1 Introduction

Acceptability judgments are the standard methodology for investigating morphosyntax (e.g. Schütze 1996). In recent years, acceptability judgment methodologies have been combined with best practice from sociolinguistics in order to investigate the morphosyntax of non-standard varieties (e.g. Cornips & Poletto 2005; Barbiers & Bennis 2007; Zanuttini et al. 2018; Smith et al. 2019) and have been shown to be reliable ways to capture variation within communities (Thoms 2014; Zanuttini et al. 2018).

While these methodologies have been successful at investigating cases of stable variation, work on syntactic microvariation using judgments does not tend to incorporate discussion of ongoing change in these varieties. From sociolinguistic research, we know that the final stages of a change can be rapid and less orderly than the earlier stages of a change, with unexpected production patterns (e.g. Dorian 1977; Campbell & Muntzel 1989; Hill 1989; Schilling-Estes & Wolfram 1999; Labov 2001). Furthermore, Dorian (1977) presents evidence that “semi-speakers” of East Sutherland Gaelic, those at the tail end of the loss of the variety, are “unreliable” when asked to perform elicitation tasks. However, Adger (2017) argues that the “unreliability” of elicitations in East Sutherland Gaelic is due to the breakdown and subsequent reorganisation of the syntax in the ongoing change process, with semi-speakers using sparse data to reconstruct a system that fits the underlying “syntactic ecology” of the language.

We can therefore ask: how do speakers in the late stages of a syntactic change behave when completing judgment tasks? Is there unpredictability? Do we see an organised restructuring process? Understanding how speakers behave in these contexts can shed light on the nature of the grammatical system at the end of a change, and how this is affected by the potential social pressures of a metalinguistic task.

In this paper, I focus on one community where a number of forms in the dialect are said to be undergoing obsolescence: the Shetland dialect of Scots. Linguistic change and the loss of features in this community is argued to be happening rapidly – even from one generation to the next (Smith & Durham 2011, 2012; Durham 2013). Shetland dialect is thus an ideal testbed for seeing how a community of speakers who are undergoing potential language loss, and exhibiting the unpredictable patterns seen at the end of a change, behave when presented with acceptability judgment tasks which ask them to make explicit judgments on dialect variants – in this case, –n, a particle used in certain types of biased questions. I will also consider whether unpredictability extends to judgments on stable acceptable and unacceptable features.

2 Dialect syntax and acceptability judgments

Since at least Chomsky (1966), the standard data collection procedure for morphosyntactic data has been acceptability judgments. While these began as introspective binary judgments, considerable criticism of this methodology (Quirk & Greenbaum 1970; Labov 1972; Newmeyer 1983; Rizzi 1990; Nagata 1992; Bard, Robertson & Sorace 1996) has led to the development of alternative, gradient methodologies for syntactic acceptability, such as Likert scales or magnitude estimation tasks (e.g. Bard, Robertson & Sorace 1996; Schütze 1996; Sprouse 2007; Bader & Häussler 2010). It is these gradient methodologies that have been incorporated with best practice from sociolinguistics (e.g. Labov 1966; 1972) in more recent work that has aimed to tap into vernacular variation in morphosyntactic phenomena (e.g. Cornips & Poletto 2005; Barbiers & Bennis 2007; Zanuttini et al. 2018; Smith et al. 2019). In English varieties alone, acceptability judgment data has led to insights into a wide variety of syntactic phenomena, such as imperatives (e.g. Henry 1995; Weir 2013), preposition dropping (e.g. Myler 2013; Biggs 2015), datives (e.g. Horn 2013; Wood & Zanuttini 2018), possessives (Tsoulas & Woods 2019) and more.

Gradient judgment tasks have been shown to be reliable indicators of acceptability in cases where there is no expected variation. Here, inter-speaker judgment reliability is high (Sprouse, Schütze & Almeida 2013; Sprouse & Almeida 2017; Langsford et al. 2018, though see Linzen & Oseki 2018 on Hebrew and Japanese judgments). Gradient judgment tasks have also been shown to be reliable in cases of stable variation in a community, where speakers using both a dialect variant and a standard variant (Thoms 2014; Zanuttini et al. 2018). However, when there are multiple grammars of a non-standard dialect within an individual community, Henry (1995) observes that speakers with one grammar may form incorrect hypotheses about speakers of the other. She notes (1995: 56) that it is “particularly important to get grammaticality judgments from speakers who themselves use the dialect concerned” in order to avoid this.

