1 The canonical Welsh mutation system

Mutation involves a change in the initial phoneme of a morpheme, usually triggered by an adjacent morpheme. There are three basic forms of mutation in canonical registers, traditionally named soft, nasal, and aspirate, along with a mixed category that combines aspirate and soft mutation. The outcomes of the various mutation types are summarised in Table 1; in addition to the listed effects on consonants, certain triggers also cause h-prothesis before initial vowels (e.g., ein “our” + afal “apple” -> ein hafal).

Table 1

Canonical Welsh mutation patterns. The symbol ∅ indicates that the consonant in question disappears, while a dash (—) indicates that the consonant remains unchanged.

unmutated		soft (SM)		nasal (NM)		aspirate (AM)		mixed (MM)
orthog.	IPA	orthog.	IPA	orthog.	IPA	orthog.	IPA	orthog.	IPA
p	p	b	b	mh		ph	f	ph	F
t	t	d	d	nh		th	θ	th	θ
c	k	g	g	ngh		ch	χ	ch	χ
b	b	f	v	m	m	—	—	f	v
d	d	dd	ð	n	n	—	—	dd	ð
g	g	∅	∅	ng	ŋ	—	—	∅	∅
m	m	f	v	—	—	—	—	f	v
ll	ɬ	l	l	—	—	—	—	l	l
rh		r	r	—	—	—	—	r	r

As can be seen in Table 1, soft mutation—of the three basic forms of mutation—affects the most consonants; it also has around 30 different lexical and grammatical triggers (along with several prefixes)—roughly double the number of triggers for the other mutation types combined. Given this, and given that its triggers include all inflected personal verbs, all subject NPs, and all singular feminine nouns and articles, soft mutation accounts for the vast majority of mutations.

Mutation occurs within complex words (e.g., an- “non-” + dibynnol “dependent” -> annibynnol “independent”) but also—and more notably—across word boundaries (e.g., fy “my” + cath “cat” + gwirion “silly” -> fy nghath wirion “my silly cat”). Historically mutation arose out of phonetically conditioned alternations. Soft mutation, for instance, is chiefly a vestige of intervocalic lenition (Ball & Müller 2002: pp. 53–72). In the modern language, however, the original phonetic conditioning environments have generally long since disappeared, mainly as a result of the loss of final unstressed syllables as Old Welsh emerged from its ancestor Brythonic (Ball & Müller 2002; Willis 2010). This means, for example, that soft mutation may be triggered by morphemes ending in voiceless consonants (as with cath wirion [ka:θ ˈwırjɔn] “silly cat”). Another consequence of this process is that homophonous morphemes may trigger different mutation patterns. A good example of this is provided by the third-person possessive proclitics ei “her”, ei “his”, and eu “their”, all of which are pronounced [i] in ordinary speech.² They are distinguished, however, by the mutation of the following word. This is shown in Examples 1–3, which also serve to illustrate that the scope of a mutation trigger consists only of the immediately adjacent word.

Mutation triggers can be either lexical or morphosyntactic. Most triggers are in the former category, which includes a variety of prepositions, numerals, determiners, conjunctions, clitics, particles, and other function words. Because the original phonetic contexts triggering mutation have disappeared, lexical mutation triggers cannot be identified by form; they must be learnt individually. Nor can they be identified by morphosyntactic category. Many prepositions trigger soft mutation, for instance, but some trigger nasal or aspirate mutation, while others trigger no mutation at all.

Our survey focused on morphosyntactic triggers. Morphosyntactic triggers constitute a considerably smaller set than lexical ones;³ by their nature, however, they involve a relatively large number of lexical items (e.g., all singular feminine nouns or inflected personal verbs) so are not infrequent. Membership of a mutation-relevant morphosyntactic class is not necessarily predictable, however; aside from some suffixed words, for instance, feminine nouns in Welsh are not reliably recognisable by form.

Morphosyntactically triggered soft mutation occurs on:⁴

1. nouns following pre-nominal adjectives

2. adjectives following feminine singular nouns

3. singular feminine nouns following the definite article

4. noun phrases used adverbially

5. items directly following an overt subject NP

6. the direct object of an immediately preceding finite personal verb

7. items following an intervening phrase (e.g., a sentence modifier intervening between a verb and its object)

Such triggers form an interesting test case for looking at variation in the mutation system for a number of reasons. One reason is that their application across many lexical items means that they are frequent across sentences but abstractly defined. This has theoretical implications. It is fairly typical to treat triggers 5 and 6 as the same trigger: subject NPs, which can be overt or (especially in more formal and literary registers) not overt. This suggests that Welsh speakers should treat both environments as equivalent with regard to mutation. Some researchers have gone further and accounted for 5–7 by arguing that any c-commanding maximal projection (overt or not) triggers soft mutation (for a review and discussion of this and related accounts, see Borsley, Tallerman, & Willis 2007; pp. 226–230). If so, we should expect that users of Welsh should be equally likely to implement and to expect soft mutation in response to any such phrase. On the other hand, we might expect that some triggers are “stronger” than others, that is, more likely to trigger mutation reliably. A potentially relevant feature here is the transparency of the trigger in question, by which we mean the degree to which the trigger is easy to learn and identify, especially by L2 speakers (see Section 3.1). The adverbial use of noun phrases is, for example, a particularly opaque trigger. That is, the mutation is triggered not (as for most instances of mutation) by an adjacent overt element, but by the abstract role the noun phrase is playing in the sentence. There are also transparency differences within triggers. Natural-gender feminine nouns (like merch “girl”), for instance, can be thought of as more transparently feminine than inanimate feminine nouns (like pont “bridge”). If, as a result of change, speakers mutate less reliably, then there may be evidence of variation between triggers with respect to their “strength” or reliability, which in turn may be related to their relative transparency. If the maximal-projection account is correct, however, then relevant triggers should not vary in reliability (because, underlyingly, they are the same trigger), regardless of transparency.

In the study presented here we investigated morphosyntactic triggers 2–6 listed above. For the gender-based triggers (2 and 3) we used natural-gender nouns and inanimate nouns to investigate transparency. We also compared mutation of direct objects of finite personal verbs with overt and covert subject noun phrases so as to investigate the maximal-projection account. Our key questions are thus as follows: To what extent do Welsh speakers’ expectations of and preferences for mutation vary, to what extent does variation depend on demographic factors, and to what extent is there variation between triggers? There is certainly long-standing evidence of variation and change in Welsh speakers’ production of mutation, which we discuss in Section 2.

