2.1 Corpus evidence for variant repetitiveness

Following Szmrecsanyi (2006), we use the term persistence to describe a speaker’s tendency to repeat the same variant in their own naturalistic production. At the syntactic level, persistence has been documented for a wide range of phenomena. In English, for example, it has been found for word-order alternations, including the passive (Weiner & Labov, 1983), ellipsis (Nykiel, 2015), and genitives, particle placement, and complementation (Szmrecsanyi, 2006). It is also a well-established factor influencing Spanish subject pronoun omission (Cameron, 1992; Cameron & Flores-Ferrán, 2004; Travis, 2007; Travis & Torres Cacoullos, 2012) and noun phrase agreement (Poplack, 1980, 1984; Scherre & Naro, 1991, 1992), and has more recently been shown to influence Mandarin locative variation (Li et al., 2023). In the domain of phonological variation, persistence has been shown to influence American English coronal stop deletion (e.g., old ∼ ol’), DH-stopping (e.g., them ∼ dem), and -ing ∼ -in’ variation (Tamminga, 2014, 2016b), as well as TH-fronting (e.g., think ∼ fink) in Scottish English (Clark & Walsh, 2015) and /t/-flapping in New Zealand English (Clark, 2018).

Corpus data also shows that speakers tend to repeat variants they have recently heard from other speakers, sometimes called convergence. There are sizable literatures on convergence in sociolinguistics (where it is also called accommodation) and phonetics (where it is also called imitation). As with persistence, convergence has been documented across languages and across levels of the grammar (Pellegrino et al., 2024). Syntactic convergence in conversational data has also been viewed in the syntactic priming literature as relevant naturalistic evidence for syntactic priming, as the experimental data shows considerable common ground between production-to-production priming and comprehension-to-production priming in syntactic alternations (Mahowald et al., 2016; Tooley & Traxler, 2010). There is also some evidence for this kind of parallel in corpus data: Li et al. (2023) found both persistence and convergence in Mandarin locative variation, meaning speakers tended to repeat a recently-used variant whether the prior instance came from the same speaker or a different speaker. Moreover, this paper showed that the empirical properties of persistence and convergence for the same variable in the same conversational data show broad qualitative similarities, including similar decay profiles where the effects persevere across multiple intervening sentences.

2.2 Connecting corpus repetitiveness to experimental evidence

A longstanding question in the study of corpus repetitiveness, both within and across speakers, is whether it should be attributed to automatic processing mechanisms such as priming, or to socially-motivated behavior such as style-shifting and accommodation (Pickering & Garrod, 2017; Szmrecsanyi, 2006; Tamminga, 2014). For syntactic variation, many researchers have taken syntactic priming to be the source of conversational repetitiveness, with evidence for their shared empirical properties being an important line of evidence for this link. One such property is temporal durability. It has been consistently observed that corpus persistence decays over time, but only gradually (e.g., Gries, 2005; Szmrecsanyi, 2005). For example, in the case of English dative alternation, Gries (2005) found in a corpus analysis that priming did not decline significantly over the course of an intervening sentence. Experimental syntactic priming studies have similarly found long-lived effects that endure beyond adjacent sentences, but do decay gradually over time (Bock & Griffin, 2000; Gries, 2005; Pickering & Garrod, 2004). Given that decay is an important signature of structural priming, the fact that conversational syntactic repetitiveness also undergoes decay with increasing distance between prime and target further strengthens the attribution of corpus repetitiveness in syntactic variation to syntactic priming (Bock & Griffin, 2000; Clark, 2018; Pickering & Garrod, 2004; Tamminga, forthcoming).

However, the argument that variant repetitiveness is driven by priming has also been extended to cases of phonological variation, despite the fact that there does not yet exist corresponding experimental evidence for phonological variant priming. It has been argued that repetitiveness in phonological variation reflects priming in conversational speech on the grounds that phonological variables also exhibit decay similar to that observed in syntactic priming (Clark, 2018; Tamminga et al., 2016). However, we might expect the repetition of syntactic constructions to behave differently from the repetition of phonological elements, operate over different time scales, and implicate different kinds of representations and sets of options (Levelt, 2001).