How does this work in situations of linguistic change? There has been little discussion of the reliability of acceptability judgment methodologies in situations of linguistic change.¹ On one hand, languages are always changing (e.g. Labov 1994, 2001). After an initial innovation, changes progress constantly and systematically (Weinrich, Labov & Herzog 1968). We would therefore expect that as long as we are able to target speakers of a particular grammar within the change, judgments should be reliable for that grammar. However, at the end stages of linguistic change, speakers’ behaviour with regard to the feature being lost is argued to be somewhat unpredictable. This can be seen with individual linguistic features in an otherwise “healthy” linguistic variety (e.g. Ingason 2010), but is amplified when linguistic change is happening to multiple features at once, leading to potential language loss (e.g. Dorian 1977; Christian, Wolfram & Dube 1988; Campbell & Muntzel 1989; Hill 1989; Labov 1994, 2001; Schilling-Estes & Wolfram 1999).

This unpredictability is reflected in the behaviour of speakers at the end of a change in other types of linguistic data gathering methodologies. In recorded production data, we see “dramatic increases in variability due to incongruent and idiosyncratic change” (Cook 1989: 235) across phonological (e.g. Wolfram & Schilling-Estes 1995; Schilling-Estes & Wolfram 1999) and morphosyntactic (e.g. Dorian 1973) variables.

Furthermore, we see unpredictability in the results of more structured data gathering tasks with speakers undergoing change in Dorian’s (1977) use of spoken elicitation tasks with “semi-speakers” of East Sutherland Gaelic. “Semi-speakers” were younger members of the East Sutherland community whose Gaelic was considered “imperfect”, though they could still make themselves understood (Dorian 1977; 1994). Dorian presented 16 members of the East Sutherland community with 115 English sentences and asked them to translate the sentences into Gaelic. The semi-speakers produced analogical levelling across a range of paradigms, including verbal agreement and irregular noun formation, in an inconsistent manner. She concludes that reduced use of features applies “essentially to the individual speaker, not the community” (Dorian 1977: 30), with different speakers producing different patterns seemingly randomly. However, Adger (2017) argues that the levelling in East Sutherland Gaelic can be understood in terms of an abstract restructuring of the syntactic system, reconstructed by the semi-speakers on the basis of sparse input. Rather than acting at random, speakers are creating a new system within the confines of underlying constraints.

Is this flux that we see for speakers at the end of a change in both production and elicitation data reflected in acceptability judgment data, and if so, how? Do we see idiosyncratic behaviour, or evidence of a restructured system? In the next section I introduce the community that we will use to test potentially obsolescing speakers’ behaviour with regards to the feature –n when presented with acceptability judgment tasks.

3 Shetland dialect

The Shetland Islands are the most northerly region of Scotland, located over 200 miles north of Aberdeen (see Figure 1). The islands have a relatively stable population of just over 23,000 (Shetland Islands Council 2017).

Figure 1

Map of Scotland highlighting the Shetland Islands (CC-BY-SA-3.0).

The dialect of Scots spoken in the Shetland Islands² in the present day is known for being especially distinctive among Scots varieties, retaining older Scots features presumably due in part to its remoteness (Melchers 2004) as well as features which are argued to have resulted from the period of language contact between Scots and Norn (Millar 2008), the west Norse language spoken in the isles until the mid-eighteenth century (e.g. Barnes 1998).

However, a number of recent works have noted that Shetland dialect appears to be obsolescing (van Leyden 2004; Melchers & Sundkvist 2010; Smith & Durham 2011; 2012; Sundkvist 2011; Durham 2013; 2014; 2017; Jamieson 2015). In particular, Smith & Durham’s (2011; 2012) research demonstrates that young speakers of the dialect have reduced rates of both Shetland-specific and Scotland-wide lexical, phonological and syntactic variants as compared to both older and middle-aged speakers. Their research also shows some seemingly unpredictable behaviour among the younger speakers – for example, with demonstrative distal yon. Yon participates in a distal deictic system in Shetland dialect alongside this and that, and can either be used in a determiner-type role (1) or as a pronominal (2) (Smith & Durham 2011: 209).

