later that evidence from implicational hierarchies of speaker responses supports this conclusion.

Finally, the pattern of diffusion here is weak. There is little geospatial variability in the data and, in so far as we can tell, the innovation seems to begin at much the same time and progresses to much the same extent across much of the central cau region.

3.6 Implicational hierarchies

Another way of making sense of the variation is to look at the grammars of individual speakers, to see what combinations of structures they accept. Of the 96 speakers in the northern part of the questionnaire, 19 lacked cau entirely. The distribution of the grammars of the remaining 77 speakers is given in Table 1. Numbers in parentheses refer to the number of speakers who used negative concord actively (adding it to at least one item where it was not offered) rather than passively (merely repeating it as offered). Combinations not listed were not attested.

Note that one of the advantages of the systematic investigation with a structured questionnaire carried out for a syntactic dialect atlas is that it allows grammatical systems for individual speakers to be established in this way. This would not be so readily possible for data from a spoken corpus, since most speakers present in a corpus will provide relevant data only for a subset of environments, possibly only one.

As Table 1 shows, the possible combinations within a single grammar are rather constrained. These combinations yield the implicational hierarchy (implicational scale) given in (31). This shows, for instance, that a speaker with n-words in subject position of cau-clauses always has weak negative concord with cau, while the relationship between cau-stranding and allowing n-words in subject position of cau-clauses is free and neither implicates the other.

This distribution of grammars suggests that the emergence of cau-stranding is independent of the development of negative concord, while the other innovations are all syntactically interconnected. This is consistent with the suggestive evidence from the rate of change of the GWR model developed above, in which we found the rate of change in cau-stranding to be substantially slower than in the other two cases examined.

In the next section, we will develop a formal syntactic account of the innovation consistent with these empirical observations.

4.1 Background assumptions about Welsh grammar

Declarative main clauses in Welsh manifest VSO word order. Where the inflected verb is an auxiliary, this is manifested as AuxSVO. In generative analyses, this word order is assumed to be derived by head movement of the finite verb (whether an auxiliary or a lexical verb) from V to some projection beyond TP. This position is Aux for Jones & Thomas (1977: 19–23), I (Infl) for Sproat (1985) and Rouveret (1994: 51–90), and AgrS for Willis (1998: 19–26) and Roberts (2005: 7–48). The evidence for this movement comes from the position of the finite verb and subject, both of which precede the negative marker:

Given this ordering, we must assume that both subject and inflected verb raise out of the VP beyond the polarity phrase, PolP, to a position preceding the negative marker. For the sake of concreteness, it is assumed here that this movement takes subjects, being projected initially to the specifier of an aspectual light verb projection labelled here as AspP/vP, to Spec, TP, the projection immediately dominating PolP. Consistent with this, the verb must raise via T to C. This would be a low head position, such as Fin, in a system in which C was split into multiple projections (Rizzi 1997). This system is shown in the tree in (33), representing the sentence in (32).

This commits us to the view that preverbal particles, such as the main-clause affirmative marker mi in (34), in so far as they survive in spoken Welsh, are proclitics or prefixes to the finite verb and thus do not block movement of the verb into the C-domain (cf. Willis 2007: 437).

Adverb placement suggests that some non-specific indefinite subjects remain in AspP/vP; however, this is not relevant to the current discussion. Aspectual particles such as the unmarked particle yn or the perfect particle wedi are the spellout of the head of this projection. For further detail on the system of Welsh clause structure assumed here, see Borsley, Tallerman & Willis (2007: 48–52).

Furthermore, I assume an account of negative concord in Welsh that extends and adapts Zeijlstra’s (2004) account of negative concord in a range of mostly Romance, Germanic and Slavic languages. On this approach, in Welsh, a clause is negative if Spec, PolP contains an interpretable negative operator [iPol: +NEG], generally spelled out as ddim. Other elements that participate in the negative concord system bear uninterpretable features that agree (including upwards Agree) with this element. The auxiliary bod ‘be’ will therefore be specified either as [uPol: +NEG], spelled out as dydi, if the negative operator is present, or as [uPol: –NEG], spelled out as mae, if it is not. This accounts for the difference between (3) and (4) above.

Welsh indefinites of the main neb-series (neb ‘anyone, no one’, dim byd ‘anything, nothing’ etc.) are n-words in the sense of Laka (1990: 107–9): they convey a negative meaning when used as sentence fragments, but must be accompanied by some other negative element in a negative-concord configuration in finite clauses. On the approach adopted here, these n-words are reflexes of indefinites that have undergone agreement with this operator. So, in (7), neb is the spellout of the [+human] indefinite that is [uPol: +NEG] by virtue of having undergone agreement with ddim, which is marked [iPol: +NEG] in Spec, PolP. In (8), the same process applies, except that the negative operator in Spec, PolP is not spelled out as ddim, but rather is covert. Spec, PolP is generally spelled out as ddim, but the spellout rules allow for a sequence of two negative elements to be spelled out as one: ddim neb ‘not no one’ is spelled out as neb ‘no one’ and ddim dim ‘not nothing’ is spelled out as ddim ‘nothing’. The exact conditions in which this happens and the mechanism for producing it are not immediately relevant here.

4.2 A formal analysis of the development of cau

Having laid out the relevant background assumptions, we can now consider the status of cau at the more conservative stage. In the initial grammatical system, attested in the nineteenth century, cau is a phonologically reduced form of nacáu ‘refuse’, a control verb of a familiar type. The structure is therefore biclausal, as given in (35), representing a typical example sentence such as (36).