2 How is the Welsh mutation system changing?

The idea that the mutation system might be changing is not novel. It was stated in the November 1904 issue of the literary magazine The Athanæum (p. 621) that “If Welsh loses its mutations as South Wales is doing slightly, we shall be sorry.” Nor is change surprising. Linguistic systems are typically dynamic and subject to change. We should thus expect some degree of change in the Welsh mutation system even in the absence of contact with English. All the same, Welsh is a minority language whose speakers are almost without exception also speakers of major world languages; such levels of contact are well established to cause change in the minority language, including the simplification of complex grammatical systems (Dorian 1978; 1980; Wolfram 2004; Matras 2020).

Changes to the Welsh mutation system have been systematically documented only since the 1970s. Before then, relevant observations were to be found mostly in theses documenting specific regional dialects of Welsh and were typically strictly descriptive, with little attempt to draw broader inferences from the data (M. C. Jones 1998, p. 3). One of the first systematic approaches came with the work of M. Jones and Thomas (1977), who (among other points) noted the variability of aspirate mutation in the spoken language. They also commented on the shift from mixed to soft mutation for negated verbs and the replacement in some varieties of nasal mutation with soft mutation. Similarly, Awbery (1986) suggested that the Welsh mutation system was simplifying from a four-way system to a two-way system, whereby aspirate and nasal mutation fall gradually out of use, with soft mutation replacing them in at least some environments. M. C. Jones (1998) introduced an apparent-time dimension, finding that—in two different communities—older adults retained and used the traditional mutation system better and more consistently than younger adults.

Besides the loss of aspirate and nasal mutation in favour of soft mutation, the best-explored change to the mutation system involves grammatical gender (Thomas, 2001). The loss of grammatical gender is well-documented in many languages, and it is therefore unsurprising that grammatical gender has been a topic of interest for language change in Welsh (Gathercole & Thomas, 2005; Gathercole, Thomas, & Laporte, 2001; Hammond, 2016; Thomas & Gathercole, 2005). That this coincides with studying change in the mutation system provides an exciting opportunity for understanding change in each of the two systems in the context of the other (in line with Weinreich, Labov, and Herzog’s, 1968, embedding problem). However, it also creates a potential confound for empirical research: If speakers do not mutate a given feminine noun or adjective, is this because they do not mutate feminine nouns or adjectives or because they do not treat the noun in question as feminine?

Gathercole et al. (2001) studied the productive abilities of children aged 4–10 in Welsh-only homes, using a battery of tests to investigate children’s interpretation and use of the Welsh gender system. They found that children both overextended the mutation of adjectives to masculine contexts and underextended the mutation of nouns in feminine contexts. The authors suggested that their participants might be acquiring mutation patterns item by item rather than as generalisable rules. Thomas and Gathercole (2005) used a story-telling task to elicit semi-naturalistic speech data from Welsh adults and children. Looking at article + noun + adjective constructions, they found that adults had a relatively good productive knowledge of the mutation and gender system, producing soft mutation for feminine nouns 90% of the time. They nonetheless observed a certain amount of variability for soft-mutation, noting that even the same speaker might mutate a feminine noun in one sentence but not mutate the same noun in the next. Other speakers were found to mutate the noun after the article, but not the following adjective (e.g., y gath du for y gath ddu “the black cat”), with a production experiment indicating that only around 60% of adults mutate adjectives after feminine nouns. They further noted that a few of the adults had what were canonically soft mutated forms as base forms (dylluan and gwningen instead of tylluan and cwningen), and therefore “double mutated” the nouns when soft mutation was appropriate (resulting in y ddylluan or yr wningen instead of y dylluan “the owl” and y gwningen “the rabbit”). Gathercole and Thomas (2005) studied the production of soft mutation on native Welsh nouns, and on borrowed nouns from English. They found that participants were very reluctant to mutate borrowed nouns (see also Bellin, 1988). They also compared animate and inanimate nouns (the gender of the former being more transparent) and found that this influenced responses for some tasks.

As noted above, how gender is treated matters, as it brings with it a potentially serious confound, given that mutation constitutes the only way in which gender is reliably morphologically marked in Welsh (Hammond, 2016). This means that change in mutation behaviour in the context of gender may reflect change in the gender system, change in the mutation system, or both. This is exemplified in a study by Hammond (2016) showing through statistical natural language processing techniques that soft mutation is the only remaining factor that allows learners to disambiguate grammatical gender in Welsh. Hammond argued that the weakening of the mutation system will mean that learners guess masculine more and more often, eventually leading all nouns to move into the masculine category, removing the functional utility of mutation as a marker for gender altogether. The decline in use of grammatical gender and decline in use of mutation are thus closely connected.

As well as addressing this question in their 2005 work by comparing responses to animate and inanimate nouns, Thomas and Gathercole (2007) suggested that to understand the acquisition of the grammatical gender system, we must look at the acquisition of the mutation system in non-gender-related contexts. By comparing results from studies on mutation in these contexts (e.g., Ball, 1988; Ball & Müller, 2002) to studies on the grammatical gender system, it becomes possible to compare the two. While this is useful in allowing us to get a global idea of how the trajectories of these two systems compare, the comparison is not ideal as non-gender-related mutation triggers differ in frequency, salience, and complexity from gender-related triggers. It is very possible that some mutation subsystems are changing less rapidly than others.

Taken together, production studies on the acquisition and use of mutation in Welsh suggest that the system is undergoing change: aspirate and nasal mutation are declining in use, and the use of the soft mutation system is becoming more variable, including for many adults. However, some evidence, such as Thomas and Gathercole’s (2005) and Gathercole et al.’s (2001) over-mutation finding, suggest that, instead of simply declining in use, the system may instead be undergoing systematic restructuring and simplification. This can be better understood if we understand the reasons for change in the system.

3 Why is the Welsh mutation system changing?

In bilingual contexts where the utility of the minority language is in decline due to a prestige language like English, simplifications of grammatical structures is common, especially if they are not shared with the prestige language (e.g., Dorian, 1978 on Scottish Gaelic). Welsh speakers are almost categorically bilingual, and since Welsh is the minority language it fits this profile well. Additionally, the Welsh mutation system is complex to learn and carries a low functional load. Simplification of the mutation system should therefore not be surprising. Section 3.1 discusses simplification of linguistic structures in bilingual contexts. Section 3.2 discusses the present sociolinguistic situation in Wales. Section 3.3, finally, explains how the learnability and functional load of the mutation system also contribute to why it is changing.

3.2 The health and sociolinguistic situation of Welsh

Although certainly a minority language, Welsh is by far the healthiest of the Celtic languages. It is the only one not listed by the UNESCO Atlas of World Languages in Danger (UAWLD) as endangered, being listed merely as vulnerable (Moseley, 2010). The 2013–2015 Welsh-Language Use Survey found that 318,800 people in Wales (11% of the population; 47% of the total number of speakers, as estimated by the same survey) considered themselves fluent speakers of Welsh, while a further 21% of speakers considered themselves to know “a fair amount”; 360,000 (13% of the population of Wales) claimed to speak Welsh daily; only 5% of speakers claimed to never use the language. This contrasts clearly with the situation of all the other Celtic languages. The next healthiest is Irish, for which the 2016 Irish census (Central Statistics Office, 2020) reported that only 10% of speakers used Irish at least weekly, while 23.8% said that they never speak it.