Smith & Durham investigated speakers’ usage rates of demonstratives in both determiner and pronominal positions, in singular distal contexts – contexts where there is variation between yon and that.³ There were generally low rates of yon across the three age groups that Smith & Durham studied but, surprisingly, there was a higher percentage of yon as compared to that in the younger group than the middle or older groups. Looking at this between speakers, however, only four of the eight younger speakers used any yon at all, with four speakers using that 100% of the time. Furthermore, two of the eight were especially prolific yon users, with around 36% yon usage – compared to < 20% yon across the older and middle groups – thus bumping up the overall younger speakers’ mean usage rates. The division in younger speakers’ usage rates did not correlate with any expected social factors, such as gender, social network, or attitude towards the islands, leading Smith & Durham to conclude that the variation seen is the kind of “personal pattern variation” found by Dorian (1994) in East Sutherland Gaelic, and that the dialect “may be facing rapid dialect attrition” (Smith & Durham 2011: 220).

Shetland dialect is therefore well placed as a community in which to investigate how speakers behave in acceptability judgment tasks when presented with a potentially obsolescing feature. In the next section, I introduce the acceptability judgment tasks that speakers completed in this research.

5 Quantitative results

5.1 Analysing Likert scale data

As described in Section 4.2 above, participants were asked to judge each item in the research on a 5-point Likert scale. 1 was described as something that was “unusual; sounds weird; no one says that”, while 5 was described as being “totally natural; I say that; people around me say that”. No particular descriptors were attached to the middle of the scale.

Although this is a numerical scale, we cannot assume that the distances between the points of the scale are equivalent – e.g. is the difference in acceptability between and 4 and 5 of the same magnitude as the difference in acceptability between 3 and 4? The question thus arises as to whether scales like these can be interpreted as continuous, or if they must be analysed as discrete, ordinal points. The decision about how to categorise the data affects the choice of statistical test used.

If the data is interpreted as continuous, a t-test can be used to establish if there are significant differences in the mean ratings between two groups. If, however, the data is interpreted as ordinal, a non-parametric test e.g. the Mann-Whitney-Wilcoxon rank sum test would be more appropriate. As opposed to comparing means, the Mann-Whitney-Wilcoxon test ranks the judgment scores given by participants and calculates whether there is a different distribution of the data given the median of the two populations.

A test for normality can be conducted in order to establish whether the data resembles a normal distribution, and thus can be interpreted as continuous. Here, I conducted Shapiro-Wilk tests for normality. For each of the data cases presented below, the test returned a significant result, indicating that the distribution of scores did not align with a normal distribution. This is an initial indication that non-parametric tests would be better suited to the data. However, there is considerable research suggesting that even when continuous data are non-normal, t-tests are most appropriate when the sample size is large enough (e.g. Zimmerman & Zumbo 1993; Stonehouse & Forrester 1998; Skovlund & Fenstad 2001).

However, the research cited did not specifically include the interpretation of Likert scales. de Winter and Dodou (2010) conducted a comparison of levels of Type I (false positive) and Type II (false negative) errors between t-test results and Mann-Whitney-Wilcoxon results carried out on 5-point Likert scale data with different distributions. The authors found that the statistical power of both parametric and non-parametric tests was equivalent across most distributions, with two crucial exceptions. When the data was heavily skewed (e.g. almost all ratings were 5), non-parametric tests held more statistical power; on the other hand, when the distribution was strongly bimodal (e.g. ratings were fairly equally split between 1 and 5), parametric tests held more power.

Looking at the descriptive statistics for positive polar questions (e.g. Figure 2), we can see that the data is heavily skewed, with almost all cases being rated 5. Throughout the rest of the data, results are either skewed in this way, or have somewhat more flat distributions, where both t-tests and Mann-Whitney-Wilcoxon tests had equal power in de Winter and Dodou’s (2010) work.

Figure 2

Acceptability judgment results for positive polar questions, by age group.

As non-parametric tests were shown to hold more power for the heavily skewed data, and other data were shown to have non-normal distributions through the Shapiro-Wilk test, I opt for non-parametric tests here. This will take the form of both Mann-Whitney-Wilcoxon rank sum tests and ordinal logistic regression models.

I will first present the results for a variety of examples expected to have different relationships to change. Firstly, I will present the results for a set of stable acceptable examples – positive polar questions, which in Shetland dialect are taken to be the same as in English,⁵ e.g. (16).

(16)	Do you have a dog?

I will then present results for two different sets of presumed unacceptable examples. Firstly, a set of unacceptable examples that don’t contain any ostensibly dialectal features: wh-questions with incorrect word orders e.g. (17).

(17)	What do want you?

Secondly, a set of unacceptable examples that include dialect features, but that are ungrammatical in the dialect – specifically, negative imperatives with do no,⁶ e.g. (18). Unlike English negative imperatives where both don’t and do not can be used (Potsdam 1996; Rupp 2003), negative imperatives in Scots varieties like Shetland dialect do not permit both options. Only dunna/dinnae (depending on location) is available; do no is unacceptable (Weir 2013: 9).