One particular reason we might not expect phonological persistence to mirror syntactic priming is that experimental work on basic phonological priming suggests quite different temporal properties. For example, facilitation effects in primed lexical decisions between phonologically-related words such as dough → snow tend to be short-lived (Dufour, 2008; Wilder et al., 2019). For instance, Wilder et al. (2019) examined the effect of phonological similarity between the prime and target (e.g., gray → grape) in lexical processing and found that phonological priming appears to be restricted to immediately sequential trials as opposed to persisting across a number of intervening trials. With respect to phonological variation specifically, recent work found that when variable ING is manipulated in a primed lexical decision experiment, -in’ → -in’ prime–target pairs enjoy a larger priming boost than the other pair types (White, 2021; White et al., 2024), but this extra boost does not last across an intervening filler trial. There has also been some work on repetition priming and semantic priming that involves different pronunciation variants (Sumner, 2013; Van de Ven et al., 2011). However, because almost all of these results come from experiments on how variation influences word recognition speed (as opposed to the choice between variants being the dependent variable), they are not as comparable to corpus phonological repetitiveness as the syntactic priming paradigm is to corpus syntactic repetitiveness.

When it comes to phonological variant repetitiveness in corpus data, though, we see a complex picture. Some research points to remarkably durable persistence effects. For the ING variable in particular, Tamminga (2016a) finds that in conversational speech, although ING persistence does decay significantly over time, this decay proceeds so slowly that it can last for over two minutes. In the same larger study, Tamminga (2014) also finds significant but slow decay for /ð/-stopping and /ð/-deletion, both lasting around 30 seconds. The decay picture for /t,d/ deletion is less clear, with no evidence for decay in monomorphemes and seemingly faster decay in past tense tokens; however, /t,d/ primes and targets are on average much closer together, making it difficult to assess the effect’s durability. Other studies report much shorter-lived corpus phonological persistence effects. For example, persistence for /t/-flapping (Clark, 2018) and the short front vowel shift (Villarreal et al., 2021) in New Zealand English is lost rapidly after about 3 seconds. Crucially, the decay effect was explicitly explored in these studies because it has been treated as a marker of priming; these studies argue that rapid decay is a hallmark of activation-based priming mechanisms. This contrast between the temporal properties of experimental phonological priming and existence of long-lasting corpus phonological persistence for at least some variables further motivates the investigation of decay in phonological variant priming. After all, even if phonological variant priming does exist, it may not share the same decay profile as the basic phonological priming effects. If phonological variant priming for ING turns out to be short-lived, it would cast doubt on variant priming as the primary cause of the long-lived persistence effect in corpus data; conversely, if phonological variant priming for ING is longer-lasting, it would strengthen the connection between persistence and priming.¹

Against this background, our goal is thus to test whether phonological variants can be primed in a controlled experiment, making use of an “inference under uncertainty” paradigm (Allen et al., 2003; Kleinschmidt & Jaeger, 2015; Levon et al., 2022; Newman et al., 2001; Peterson & Barney, 1952; Xie et al., 2023). Inference under uncertainty refers to the idea that as speech signals unfold over time, listeners constantly draw linguistic inferences, update expectations and generalize knowledge across words, sentences and talkers, given their pre-existing knowledge of probabilistic distributions of acoustic, linguistic and social cues (Kleinschmidt et al., 2018; Norris et al., 2003). For instance, in a typical phoneme categorization task, listeners would hear an ambiguous sound halfway along a continuum, and their choice of perceiving the ambiguous sound as either A or B would be modulated by which cue distributions they were exposed to and how they upweighted or downweighted certain cues (Kraljic & Samuel, 2005; Norris et al., 2003). In the current case, by asking participants to categorize ambiguous instances of the ING variable, we create uncertainty and the conditions in which a preference for a recently-primed variant may be revealed. Our experiment is thus closely related to more familiar tasks, such as phoneme categorization. However, a typical phoneme categorization task leads listeners to access distinct words with different meanings or perhaps even to perceive ungrammatical forms. In contrast, in our case, the ING variable, the competitors (-ing and -in’) are equally well-formed options that do not differ in their meaning, although, of course, they differ stylistically. We discuss this variable next.