In this grammar, cau itself has full argument structure, introducing an agent external argument, Aled, in Spec, AspP/vP in (35). By projecting an agent, it imposes the usual semantic restrictions on this argument, namely that it is volitional and thus capable of functioning as an agent. Each of the two clauses has its own independent polarity value. Formally, polarity is represented in (35) by a polarity projection PolP, whose head bears the feature [–NEG]. Negative concord on the finite verb is treated as the morphological reflex of a polarity feature on T. Cau is a semantically adversative predicate with negative connotations, comparable to adversative predicates in English such as refuse, deny, be afraid to, be reluctant to (Klima 1964: 313–14), which license negative polarity items in their complements. However, it is not grammatically negative and therefore does not trigger negative concord within its clause: polarity items are semantically licensed, while negative concord is based on a syntactic configuration. The Pol head in the main clause in (35) is therefore marked as [–NEG]. The unvalued Pol feature on the auxiliary in T agrees with this, taking the value [Pol: –NEG], before itself moving to C. The spellout of the auxiliary is therefore the [–NEG] form, mae.

Finally, the structure is a biclausal control structure. This is represented as a second predication in a complement AspP/vP with PRO as the external argument of dawnsio ‘dance’. Nothing crucial, however, depends on the exact representation of control and additional functional projections in the complement or a movement analysis of control (Hornstein 1999) could be accommodated without changing the essential intuition.

Having established an analysis for the historical input system, we can now consider the syntactic representation of the reanalysis. Reanalysis of cau as a modal is reflected in all the dialects under investigation here. The new structure is shown in (37).

In the new grammar, the structure is acquired as monoclausal. Cau is no longer a lexical verb, and does not introduce an external argument of its own. Instead, it effectively acts as a raising verb (assuming auxiliaries in general essentially to be a type of raising verb) with a modality feature (identified by its treatment as a modality head M in (37)). The external argument is merged as the specifier of vP and moves to become the subject in Spec, TP, while the finite auxiliary mae ‘is’ moves from T to C. Less structure is projected than previously, as is typical of reanalysis involved in grammaticalization. This reanalysis is purely a case of auxiliation, interpreted as the failure of acquirers to posit the entirety of the syntactic structure of their input data (loss of structure, and, in a sense, upwards reanalysis, cf. Willis 1999; 2000 on the Russian conditional; Roberts & Roussou 2002; 2003: 35–72 on the Romance and Greek future and English modals). Its surface effects are above all semantic, in that the emergent transparency of the new modal to the assignment of argument structure means that semantic restrictions on the subject are relaxed to include only those imposed by the lexical verb. Thus sentences like (2) above, with non-agentive subjects, are possible if the lexical verb is semantically consistent with such subjects. These are possible in all contemporary dialects with cau; hence, we can be confident that the reanalysed structure has been adopted in all these dialects.

Polarity here operates in the same way as before: T bears an uninterpretable polarity feature, which is valued under Agree with an interpretable polarity feature on the Pol head. In this case, the Pol head retains the [–NEG] value of the previous generation’s grammar, presumably on the basis of observed instances of affirmative rather than negative forms of the verb in clauses containing cau.

Note that, in this grammar, n-words are ungrammatical in subject position:

The structure of (38) will be the same as that in (37), with the appropriate lexical substitutions. This structure fails to spell out as (38), however, because the source of the polarity is lexical rather than grammatical in this structure. Consequently, the subject indefinite is marked as [uPol: –NEG]. Since neb is the spellout of [uPol: +NEG], it fails to be generated. Since there is no way to generate (38) in this grammar, this analysis predicts that no dialect can have n-words in cau-clauses (without a double negation interpretation). This is indeed what was found in sections 3.5.2 and 3.6 above.

Next, we turn to the grammatical system with innovated negative concord. Here, we can suggest that the negative semantic element of cau has grammaticalized, that is, gone from being represented lexically in the semantic representation of cau, to being represented in the grammatical specification of the item (cf. van Gelderen’s 2015 postulation of the diachronic linguistic cycle as a shift in feature status from semantic to interpretable to uninterpretable). Syntactically, this means that it gains an [iPol: +NEG] feature, which triggers negative concord throughout the clause. One problem with this is that it predicts that n-words, being the reflex of agreement with [iPol: +NEG] should immediately become available in such a grammar. We have seen that only a subset of speakers with negative concord allow n-words in cau-clauses. Further research is needed to establish whether this is a function of the particular sentences in the questionnaire or a wider finding, motivating either a more fine-grained analysis or a treatment in terms of the gradual actualization of the innovations, with n-words being a relatively resistant context.²⁰

What triggers this shift to integration of cau into the polarity system? On the one hand, it is of course expected from the general direction of grammaticalization and is accounted for in the same way that this pattern is accounted for more generally. However, specific to this particular instance is the syntactic analogy of other quasi-negative elements in the Welsh grammatical system. Parallel variation in negative concord exists with other negative modals and aspect markers in Welsh dialects, specifically, in (39), the negative modal methu ‘be unable’, and, in (40), the negative perfect aspect marker heb (<preposition heb ‘without’).

The dialect distribution of this feature has never been investigated, but it seems to be widely distributed geographically.