(18)	Dinnae smoke in here!

(19)	*Do no smoke in here!

In these three contexts, I present the results from Mann-Whitney-Wilcoxon rank sum tests in order to test whether there were differences by age group.

I will then present results for the changing –n feature. Firstly, with tag questions like (9), above, where we see acceptability. Secondly, for tag questions like (10), above, where –n has never been attested in the dialect. Finally, for biased questions like (11), above, which appear to be undergoing change. The examples with –n will be compared with judgments for the standard Scots constructions that could be expected to be used in all three contexts (20–22).

(20)	You can drive, can you no?

(21)	She hasna gone and gotten a dog, has she?

(22)	Did you no like that pizza place?

To compare the –n variables with the relevant standard constructions, I use ordinal logistic regression models – a non-parametric variant of linear regression designed to deal with multiple, discrete response categories such as the Likert scale data presented here.

Regression models allow us to test to what extent a range of predictor variables (e.g. type of construction, age group, gender) affect a result (e.g. the judgment score given), and predict to what extent a change in a variable should affect the outcome for the predictor. For example, to what extent does being in the older age group affect the scores you give overall?

Here, I present results from cumulative link mixed models, run using the ordinal package (Christensen 2018) for R. Cumulative link mixed models are a form of ordinal logistic regression, which include mixed effects to allow potentially random effects (e.g. differences between individual participants, or responses to individual experimental stimuli) to be accounted for.

For each context, I ran models with random intercepts for participant and example, with construction type, age group and evidential context as predictor variables.⁷ At no point was gender found to be a significant factor for any of the contexts, and so this was removed from the models.

5.3 Ungrammatical: Wh-questions with incorrect word order

The second example we will look at is an example which is presumed unacceptable and does not include any notable “dialect” features: specifically, wh-questions with incorrect word orders, as in (24). Each participant judged 8 examples of questions like these, giving 80 judgments per age group.

(24)	You’re going to the shop. I ask if you can get me something and you say:
	What do want you?

As can be seen in Figure 3, speakers in both age groups rate these constructions low: primarily giving ratings of 1. The median score for both age groups is 1; there is a slight difference in mean ratings (younger = 1.275, older = 1.375) but no statistically significant difference (W = 782, n.s.). Ungrammatical examples are given “floor” level ratings, as Bader & Häussler (2010) found for ungrammatical word orders in German. It therefore seems that judgments continue to be reliable even at the end stages of change. However, the picture is more complicated for ungrammatical examples that include dialect features.

Figure 3

Acceptability judgments for ungrammatical wh-questions, by age group.

5.4 Ungrammatical: Negative imperatives

Each participant judged four examples of negative imperatives which were constructed with do no in place of dunna, such as example (25), giving 40 judgments per age group. Recall that while no is the Scots form of not, Scots imperatives cannot be formed with do no, unlike English do not. Only dunna (equivalent of don’t) is available.

(25)	I have just arrived at your house and I’m really hungry. There’s some fresh scones sitting out, and I’m eyeing them up, but you say:
	Do no touch them! They’re for later.

Speakers in both age groups did not rate these examples in the “grammatical” range, as set out above. In Figure 4, we see that there are very few 5 ratings in either age group. However, despite the fact that speakers did not consider these constructions acceptable, there were clear differences in the ratings between the age groups in both communities. In the 18–30 group, we see a higher number of 3 and 4 ratings, as compared to the higher numbers of 1 and 2 ratings seen in the 55+ group.

Figure 4

Acceptability judgment results for do no in negative imperatives, by age group.

While neither of these groups rate the examples at ceiling, there are statistically significant differences with respect to how low the constructions are rated. Again, there were differences in mean ratings (younger = 2.75, older = 2.0) and in the median. The median for younger speakers was 3, while for older speakers it was 1. This difference was significant (W = 536, p = .008). For younger speakers, there seems to be an unwillingness to always fully rule out an ungrammatical example: instead, sometimes a more mid-range rating is given.

This holds across participants quite generally: of the 10 1 ratings given here in the younger age group, 7 of them come from 2 participants, with the rest of the participants operating primarily across the mid-range. On the other hand, in the older age group, 6 of the 10 participants gave 1 ratings in at least 3 of the 4 test examples.

So, when it comes to dialect features, younger speakers seem to be doing something different. We can then turn to look at the changing variable in question, –n, and see how participants behave.