2.3 The sociolinguistic variable ING

The sociolinguistic variable ING refers to the alternation between word-final velar nasal (/ŋ/) and alveolar nasal (/n/) after unstressed /i/, as in words like thinking ∼ thinkin’. We chose ING as our test case for this study in part because it is one of the best-studied sociolinguistic variables in English, meaning that we can rely on existing knowledge about how its production and perception are conditioned by different social and linguistic factors (e.g., Campbell-Kibler, 2007, 2011; Fischer, 1958; Hazen, 2008; Houston, 1985; Labov, 1972). ING also offers a number of desirable practical properties for use in the lab: It is easy for a model talker to produce the variants, it is phonetically straightforward to cross-splice the variants across verb stems, and it is possible to represent the variants orthographically in a familiar way. Most importantly, there is also prior corpus work documenting ING persistence in conversational speech (Abramowicz, 2007; Tamminga, 2014, 2016a, 2016b), giving us a point of comparison in line with our broad motivation to connect corpus and experimental work on phonological priming. This prior work shows that ING persistence is a very large effect that decays significantly but slowly over time: The effect is still detectable after two minutes of intervening speech (Tamminga, 2016a).

3.1 Methods

The experiment combines two established methodological paradigms: the lexical decision task and the two-alternative forced-choice task (2AFC). Lexical decision trials, where the participant gives a word/non-word response by pressing a key, were used to expose participants to critical prime items containing clear -ing or clear -in’. The lexical decision response confirms that the participant recognized the form as a word, even if it contained a nonstandard variant. Critical ING targets, differing from their primes in the verb stem, are manipulated to be ambiguous, and participants are asked to respond with a 2AFC identification of which word pronunciation they heard (with the options presented orthographically, as in [WORKING WORKIN’]). The prediction is that the variant used in the prime should influence participants to favor that variant in their identification of the ambiguous target.

While every trial involves some kind of response, critical sequences occur as triplets (that participants are not made aware of): two lexical decision trials followed by the target 2AFC trial. The second of those trials (immediately preceding the target trial) is the ING prime, while the first trial in the sequence contains a different ING word with the opposite variant to the prime. This avoids a possible confound, when during -ing-primed trials the participant has also heard more -ing overall, and vice versa. We refer to the trial preceding the prime as the “pre-prime.” Note that the inclusion of the pre-prime may bias participant behavior against our hypothesis, as the two variants could cancel each other out if their effects both extend to the prime.

A within-subjects design with two critical conditions is adopted. In one condition, participants encounter sequences of -ing → -in’ → ING (-in’-primed condition, with ING indicating the ambiguous suffix), whereas in another condition, participants have sequences of -in’ → -ing → ING (-ing-primed condition). This is further illustrated using concrete examples in Table 1.

3.2. Analysis and results

Responses during lexical decision trials (including both critical and filler trials) were coded for accuracy: whether a word was correctly identified as a real word or whether a nonword was correctly identified as a nonword. Responses to critical categorization trial targets were coded for each ING variant chosen by participants. A total of 31 participants were removed because their overall accuracy in the lexical decision task was lower than 80%. Data from the remaining 71 participants was analyzed (number of participants in each list: List 1 = 22, List 2 = 14, List 3 = 15, List 4 = 20). Critical sequences, where participants judged -in’-containing prime and pre-primes to be nonwords, were also excluded from the final analysis, since failure to recognize the item as a word in these cases might interfere with the priming effect. In the end, the final statistical analysis included 2,197 observations (75% of the original data). The overall mean accuracy for pre-primes (88%) and primes (89%) was nearly identical. For primes, the mean accuracy for -in’-suffixed items was 85%, and the mean accuracy for -ing-suffixed ones was 96%. In terms of pre-primes, the mean accuracy for -in’-suffixed and -ing-suffixed items was 82.9% and 94.3%, respectively. The high accuracy results indicate that participants successfully recognized the prime and pre-prime as words. This is crucial, as recognition is not always guaranteed for isolated words containing nonstandard variants.