Nonetheless, there remains cause for concern for the future of Welsh. Figure 1 shows the proportion of the Welsh population that speaks Welsh between 1891 and 2011. In spite of a recent plateau, the overall trajectory since 1891 has very clearly been downwards. The number of fluent speakers as recorded in the 2013–2015 Welsh-Language Use Survey was lower than in the 2004–2006 survey (although the number of non-fluent speakers had increased).

Figure 1

Welsh speakers as percentage of Welsh population (including all speakers and monoglot speakers) based on census data for every decade between 1891 and 2011 (no census was conducted in 1941). To save space the y axis runs from 0 to 60%.

Not surprisingly, decline in the use of Welsh is accompanied by an increase in the use of English. With the exception of some preschool children and dementia sufferers with second-language attrition, all Welsh speakers in Wales also report speaking English. Over 50% of Welsh speakers lack a parent who speaks Welsh fluently, and a little over 50% of speakers report being more comfortable speaking English than Welsh. This situation poses a danger to the long-term survival of Welsh, and a more immediate danger to the survival of complex grammatical systems such as mutation.

The main counter-pressure to these demographic forces is education. All school pupils attend Welsh classes, and in the 2019/2020 school year 108,638 school pupils (23%) were educated in other subjects through the medium of Welsh (StatsWales, 2020).⁵ Moreover, the success of Welsh-medium education and the Welsh government’s commitment to supporting the language mean that the trend is for more Welsh-medium schools to open (or for traditionally English-medium schools to introduce more Welsh instruction) rather than the reverse. While this is a positive trend in terms of not only maintaining but even increasing the number of Welsh speakers, it has the inevitable consequence of increasing the proportion of Welsh speakers who acquired the language in school rather than at home. As noted above, less than half of Welsh speakers have a parent who speaks the language fluently.

Differences between earlier and later learners of minority languages have been noted elsewhere. In a study of the revitalisation of Scottish Gaelic, Nance (2015) showed that young learners (or “new speakers”) of Scottish Gaelic sound different from “traditional speakers” of the language. She concludes that new speakers use different forms of the language and use the language for different communicative reasons than traditional speakers.

This work is supported by game theoretic models of minority language shift that show that when bilinguals of a majority and minority language communicate they tend to use the majority language if they are unsure about whether their interlocutor is bilingual, even if they are motivated to use the minority language as often as possible (Iriberri & Uriarte, 2012; Uriarte & Sperlich, 2021). Within bilingual populations, speakers converge to a stable equilibrium in which many bilinguals shift to use the majority language even if the minority language would have been communicatively available for use. Uriarte and Sperlich (2021) showed that this communicative convention aligns with data on the daily use of Welsh as determined by The Census of Wales 2004–06 and 2013–15.

There are two main routes for young L2 Welsh learners to acquire mutation: through explicit instruction in school and through exposure to conservative varieties of Welsh. The first is straightforward: all school pupils in Wales have Welsh classes, in which the mutation system is taught explicitly at least to some extent. The second is more complex; in some regions exposure to conservative speakers may be rare. Formal literary Welsh would provide the most consistent exposure to canonical mutation patterns, but—even assuming reliable exposure to literary Welsh on the part of young learners—it differs substantially from colloquial registers and in that respect provides a poor model for them (cf. Ball, 1992). Less formal standardised varieties, as used in broadcasting by the BBC and the Welsh-language television channel S4C, provide better and more accessible models. However, broadcast Welsh varies considerably not only in register but also in the extent to which canonical mutation patterns are observed. (It is notable in this context that subtitles on BBC videos often “correct” non-canonical mutation patterns).

Regional variation may be relevant here. Welsh is divided into a number of regional dialects, with the clearest distinction being a north–south divide. Although, according to the 2011 census, roughly 58% of Welsh speakers are located in the more densely populated south, the areas with the highest proportions of Welsh speakers are located in the north, and the north of Wales is typically perceived as being more Welsh-speaking. Some work has related changes in the mutation system to specific regional varieties. Awbery (1986), for instance, treated simplification of the mutation system—realised as the loss of aspirate and nasal mutation accompanied by the spreading of soft mutation to new contexts—as a feature of (some) southern dialects. However, simplification also occurs in northern varieties, and it is not clear whether the apparent association with southern varieties might be better accounted for in terms of other sociological variables, such as parental fluency or age of acquisition, which also vary regionally.

In summary, the bilingual sociolinguistic situation in Wales seems likely to explain a great deal of the variability observed in studies like that of Gathercole et al. (2001). In the perception study we report in this paper we thus expected to find more variable acceptability judgements from younger “new” speakers, and more conservative judgements from older “traditional” speakers (Dorian, 1994; Nance, 2015; O’Rourke & Pujolar, 2015). It may also be (though it is less certain) that there are regional differences between north and south.

4 Predictions

Previous production studies on the Welsh mutation system have found variability in the system, which is in line with long-standing literature on language change in bilingual minority-language contexts (Dorian, 1978, 2010; Dressler, 1972; Lavandera, 1978). Our study is an online acceptability judgement survey focusing on respondents’ perceptions and expectations of mutation in Welsh.

We had several predictions. The first concerns demographic patterns. Although change occurs in the absence of contact, we should expect the influence of bilingualism to be key to understanding change and variability in the modern Welsh mutation system. This leads us to predict that respondents’ L1 status (i.e., whether they consider Welsh or English, or both, to be their first language) should play an important role in explaining variability in our data. In line with work on Scottish Gaelic by Nance (2015), among others, we expect that L1 Welsh speakers will show more conservative acceptability judgements than L2 Welsh speakers and bilinguals. Other demographic factors that can be expected to play a role include age and regional background. We should expect age to matter if there is a change in progress, with older speakers preferring more conservative mutation patterns to younger speakers. Finally, we should expect regional background (i.e., whether a respondent is from the north or the south of Wales) to matter (a) if there is regional variation in the mutation system, as suggested by Awbery (1986), and (b) because Welsh speakers make up a larger proportion of the population in the north of the country.

Our second prediction concerns the nature of the mutation trigger. The maximal-projection account, as discussed in Section 1, holds that a number of potentially distinct triggers are in fact examples of the same trigger, specifically c-commanding maximal projections (Borsley et al., 2007; Tallerman, 2006). This implies that triggers that can be treated as falling under this category should be treated the same by speakers. Alternatively, it could be that the triggers involved are learnt and treated as distinct, at least by many speakers. To get at this we included sentences in our data in which direct objects directly follow inflected personal verbs with no overt subject and sentences in which they directly follow an overt subject. If responses differ for these two sentence types, that suggests that these may operate as distinct triggers.