5.5 –n in tag questions with positive anchors

In these cases, participants judged examples like (9), repeated here as (26), and general Scots examples like (27). There were 4 examples that were clearly local (40 judgments per age group), 12 examples that could be confounded with the standard (120 judgments per age group) and 16 general Scots examples (160 judgments per age group). These Ns are reflected in the width of the bars in Figure 5.

Figure 5

Acceptability judgment results for tag questions on positive anchors, by age group.

(26)	You can drive, can’n you?

(27)	You can drive, can you no?

As can be seen in Figure 5, in cases where the –n particle is potentially confounded with a standard variant, participants in both age groups give a high percentage of 5 ratings; the same is also true in the standard contexts. For speakers in the 55+ group, the clearly local contexts are also given a high percentage of 5 ratings; however, the younger group are more evenly split between ratings of 4 and ratings of 5.

I ran a cumulative link mixed model on the data, presented in Figure 6. The estimated coefficients in the left column of the main table indicate the inferred effect of the predictor variable on the participants’ judgments on the Likert scale in log odds, with the z value (second rightmost column) indicating whether or not this is a statistically significant predictor. A z value over 2 generally indicates a significant predictor.

Figure 6

Cumulative link mixed model results for tag questions on positive anchors.

Figure 6 shows that there is a general effect of whether or not the example is a clearly local one (particleN.LOCAL), as well as an interaction between age group and the judgments given to clearly local instances of –n (particleN.LOCAL:AGEY). There is also a general effect of –n in confounded contexts (particleN.OTHER).⁸ For older speakers, it appears that clearly local variants are as acceptable as the potentially confounded variants, while that is not quite the case for younger speakers. However, it is worth noting that both groups still rated –n highly, even in clearly local contexts.

5.6 –n in tag questions with negative anchors

–n on tag questions with negative anchors, such as (28), was compared with standard constructions like (29). There were 16 examples of each type of construction (so 160 judgments per age group for each construction).

(28)	She hasna gone and gotten a dog, has’n she?

(29)	She hasna gone and gotten a dog, has she?

The picture is complicated in a similar way as it was for the clearly ungrammatical examples. From Figure 7, we can see that standard examples are rated highly by speakers in both age groups. For –n, on the other hand, younger speakers give more 3 and 4 ratings as compared to the older speakers’ 1 and 2 ratings. Neither group is rating these examples in the “acceptable” range, but their behaviour regarding something being “unacceptable” seems to vary. This is backed up by the results of the cumulative link mixed model (Figure 8).⁹ While there is an overall effect of the construction type (particleN), there is also an interaction between age and construction type (particleN:AGEY), indicating that younger speakers rated examples with –n in this unacceptable context significantly higher than older speakers did.

Figure 7

Acceptability judgment results for tag questions on negative anchors, by age group.

Figure 8

Cumulative link mixed models for tag questions on negative anchors.

5.7 –n in matrix biased questions

With examples of –n in matrix biased questions, like (30), the picture is complicated. Speakers also judged standard examples like (31). There were 4 examples that were clearly local (40 judgments per age group), 12 examples that could be confounded with the standard (120 judgments per age group), and 16 general Scots examples (160 judgments per age group). These Ns are reflected in the width of the bars in Figure 9.

Figure 9

Acceptability judgment results for biased questions, by age group.

(30)	Did’n you like that pizza place?

(31)	Did you no like that pizza place?

The results can be seen in Figure 9. As expected, standard examples are rated highly. There are also very high ratings for –n in potentially confounded contexts. From the results of the cumulative link mixed model (Figure 10), we can see that there are overall effects of the different types of –n as compared to the standard construction (particleN.LOCAL and particleN.OTHER); there is also an interaction between the clearly local cases of –n and the age group.

Figure 10

Cumulative link mixed model results for biased questions.

Looking at the descriptive statistics, we see that for older speakers, the median for the N.LOCAL cases is 5, with over 50% of examples rated 5. The mean is mid-range (3.73), seemingly from a collection of 1–2 ratings (27.5% of the total) bringing down the average. On the other hand, the 18–30 group’s median is 4, with over 50% of examples rated 4. The mean is still, however, mid-range (3.63), with only 12.5% of judgments rated 1 or 2.

If we were simply to take the mean judgments of the older and younger speakers, then, we could quite easily conclude that the participants were behaving in the same way. However, the way in which the two groups reached this result was quite different, as shown by the model, and further illuminated by the descriptive statistics.

6 Discussion

A first possible analysis for the data worth entertaining is that in the face of ongoing linguistic change, young speakers in Shetland are simply unsure about giving judgments, hedging their bets in case they get something “wrong”. This would account for, for example, the data on tags on negative anchors presented in Section 5.6: younger speakers are unsure of what other dialect speakers would do. However, this would be surprising as younger speakers have been shown to have greater perceptual awareness and metalinguistic understanding than older speakers in metalinguistic tasks (e.g. Drager 2011; Carrera-Sabaté 2014; Lawrence 2017). Furthermore, in the data, younger speakers do appear to be certain about particular judgments, such as the tags on positive anchors shown in Section 5.5, and treat ungrammatical examples with dialect features, as shown in Section 5.4, differently to ungrammatical examples with no dialect features (as shown in Section 5.3). The younger speakers do not seem to be simply uncertain about their judgments.

A second possible analysis is that we are dealing with what Labov (1972) terms hyperdialectalism. Hyperdialectalism is a form of hyperadaption in which speakers become “hyperlocal” (Labov 1972), increasing their use of particular features of the traditional dialect – especially in cases where there is ongoing language attrition. Speakers may extend a feature to contexts where it did not originally appear – structural hyperdialectalism (e.g. Trudgill 1986; Britain 2009). For example, loss of rhoticity along the English/Welsh border led to speakers in traditional rhotic areas developing rhotic forms in words with no etymological <r>, such as last [la^ɹst] (Britain 2009: 135). There is also the potential for statistical hyperdialectalism, where speakers do not extend use across contexts but simply increase the usage rates of a particularly salient feature (Wolfram & Schilling-Estes 2006). Smith & Durham (2011) propose that statistical hyperdialectalism is what we see with the younger speakers in Shetland who used higher percentages of yon as compared to older speakers.

We could conclude from the quantitative results that –n exhibits hyperdialectal usage patterns in Shetland dialect for younger speakers in the present day – both structural, via extension to tags on negative anchors (Section 5.6) and statistical, via usage in matrix biased questions at a rate higher than expected (Section 5.7). However, qualitative data indicates that it is not the case that younger speakers believe they produce –n in contexts that older speakers do not. Discussion with participants makes it clear that younger participants are giving mid-range ratings, as we saw for –n on negative anchors and in biased questions, based on the perception that other people in the community might say something, even though they themselves wouldn’t. This is shown in example (32), in which a younger speaker rates an example of –n in a rhetorical wh-question, another context in which –n is not attested in the dialect, nor acceptable for older speakers.

Here, SY07 explicitly states that she would not use the example given, and gives the alternative that she would use, with standard Scots –na negation: canna. However, she still gives the example a rating of 4, the second highest rating. SY07 justifies this in a number of ways, drawing a clear distinction between herself and speakers who are “really broad”. By stating that (older) members of her own family would use a variant, SY07 implicitly defends her position as an in-group member. Furthermore, in mentioning a specific area (North Roe, a rural area at the northmost point of mainland Shetland) where a particular feature could be used, SY07 is marking out her perceived understanding of the dialect. However, examples like the one in (32) received a mean rating of 2.23 from the older speakers, with a mode of 1. The younger speaker’s understanding of the variation in the community, therefore, does not seem to be an accurate perception and would extend the potential for –n to a context where it is not accepted for users of a more traditional form of the dialect.

This type of justification given by the participant also appeared to extend to other ungrammatical examples, such as the instance in (33). Participant SY05 was one of the youngest participants at age 18, but was also from Whalsay, a distinctive dialect area which is reputedly particularly “broad” (Cohen 1978; Melchers 1985; Durham 2017).

Despite the fact that the example shown in (33) is entirely ungrammatical in terms of its word order, both in standard English and in Shetland dialect, SY05 states that this is something she has “heard”. However, she clearly goes on to distance herself from it – “I wouldn’t say it myself”. The participant here may have been commenting on the use of the be-perfect in this example,¹⁰ a traditional feature of Shetland dialect that SY05 consistently gave ratings of 5 to in assumed grammatical filler examples (e.g. Are you been to Majorca?). It may be that the participant has an awareness of the be-perfect and is suggesting that other speakers use it in different ways (e.g. with this unusual word order).

Younger participants also made similar comments about imperatives with do no, which was discussed in Section 5.4. For example, participant SY02 said that the example presented in (24) was something that “older” speakers would use, and that while she herself would use dunna, she nonetheless gave it a rating of 4.