Our third prediction concerns trigger transparency. Given the demographic prediction above and in the context of ongoing change in the mutation system, it seems likely that not all triggers are changing to the same extent. One possibility is that this is due to the functional load of the trigger (Boon, 2014). Another (though not mutually exclusive) possibility is that this is due to the transparency of the trigger. That is, how straightforward is it to learn and identify the trigger, especially for L2 speakers? We investigated this with respect to gender triggers by including natural-gender nouns (where natural gender matches grammatical gender and may aid recognition of the latter) and inanimate nouns. The natural-gender nouns, we anticipate, should—by virtue of greater transparency—be a more reliable trigger than the inanimate nouns. We also included sentences with adjectives following natural-gender nouns; since the adjectives themselves are gendered only on the basis of the noun being modified (and thus often occur unmutated in the same grammatical context, i.e., directly following a noun) we predicted that this would be an even less reliable trigger. For comparison, we also included adverbial noun phrases, already known as a variable trigger (even in canonical contexts), and one that is very abstractly defined.

We took two approaches with our analysis. With respect to our first prediction, we looked at the entirety of the data and tested the role of different demographic factors using a linear mixed-effects model. For the second and third predictions, which required that we separate out data for different triggers, we did not have enough data for any given trigger for the linear-model analysis to be appropriate. In this case we therefore performed a k-means clustering analysis to identify patterns in the data for comparison. The goal was primarily to provide preliminary observations as a basis for future investigation focusing on specific triggers with larger samples.

6 Results

Data are available at https://osf.io/hqrz6/. In what follows we will start by presenting an analysis of general patterns in the data, collapsing all triggers together (Section 6.1). This will include both a clustering analysis (Section 6.1.1) and a linear-model-based analysis (Section 6.1.2). We will then present a clustering analysis of each trigger separately (Section 6.2).

6.1 General patterns

Our first analysis was an analysis of the overall non-trigger-specific patterns in which we collapsed respondents’ responses to the different sentences, and simply asked how strong respondents’ preference for canonical mutation patterns was, both when mutation is canonically required and when it is not.

6.1.1 Clustering analysis

Figure 2 displays mean extended non-normativity scores for all respondents across sentence types. We used shape and colour to indicate the language that respondents reported as their first language. The x-axis of this graph represents extended non-normativity scores for sentences in which mutation would be expected (e.g., feminine nouns following the definite article). The leftmost end of this axis denotes a strong preference for mutating canonically, while the rightmost end denotes a strong preference for “under-mutation”, that is, omitting mutation that would canonically be required. The y-axis represents extended non-normativity scores for sentences in which non-mutation is canonically required (e.g., masculine nouns for the gender-based triggers). The bottom end of the axis denotes a strong preference for canonical non-mutation, while the topmost end denotes a preference for “over-mutation”, that is, extending mutation to a non-canonical context.

Figure 2

Respondents’ overall mutation preferences; colours and shapes indicate first language. Each point on the graph represents a respondent’s mean extended non-normativity score. Respondents in the bottom-left prefer canonical mutation patterns. Respondents in the bottom-right do not mutate in canonical contexts. Respondents in the upper-left do mutate in non-canonical contexts. Respondents around the centre of the graph show variable preferences.

This axis structure gives us a graph that can be divided into four quadrants. Respondents who prefer canonical mutation patterns should fall into the bottom-left quadrant of this graph (the “Conservative” quadrant). Speakers who use very little mutation in general should fall into the bottom-right quadrant (the “Under-mutate” quadrant). Speakers who over-mutate, such as mutating masculine nouns after the article, should fall into the top-left quadrant (the “Over-mutate” quadrant). Speakers who vary in whether or not they mutate canonically should fall in the middle of the graph, around the centre-crux of the quadrants (the “Variable” area). We should not expect the top-right quadrant of this graph to be occupied by many speakers, since it denotes a preference for not mutating when mutation is canonically expected and applying mutation when it is not canonically expected; such a direct reversal of canonical patterns would be surprising.

We plotted responses for our two questions (Q1. Would you mutate like this yourself? Q2. Would you be surprised if someone else mutated like this?) separately. We expected responses to Q1 and Q2 to be rather similar but for respondents to have stronger opinions about their own production (Q1) than about the use of mutation by others (Q2). This was the pattern we found. (See supplementary materials for overall responses to Q2 at https://osf.io/hqrz6/.) Therefore the analysis in the remainder of this paper will focus primarily on Q1 data.

The clustering analysis identified two clusters in the overall Q1 data, which are displayed in Figure 3.⁷ This figure shows the same data as Figure 2, except that colour is now used to distinguish clusters (to which we also fitted ellipses based on a 95% confidence level for a multivariate t-distribution). One cluster is located in the bottom of the graph, mostly in the bottom-left Conservative quadrant, though including a few respondents with some tendency for under-mutation (all located towards the bottom left of the under-mutation quadrant, indicating that this tendency is not strong). Broadly speaking, respondents in this cluster prefer canonical mutation rules, rarely over-apply mutation, and only occasionally under-apply it. The second cluster is a diagonally oriented cluster of respondents running through the Variable middle section of the graph. This cluster is smaller than the other cluster, but not very much smaller. It constitutes 42% of the data, an indicator of the extent of variability in the current mutation system. In summary, if all triggers are considered together, our respondents tend to be either rather conservative (i.e., canonical) in their use of mutation, or variable. Very few respondents consistently over-mutate or under-mutate. But are there demographic patterns behind this result?

Figure 3

Respondents’ overall mutation preferences with k-mean clusters indicated by colour and ellipses; shapes indicate first language.