On the other hand, older participants did not attempt to attribute ungrammatical examples to either younger speakers or to a previous generation of dialect speakers.¹¹ They were also generally more disparaging in their comments about ungrammatical examples, describing –n in contexts like (32) as “doesn’t make any sense” and “daft”, and tags on negative anchors as in (28) as “bad” or “contradictory”.

The results of the younger speakers’ acceptability judgments therefore do not seem to reflect hyperdialectal usage patterns, despite the fact that these are claimed to be part of the variety. However, I interpret these results as an instance of hyperdialectalism within the realm of perceptual dialectology. While the younger speakers use the –n particle in certain contexts (e.g. tags on positive anchors), they are aware that, traditionally, –n can be used in more contexts than this. However, they aren’t accurately able to pinpoint what those contexts are. They therefore extend the possibility of a variant to new contexts (e.g. tags on negative anchors) and lose the defined original environments for use (e.g. matrix biased questions) – just as we see in hyperdialectalism in usage patterns. This then extends to slightly higher ratings for plausibly dialectal ungrammatical examples (e.g. imperatives with do no), through a general willingness to accept that there may be dialect features that they as younger speakers do not use.

This phenomenon does not appear to be limited to the –n particle. Earlier judgment work on Shetland dialect also shows younger speakers exhibiting these patterns with respect to a different set of variables. In work investigating the loss of verb raising in imperatives and questions in Shetland, Jamieson (2015) finds that speakers in their younger (18–30) age bracket give mid-range ratings to examples of verb raised questions, as in (34). Participants gave the same kinds of comments about who exactly would use these examples – for example, participant 01A cites an older family member as someone who would use the verb raised example.

Speakers in Jamieson’s older age bracket (60+) also did not rate these questions at ceiling level. However, these speakers’ comments indicated they would use these variants sometimes, rather than attributing use of them to other community members (35–36).

This is similar to what we see with –n in matrix biased question contexts in this research: though there was no statistical difference in younger and older speakers’ judgments (see Section 5.6), the information conveyed by the numerical judgment was different, as reflected by the qualitative comments.

Taking the qualitative comments into account, younger speakers of the obsolescing Shetland dialect often give judgments that do not seem to reflect their own grammars. Beyond what they know they definitely say, they rely on their hypotheses about older or more rural speakers’ grammars when giving judgments. This is in sharp contrast to older speakers of the same variety, whose judgments are more definitive and whose comments appeared to reflect their own usage, rather than hypothesising about other speakers’ dialects. I term the phenomenon seen here in the Shetland results perceptual hyperdialectalism – exhibiting the same sorts of structural and statistical effects as in hyperdialectal usage, but through the perspective of a metalinguistic task.

The behaviour of the younger speakers in this research is reminiscent of Henry’s (1995) speakers of Dialect A, a subset grammar of Dialect B, who formed “incorrect hypotheses” about what Dialect B speakers could do. The results here support Henry’s point that in dialect syntax research, in order to access a grammar it is crucial to get judgments from speakers for whom that is their grammar. However, the results here from Shetland dialect also show younger speakers forming incorrect hypotheses about features that are not undergoing change (i.e. do no), indicating that there is something broader at play.

This final point also suggests that it is not that younger speakers are reorganising the system following the underlying syntactic ecology of the language, as Adger (2017) argues for East Sutherland Gaelic. Younger speakers do not claim to produce –n in new contexts; rather, they seem to be attempting to establish their knowledge of traditional dialect more generally, claiming awareness of other speakers’ dialect use.

I therefore propose that the effects of perceptual hyperdialectalism seen in the Shetland dialect results here is due to the sort of linguistic insecurity defined by Preston (2013). Traditionally, linguistic insecurity (Labov 1966) has been defined as occurring when speakers feel “that the variety they use is somehow inferior, ugly or bad” (Meyerhoff 2006: 292). However, Preston (2013) argues that it can be seen in terms of an individual speaker’s relationship to their regional standard, “when one feels that they are not able to perform the linguistic job at hand” (Preston 2013: 324). More specifically, speakers may believe that their own variety is “correct”, but feel they are lacking in their ability to use it. I propose this is what is happening with the younger Shetland dialect speakers in this research.¹²

Wolfram & Schilling-Estes (1995: 698) note that “nonmainstream dialects… play a large role in the shaping of cultural identity” and that obsolescence of a dialect threatens this cultural identity (1995: 699). Children of Shetland heritage have a generally positive attitude towards the dialect and consider it to be “an important facet of their identity” (Durham 2014: 303). Furthermore, there is a strong desire for them to distance themselves from “soothmoothers” (Karam 2017), people who are not members of the community in-group. Despite these positive attitudes and the clear importance of the dialect as an tool for identity construction, however, Durham (2014, 2017) shows that younger Shetland speakers’ self-reported usage estimates lean more strongly towards English than dialect when compared to the children in Melchers’ (1985) 1983 corpus, even in informal, local contexts such as speaking to a friend who is also from Shetland. Stadler (2016) and Stadler et al. (2016) show that both younger and older speakers in Shetland were able to identify the direction of a morphosyntactic change and place themselves in relation to the change curve, with younger speakers consistently positioning themselves ahead of the curve to a greater extent than older speakers. Taken together, these findings indicate that younger speakers are aware that they are using fewer traditional dialect features, while still maintaining that dialect use is a core part of their identity as Shetlanders.

How does this relate to acceptability judgment tasks? When given acceptability judgment tasks, speakers are asked explicitly to act as “gatekeepers” for the variety (Jørgensen 2010) and are given an opportunity to make conscious choices about their linguistic resources in order to signal in-group membership (Coupland 2007). Following Le Page & Tabouret-Keller (1985: 181):

“The individual creates for themselves the patterns of their linguistic behaviour so as to resemble those of the group or groups with which from time to time they wish to be identified, or so as to be unlike those from whom they wish to be distinguished.”

Aware that their grammars are different to those of older Shetland dialect speakers, and insecure about how this fact relates to their identity as Shetlanders, younger participants turn to what they “know”, or think they know, about the linguistic features of a more traditional grammar – distinguishing their own usage (majority 5 ratings, such as –n in tags on positive anchors) with what they perceive to be possible, or overheard (majority 3–4 ratings, such as –n in matrix biased questions or in tags on negative anchors). Participants back this up by making explicit their connections to the community, legitimizing their right to be making judgments, as seen in examples like (32–34). Though in the task participants are not employing the resources in conversation, they are doing as much as they can to preserve their in-group status through “knowledge” of a more traditional form of the dialect. However, we can compare with older speakers’ behaviour and see that younger speakers’ “knowledge” of older speakers’ grammars is not necessarily accurate. Their desire to signal their in-group membership leads to judgments that do not reflect any current grammar of the dialect, and which must be carefully interpreted.

7 Conclusions

As highlighted at the beginning of this paper, acceptability judgment methodologies have been shown to be reliable sources for understanding stable variation within communities. However, the reliability of this methodology for dealing with linguistic change, particularly the unpredictable sort of variation attested at the end of a change, had not been explored. I thus investigated the reliability of acceptability judgment methodologies within a variety that is undergoing obsolescence: the Shetland dialect of Scots.

Older and younger speakers of the dialect patterned together when it came to filler examples that did not include any dialect features: grammatical examples were rated at “ceiling”, while ungrammatical examples were rated at “floor”. However, when dialect features were introduced, older and younger speakers diverged in their patterns of judgment. Examples that were truly acceptable were rated at ceiling for both age groups. For unacceptable examples, younger speakers gave mid-range ratings as compared to the older speakers’ low ratings; for examples undergoing change, younger speakers generally rated examples mid-range, while older speakers were more polarised in their responses.

It didn’t seem that participants were restructuring a system as Adger’s (2017) work found in East Sutherland Gaelic, as this rating pattern held across four different examples (ungrammatical do no imperatives, –n in tag questions on negative anchors and matrix biased questions, and the verb raising in questions found in Jamieson 2015). Instead, these results appeared to show unpredictable behaviour on behalf of the younger participants, perhaps retaining variables at surprising rates or extending contexts for use.

Incorporating speakers’ qualitative comments into the discussion, however, I argued that these younger speakers were exhibiting perceptual hyperdialectalism, following the principles of both structural and statistical hyperdialectalism in the metalinguistic domain. I argued that this is a reflex of Preston’s (2013) definition of linguistic insecurity. These speakers, who hold dialect as an important part of their cultural identity but whose grammar is different from that of the older speakers, attempt to distinguish themselves from community outsiders through their “knowledge” of that traditional dialect.

This research strongly supports Henry’s (1995) point that getting judgments from speakers of the grammar you are interested in is crucial for dialect syntax research, and that failing to do this can make the judgments “unreliable”. However, the “unreliability” of the judgments made by speakers undergoing significant linguistic change can also be enlightening, if carefully interpreted alongside qualitative (metalinguistic) data.

Abbreviations

N = biased question marker in Shetland, NEG = negation

Competing Interests

The author has no competing interests to declare.