6.1.2 Linear model analysis

We ran a linear mixed effects model on the data from Q1 with Sentence Type, Age, Gender, Region (North vs. South), and First Language as fixed effects. We found no effect of Age (β = –0.005, d = 0.002, p = 0.12), Gender (β = –0.04, d = 0.02, p = 0.85), or Region (β = –0.06, d = 0.03, p = 0.93) on people’s mutation preferences. There was, however, a significant effect of First Language with respondents who learned Welsh first differing significantly from respondents who learned English first (β = 0.50, d = 0.23, p = 0.01). This suggests that speakers whose first language was Welsh were more conservative, while speakers who grew up with English as their first language were more likely to accept non-canonical mutation patterns. Bilingual speakers’ responses did not differ from those of respondents who learned English first (β = 0.18, d = 0.08, p = 0.41) or from those from respondents who learned Welsh first (β = –0.32, d = 0.14, p = 0.07). The dependent variable was the Non-normativity score (Section 5.4) for which lower values indicate a greater preference for canonical mutation patterns and higher values indicate a greater preference for non-canonical mutation patterns. The mean scores for different mutation patterns can be seen in Table 4. The overall mean score was 1.52, close to the middle of the range. There are, however, important differences consistent with the cluster analysis discussed above. For instance, the Non-normativity score is 0.75 points lower for canonically unmutated sentences than for over-mutated sentences (β = 0.75, d = 0.33, p < 0.001), for instance, and the score for canonically mutated sentences is 0.98 points lower than for under-mutated sentences (β = 0.98, d = 0.44, p < 0.001). In general, under-mutation was preferred to over-mutation. This is suggested by the lower Non-normativity scores in Table 4 for sentences that are canonically unmutated and the higher scores for sentences with canonical mutation. It can also be seen in Figure 2. By contrast, the higher scores for canonically mutated sentences and under-mutated sentences suggest that when mutation was expected canonically, speakers were rather accepting of sentences without it. This is consistent with an account in which mutation is, or is becoming, variable, perhaps (though we cannot infer this from our data) conditional on register.

Table 4

Non-normativity scores for different mutation patterns (collapsed across triggers) for Q1. Non-normativity scores range from 0 to 4, where 0 indicates a strong preference for canonical mutation and 4 indicates a strong preference for non-canonical mutation patterns.

Mutation pattern	Non-normativity score
Not mutated canonically/ unmutated in experiment	0.90
Not mutated canonically / mutated in experiment	1.64
Mutated canonically / mutated in experiment	1.28
Mutated canonically / unmutated in experiment	2.25

The above results are all based on data from Q1, which asked whether respondents themselves use the mutation pattern they were presented with. We also generated the same graphs and results for Q2 (see the appendix in the supplementary materials at https://osf.io/hqrz6/), which asked whether respondents would be surprised if a conversation partner used the heard mutation pattern. We expected more variation for this question, and this is indeed what was found. Respondents are located less towards the corners of the plot and more towards the Variable central area. This suggests that respondents are used to hearing variable mutation, and are not surprised to hear non-canonical mutation patterns. This demonstrates that even speakers who report conservative mutation patterns in their own language are quite familiar with non-canonical patterns from others, especially under-mutation.

We ran the same model on the Q2 data, and again found no evidence for an effect of Gender (β = 0.03, d = 0.02, p = 0.74), Region (β = –0.05, d = 0.04, p = 0.63), or Age (β = –0.007, d = 0.005, p = .059) on respondents’ responses. There was an effect of First Language suggesting that L1 Welsh speakers were more conservative than L2 speakers (β = 0.32, d = 0.23, p = 0.019). Again respondents who grew up in a bilingual home did not differ from respondents who learned English first (β = 0.16, d = 0.11, p = 0.28) or those who learned Welsh first (β = –0.16, d = 0.11, p = 0.20) Again we found significant differences between responses to canonical non-mutation and to non-canonical over-mutation (β = 1.34, d = 0.95, p < .001) as well as between canonically mutated and non-canonically unmutated sentences (β = 1.98, d = 1.4, p < 0.001). Overall, results suggested very high familiarity with non-canonical mutation (see graph in the appendix in the supplementary materials, available at https://osf.io/hqrz6/).

6.2 Trigger-specific results

The above analysis collapsed different mutation triggers together, but examination of the data suggested interesting differences between responses to different triggers. we therefore split the data out by trigger and analysed each of them in isolation. Since there is not enough data per trigger to conduct an analysis on the data using the linear model, we provide only a clustering analysis. We also used only Q1 data for the trigger-specific analyses, since they apply more directly to respondents’ own use of mutation. In all of the graphs for the different triggers (Figures 4, 5, 6, 7, 8, 9) we have numbered the respondents so that they can be tracked between triggers; we also use shape to indicate first-language status (the only significant demographic predictor of responses in our linear model).

Figure 4

Preferences regarding mutation of definite natural-gender singular nouns. X-axis indicates respondents’ acceptance of under-mutating feminine nouns. Y-axis indicates acceptance of non-canonically mutating masculine nouns. Each numbered point in the graph represents a single respondent, coloured according to which of the seven clusters they were identified as belonging to. Shapes indicate first language status. Ellipses indicate clusters from on an alternative two-cluster analysis (ranked similarly to the seven-cluster analysis.

Figure 5

Preferences regarding mutation of suffixed definite singular nouns. X-axis indicates respondents’ acceptance of under-mutating feminine nouns. Y-axis indicates acceptance of non-canonically mutating masculine nouns. Colours indicate clusters. Shapes indicate first language status.

6.2.1 Natural-gender singular nouns following the definite article

Singular feminine nouns are canonically mutated after a definite article, but masculine nouns are not. Non-normativity score results are shown in Figure 4 using the same plot structure as for Figure 2 in Section 6.1. The silhouette analysis identified seven clusters in the data (indicated using colour in Figure 4), though an alternative two-cluster analysis (indicated using ellipses) was almost as highly ranked. The two cluster analysis is similar to the two clusters identified for the overall data (Figure 3), with a cluster of Conservative speakers in the bottom left and a diagonally oriented cluster of Variable speakers. The seven-cluster analysis breaks this down into smaller subsets, including a cluster of respondents in the top right of the Conservative quadrant, bridging the gap between the more conservative clusters in the bottom left (of which there are three) and the central variable cluster. Overall, this trigger could be characterised as a relatively strong or reliable trigger that is still used canonically by many speakers. This is suggested by the fact that four of the seven clusters, accounting for over 50% of respondents, are primarily located in the Conservative quadrant. Aside from these, the largest cluster is the variable cluster in the middle of graph. The Over-mutation cluster has only four members, and there are similarly few participants in the Under-mutation quadrant.

6.2.2 Suffixed singular nouns following the definite article

This trigger is the same as the preceding trigger (natural-gender nouns), except that the grammatical gender is less transparent. In these sentences, the nouns are inanimate, though they are morphologically marked (with the masculine suffix -yn or the feminine suffix -en, both of which typically contribute diminutive or singulative meaning). The cluster analysis identified three main groups of respondents. The two largest are a Conservative group in the bottom left (comprising 39.5% of respondents) and a Variable group in the center (comprising 37% of respondents). The most notable difference from the natural-gender trigger is the presence of a rather more substantial cluster of under-mutating respondents in the bottom right quadrant. This is consistent with the expectation that attitudes towards mutating feminine nouns should vary according to how obvious the gender of the noun is, and that the gender of natural-gender nouns is likely to be more obvious than that of nouns marked by an arbitrarily gendered suffix. Unlike for natural-gender nouns, there is also much more evidence of under-mutation than of over-mutation.

There are different possible interpretations of this result with regard to the dynamics of gender and mutation in Welsh. One is that the language might be shifting to a natural-gender system, in which nouns with inanimate referents are less likely to be perceived as feminine. It has previously been reported that such a change may indeed be underway for Welsh, at least for some speakers, with masculine pronouns being widely used to refer to all inanimate anaphors (B. M. Jones, 1993; M. C. Jones, 1998). This is also discussed by Thomas and Gathercole (2005), who did not find evidence for a shift towards a natural-gender system among speakers in northwestern Wales. An alternative interpretation of our data is that, to the extent that grammatical gender still operates in Welsh, there is uncertainty and variation as to which nouns are considered feminine, with animate nouns being the most stable. This second interpretation is not mutually exclusive with the first interpretation, and could be taken to constitute an early stage in the loss of grammatical gender. The difference in responses to natural-gender and inanimate nouns, however, suggests that mutation of feminine nouns (but not masculine nouns) following the definite article remains a feature of the Welsh mutation system for at least some speakers and that it varies to some extent independently of variation in the identification of gender.

6.2.3 Adjectives following indefinite natural-gender nouns

This trigger concerns mutation of adjectives after feminine nouns. This trigger does not rely on the presence of a definite article but affects adjectives directly following any feminine noun. In order to restrict the sentences to having a single mutation context, we only used unmutated indefinite nouns (all natural gender). The trend we saw for natural-gender and inanimate nouns, in line with work by Thomas and Gathercole (2005), was consistent with the expectation that mutation preferences might be influenced by gender transparency, with inanimate feminine nouns less likely to be mutated than animate ones. This trigger can be seen as less transparent still, as the gender of the adjective is dependent on that of another (in this case) unmutated, word. If this is right, we should expect respondents to be even less likely to prefer canonical patterns for gendered adjectives than for inanimate nouns.

This seems to be borne out in the data, as shown in Figure 6. The cluster analysis revealed five clusters of respondents. The most Conservative cluster is not only smaller than for inanimate nouns but also spread out horizontally in the bottom-left quadrant, with far fewer respondents in the bottom left corner than for either natural-gender or inanimate nouns. There is, however, a more substantial “Variable Conservative” cluster located in the top right of this quadrant (accounting for 28% of the data). There is a further Variable cluster located in the right side of the graph and a somewhat substantial cluster of under-mutating respondents in the bottom-right quadrant. There is no cluster of Over-mutating respondents. This is all consistent with this being a trigger for which the trend is to mutate less. Taken together with the other triggers, the data are consistent with the existence of a positive relationship between trigger transparency and preference for mutation. It is important in this regard to bear in mind that, for the adjective trigger, we looked only at adjectives following unmutated indefinite nouns. It is possible that we would have found different results for adjectives following mutated definite nouns, for which the mutation of the noun might prime mutation of the adjective.

Figure 6

Preferences regarding mutation of adjectives following masculine unmutated and feminine mutated nouns. X-axis indicates respondents’ acceptance of under-mutating feminine adjectives. Y-axis indicates acceptance of non-canonically mutating masculine adjectives. Colours indicate clusters. Shapes indicate first language status.

6.2.4 Direct objects of inflected verbs

Direct objects directly following inflected personal verbs are canonically mutated whether or not an overt subject intervenes. Many accounts treat overt and covert subject NPs as the same trigger, and some go further and treat any c-commanding maximal phrasal projection as a soft mutation trigger (Borsley et al., 2007; Tallerman, 2006). If responses vary depending on the presence of an overt subject, this might shed interesting light on how the mutation rule is encoded.

Results for sentences without an overt subject can be seen in Figure 7. For this trigger, over-mutation is impossible as mutation is always canonically required, so the quadrants have different meanings from the quadrants in Figures 4, 5, 6. The Conservative quadrant is now in the bottom-right; the top-right quadrant indicates speakers who are Variable and who accept both mutation and non-mutation for this trigger. The top-left quadrant contains respondents who prefer under-mutation and the bottom-left quadrant contains respondents who gave low ratings both to sentences with mutation and to sentences without mutation. The axis labels also reflect the wording on the original Likert scales.

Figure 7

Preferences regarding mutation of direct objects following an inflected verb without an overt subject. Respondents in the bottom-right quadrant prefer conservative mutation. Respondents in the top-left quadrant prefer under-mutation. Respondents in the top-right quadrant are variable. Colours indicate clusters. Shapes indicate first language status.

The cluster analysis identified eight clusters in total, but the most striking feature of the graph is that the vast majority of respondents are located in the right half of the graph, with only six respondents (14%) clearly located in the left half. In fact, there is only one small cluster (of four respondents) wholly located in that half, specifically in the right half of the Under-mutation quadrant. In other words, the vast majority of respondents can be identified as Conservative or Variable, suggesting that this trigger remains rather strong. There is a cluster of five respondents located in the bottom-right of the Conservative quadrant, suggesting particularly canonical preferences. This is suggestive. However, it is important to note that sentences without overt subjects are rather more common in formal and literary registers (where canonical mutation is to be expected) than in colloquial registers, with the consequence that speakers are somewhat less likely to encounter non-mutation with this trigger in everyday life. Prescriptive norms are also likely to play a more important role here, so differences here should be treated with caution.

Figure 8 shows the results for sentences with an overt subject. The cluster analysis identified seven clusters in this case. Overall, however, the pattern is very similar to the pattern in Figure 7. There are two main differences. The first is that, for sentences with overt subjects, the cluster of under-mutating respondents is located further to the left of the graph, suggesting a greater preference for not mutating. The second difference is that the Conservative respondents tend to be located closer to the midlines, with only one respondent in the bottom-right corner (as compared with the five-respondent cluster located in this region for the overt-subject trigger). In other words, respondents seem overall more accepting of non-mutation after overt subjects than after inflected personal verbs with no overt subject, suggesting that—at least for some speakers—these might not behave as the same trigger, contrary to many accounts. It should also be noted, however, that personal inflected verbs with overt subjects are considerably more common in colloquial registers than verbs without them. This variation between registers may well account for at least some of the differences observed.

Figure 8

Preferences regarding mutation of direct objects following an inflected verb with an overt subject. Respondents in the bottom-right quadrant prefer conservative mutation. Respondents in the top-left quadrant prefer under-mutation. Respondents in the top-right quadrant are variable. Colours indicate clusters. Shapes indicate first language status.

One other cluster of six respondents is worth commenting on. This is located around the lower half of the vertical midline. These respondents have a preference for non-mutation in this context (as indicated by their low position on the y-axis) but seem also unsure about applying mutation (as indicated by their position in the middle of the x-axis). If the verb form should be considered the primary mutation trigger in this context, sentences with intervening overt subjects can be seen as involving less direct triggers, analogous with adjectives.

6.2.5 Adverbial noun phrases

Adverbial noun phrases were our final trigger. According to canonical mutation rules, the first element in a noun phrases (e.g., mis nesaf “next month”) is mutated when it is used adverbially (fis nesaf). Historically this trigger was sufficiently consistent that certain noun phrases which occur only rarely outside an adverbial context (e.g., ddoe “yesterday”) are unfamiliar to many speakers in their unmutated form. However, the trigger is also particularly abstract, being essentially triggered by the role of the noun phrase, not by any overt morpheme; furthermore, it is canonically considered to be optional if sentence initial (Williams, 1980), and in all positions is more observed in written registers than colloquial ones. We therefore expected respondents to be rather tolerant of nonmutation for this trigger, and perhaps to be less accepting of mutation.

Figure 9 shows the results for this trigger, and a strikingly different pattern is in evidence compared with the other triggers. The cluster analysis identified two main clusters, one mainly located in the Under-mutation quadrant and the other mainly located in the Variable quadrant. There are few participants located in the Conservative quadrant (and all are identified by the cluster analysis as belonging to the Variable cluster). None are located in the most conservative corner of the quadrant. The pattern, in other words, is for variability or non-mutation, consistent with the view that this trigger is a relatively “weak” one, for which the vast majority of respondents are comfortable with non-mutation. To a small extent this may by influenced by the position of the noun phrase in question. Respondents gave their highest approval (with a mean Non-normativity score of 0.898) to sentences with an unmutated sentence-initial noun phrase, which is consistent with mutation in such sentences being canonically optional. However, sentences with mutation received similar ratings wherever they occurred (1.39 for sentence-initial and 1.61 for non-sentence-initial noun phrases; unfortunately data are too sparse for statistical comparison), and it is clear from the pattern of data in Figure 9 that the overwhelming picture was simply of high acceptance of not mutating.

Figure 9

Preferences regarding mutation of adverbial noun phrases. Respondents in the bottom-right quadrant prefer conservative mutation. Respondents in the top-left quadrant prefer under-mutation. Respondents in the top-right quadrant are variable. Colors indicate clusters. Shapes indicate first language status.

7 Discussion

We used an online survey to investigate Welsh speakers’ perceptual acceptability judgements of mutation and non-mutation across a variety of morphosyntactic triggers. We compared judgements of (a) sentences with canonical mutation patterns, (b) sentences with under-mutation (i.e., no-mutation where mutation is expected according to canonical rules), and (c) sentences with over-mutation (i.e., mutation where not canonically expected).

We had three main predictions. The first prediction was that demographic factors, primarily L1 status, but also age and region, would explain variation in the data. In fact, only L1 status had a significant effect. This was also consistent with our cluster analysis. The overall pattern, taking all triggers together, was of two similarly sized clusters of respondents: one Conservative cluster who tend to prefer canonical mutation patterns and a Variable cluster accepting of a range of different patterns. The Conservative cluster was dominated by respondents who reported having grown up speaking Welsh as a first language, and who made up 80% of this cluster (compared with 30% of the Variable cluster). Only one respondent (4%) in the Conservative cluster spoke English as a first language and only two gave both English and Welsh as first languages, compared with seven (39%) and five (28%) respondents respectively in the Variable cluster. This suggests a rather clear (and not surprising) picture: Change and variation in the Welsh mutation system seems to be primarily driven by the use of English in Wales. This general conclusion aligns with prior studies in bilingual minority-language contexts that bilingual and L2 speakers of the minority language use accept much more variability in the use of a complex language system like mutation than more traditional L1 speakers do (Dorian, 1994; Nance, 2015). This literature also led us to expect to find a similar effect for age, but we did not find one in our sample.

If we consider the mutation preferences of the Variable cluster of respondents, we find a preference for under-mutation as compared with over-mutation. The minority-language change literature suggests that complex grammatical systems can simplify either by losing features or by extending them to new contexts (i.a. Maher, 1991; Silva-Corvalán, 1986). The preference of Variable respondents for under-mutation compared with over-mutation suggests that in Welsh, omission is more common than extension. This is particularly relevant to the gender system, for which soft mutation is the only useful cue. Hammond (2016) suggested that inanimate feminine nouns are becoming masculine, and our results are consistent with this.

Our second prediction concerned direct objects directly following overt and covert subjects of inflected personal verbs. This is typically treated as the same trigger, and what we have called the maximal-projection account groups the two contexts together with other contexts to treat all c-commanding maximal-projections as triggers of soft mutation (Borsley et al., 2007; Tallerman, 2006). Contrary to this, we predicted that the presence of an overt subject NP might make a difference, at least for some speakers, implying that they treat these as distinct triggers. Our cluster analysis did indeed suggest that these two contexts are not treated identically by our respondents. This result must be treated with a certain degree of caution, however. Covert subjects in Welsh are more a feature of written or formal than of colloquial registers, so it may be that register is playing an important role here. On the one hand, if these triggers are acquired as distinct, then that may be partly a consequence of register-based differences in their distribution. On the other hand, it may be (as an anonymous reviewer suggested) that some respondents’ judgements partly reflect judgements of register-appropriateness rather than the appropriateness or naturalness of the mutation per se. While we sought to avoid this kind of effect with our instructions, it remains a real possibility, and a concern for future work.

Our third prediction concerned trigger transparency. We predicted that the remaining triggers would be ordered as follows with respect to their reliability, that is, their likelihood of eliciting more conservative responses:

natural-gender nouns > inanimate nouns > adjectives following nouns

The results of our clustering analysis were indeed consistent with this. The least conservative responses were also associated with the least transparent trigger, adverbial noun phrases, although this trigger exhibits variability even in formal registers (being optional sentence-initially). In future work researchers might like to focus on comparisons of frequency, transparency, and functional load (cf. Boon, 2014) in predicting trigger-dependent patterns of variation and change.

There are three limitations to this study that do present options for future work on this topic. First, our results on individual triggers must necessarily be considered somewhat preliminary and exploratory. While we consider our clustering analysis to have been enlightening, the amount of data for each trigger was relatively small, limiting our statistical options. We consider our results to provide a suggestive and intriguing basis for future work targeting particular triggers and variables with more power.

A second limitation concerns the stimuli. While the speaker, a Welsh-speaking linguist, was careful to aim for a natural style, as judged by the authors (and many sentences were recorded more than once so as to improve naturalness), recruitment difficulties prevented us from piloting the sentences with independent samples. This should be a consideration for future work, as should the use of a diversity of voices with different demographic characteristics.

The third limitation has more to do with the respondents themselves. As an anonymous reviewer pointed out, respondents may vary in their linguistic awareness and, consequently, in their ability to judge the appropriateness of mutation as distinct from other aspects of the sentence. Again, this is a place where future work—by asking questions in a wider variety of ways—may make progress. It is worth adding that we also provided our own respondents with prompts for open-field responses. In many cases their responses are somewhat enlightening, and we discuss them in what follows.

Several responses are relevant to the importance of L1 status. A number of respondents, for instance, indicated that they thought education and exposure to “correct” mutation patterns in the home likely play an important role. Many suggested that exposure to canonical mutation patterns at home is crucial as mutation is not formally taught until secondary school, by which time it is harder to learn. One respondent stated, for instance:

I learnt the standard patterns in secondary school (rather than in primary school), and by then my mutation patterns in speech had been ‘established’.⁸

Other respondents described how the mutation patterns they learnt at home from their parents often differed from the canonical patterns taught in school. One might expect this to result in identifiable regional variation too. As discussed, we did not find evidence for this, which might mean that region is in fact irrelevant to variation in mutation (except to the extent that it intersects with numbers of first-language speakers). However, it might also be that our regional variable (north vs. south) was simply too coarse-grained, or that our sample size was too small to identify existing patterns.

A related point concerns quality of input. Multiple respondents suggested that they learnt how to mutate through being corrected by their parents, and that their children are learning to mutate from them. One respondent commented:

I think the environment is a factor and that she’ll learn partly from how I mutate, but I certainly don’t always mutate correctly!

The variability in production implied here is very much consistent with our data: respondents who patterned as “Variable” (who tended to give high ratings both to sentences with mutation and to sentences without mutation) were one of the largest categories, and the vast majority of our respondents exhibited some variation in their responses. Nonetheless, it is worth noting that, overall, there were as many respondents in the Conservative cluster (who tended to rate canonical mutation higher than non-canonical mutation) as in the Variable cluster.

As noted above, register may play an important role in explaining some of our results. We have discussed this in the context of certain triggers in particular, but it may also be that some of our respondents associate mutation in general with more formal registers, at least to an extent. We did not manipulate register, so our survey data cannot provide much beyond hints of this. However, there is evidence from our questionnaire, in which we asked respondents if they use mutation differently in different social contexts. Many described a distinction between “lazy”/ “messy”/ “untidy”/ “relaxed” Welsh and “formal”/ “correct”/ “standard” Welsh, and suggested that in the former contexts, not paying attention to mutation is common. Crucially, most respondents described not paying attention to mutation rather than not mutating at all. It was also clear that respondents perceived a distinction between canonical, prescriptive mutation rules and acceptability in colloquial registers. Many, for instance, described themselves as using mutation “incorrectly” in informal contexts, while acknowledging (as one respondent put it) that “People use words they know are officially incorrect but are more natural to use.”

Other respondents offered support for the idea that non-canonical mutation is acceptable in informal contexts, for example:

[…] nor do they worry if they don’t speak correctly.

When I speak with the family, there’s less pressure to get it right.

A methodological point should be raised in this context. The sentences used in our survey contained many natural gender nouns. Since the list of natural-gender nouns is limited, and not all nouns begin with a sound that is subject to mutation, it was necessary for the sake of variety to include a number of nouns that are well known, but might not be especially common in everyday speech. These include nouns referring to such individuals as princesses, kings, witches, monks, and maids. It is possible that such nouns, bringing with them associations of fairytales, influenced respondents’ interpretation of register. In future work it would be important to control for such effects, and interesting to manipulate register, such as with a matched-guise task (cf. e.g., Campbell-Kibler, 2008).

Related to register is accommodation, whereby speakers adjust their language in response to their interlocutor’s linguistic behaviour or even in response to expected linguistic behaviour (Wade, 2020). In the questionnaire, multiple respondents mentioned the influence of speakers’ first- and second-language status on mutation patterns, and one respondent commented that how they mutate “depends on who is in the conversation. I adapt to go along with the person. So I would say something less correct/formal to help a learner to understand, for example.” In this context it is important to note that we asked respondents to imagine that they heard each sentence in a conversation; responses might have differed had we framed the task differently.

One respondent commented explicitly on the topic of over-mutation, in response to a question on whether they think mutation is used differently by different groups of people:

I have second-language friends in their 30–40s who over-mutate quite often, that is mutate where there’s no need, or even remutate a word that’s already been mutated.

This comment recalls a finding of Thomas and Gathercole (2005), who found adults double-mutating certain nouns, and confirms that there is metalinguistic awareness among some Welsh speakers of over-application of mutation as a phenomenon. Our overall pattern of results, however, suggested that respondents were generally more tolerant of under-mutation that over-mutation.

In general, we consider the present work to constitute an exciting starting point for new research on the status of mutation in Welsh, while providing a perceptual and judgement-based perspective to complement earlier work focused on production. The study shows that the variation in the Welsh mutation system that has been detected in production studies of Welsh children, can also be seen in the expectations of adult Welsh-English bilinguals. Balanced bilingual and L2 Welsh speakers gave lenient acceptability judgements to non-canonical mutation, especially (though not exclusively) under-mutation. L1 Welsh speakers gave more conservative judgements. These judgements pattern with expected directions of change for minority languages in bilingual contexts (e.g. Dorian, 1994).

The method—an online survey in which respondents simply gave judgements to recorded sentences—allowed acceptability-judgement data to be collected relatively easily without requiring in-person interaction with a consultant or requiring respondents to record themselves. The completely transparent approach (in which respondents were made fully aware of the purpose of the research) was also successful—we see no evidence of serious demand characteristics. In spite of the study’s simplicity, it produced suggestive results that we hope will form the basis of further investigation, ideally with larger samples. As discussed above, future work might focus in a more controlled way on the role of variables such as register and trigger transparency. This methodology could also be easily adapted to involve different voices (e.g. to examine dialect differences), to manipulate respondents’ expectations of the speakers, and to investigate the role of different kinds of triggers, or triggers of different frequencies. We hope that our study and its results might be of interest not only to researchers interested in Welsh in particular, but also to researchers with an interest in grammatical change in bilingual situations of prolonged contact between a minority and prestige language more broadly.

Abbreviations

1 = 1st person, 3 = 3rd person, am = aspirate mutation, f = feminine, inf = infinitive, m = masculine, pst = past, pl = plural, prs = present, prog = progressive, sg = singular, sm = soft mutation, sbj = subject

Data availability and supplementary materials

The stimuli, raw survey results, demographic data, and a summary of results for Q2 are available via the Open Science Foundation, along with a key describing the variables in each of these datasets. DOI: 10.17605/OSF.IO/HQRZ6

Supplementary file 1: Appendix. Overall results for Q2

Supplementary file 2: Demographic data

Supplementary file 3: Results file

Supplementary file 4: Stimuli list

Supplementary file 5: Key for reading supplementary files 2–4

Competing interests

The authors have no competing interests to declare.