Statistical analysis was conducted using R version 4.0.5 (R Core Team, 2021). Mixed effects logistic regression was run using the lme4 package version 1.1-27.1 (Bates et al., 2015), and plots were created using the ggplot2 package version 3.3.5 (Wickham, 2010) and the sjPlot package version 2.8.9 (Lüdecke et al., 2019). Data and analysis scripts are available at https://osf.io/j7c3d/.

We used the glmer() function from lme4 to fit a mixed-effects logistic regression model predicting listeners’ ING responses (-in’ = 0, -ing = 1) when identifying ambiguous targets. We included as fixed effects a critical predictor that tests our hypothesis, plus a set of control predictors to capture other factors that may influence variant identification:

• Critical predictors

  • – Prime condition: Categorical variable comparing the -in’-primed vs. -ing-primed conditions, with the contrast coded as -ing-primed = 0.5, -in’-primed = –0.5

• Control predictors

  • – Target word-specific -ing bias: Continuous predictor indicating the probability of -ing given the verb stem in the ambiguous target; z-scored.

  • – Target frequency: Continuous predictor indicating the whole-word lexical frequency in the ambiguous target, z-scored.²

  • – Trial number: Continuous predictor referring to the sequence of trials throughout the experiment, z-scored.

We selected a random effects structure following the recommendation from Sonderegger (2023) to prioritize random slopes associated with critical predictors. Accordingly, we included random intercepts by Speaker and Target, as well as correlated random slopes of Prime condition by Speaker and Prime condition by Target. We do not perform further model selection, but rather report the full results of this design-driven model.

The model output is given in Table 2. There is a significant main effect of Prime condition (β = 0.77, p< 0.001), visualized in Figure 1 along with one reference value: the -ing input rate in the experiment (the black dashed line). Participants were significantly more likely to categorize an ambiguous target as containing the variant -ing when they had just heard an -ing variant on the previous trial than when the previous trial contained -in’. No other control predictors were statistically significant.

Figure 1: Predicted probability of -ing response in categorization of ambiguous ING target, with all other predictors held at their average values. The black dashed line represents the overall -ing rate participants that were exposed to during the experiment (50%).

3.3. Interim discussion

In Experiment 1, we found that exposure to an immediately preceding variant in perception can influence the subsequent categorization of an ambiguous variant in a different word. This suggests that the identification of phonological variants can be primed, supporting the basic proposal that conversational repetitiveness in phonological variation could be caused by priming. The difference between the two conditions cannot be attributed to perceptual convergence towards the talker’s overall ING rate, because the conditions do not differ in that rate. Rather, participants were sensitive to their immediately prior perception, taking the input into account in a highly local way.

Apart from this predicted effect of prime condition, notice that in both conditions, listeners were more likely to perceive ambiguous targets as containing -in’ than they were in the perception norming studies (i.e., compared to the average -ing response baseline of 70% in the norming study). Note that this shift was not tested in the model, since the norming study is not a condition of the experiment itself. The norming bias itself may reflect participants’ stylistic awareness that words presented in isolation in a laboratory context most often contain standard pronunciations. We speculate that the apparent shift away from that bias might reflect that participants were also adapting to the overall distribution of the two variants in their perception of the model talker. It is possible that when inferring which of two variants they heard, in the -ing-primed condition, listeners were matching their -ing responses to the model talker’s overall -ing rates (the black dashed line in Figure 1), perhaps due to a global expectation during the course of the experiment.

It is also worth noting that the effect of the prime is significant even though participants heard the opposite variant as a pre-prime only one trial earlier. If there is any hold-over facilitation from the pre-prime, which we cannot rule out in this design, that effect would exert an influence in the opposite direction of the prime, attenuating the observed priming effect. Despite this possibility, we still observed the priming effect from the immediately preceding trial. To better understand whether this kind of priming effect plausibly extends across multiple trials, we now turn to Experiment 2 to ask whether variant priming decays over time. Recall that ING persistence has been found to be long-lasting in prior corpus research (Tamminga, 2016a). The fact that the pre-prime’s influence does not cancel out the prime’s influence is one hint that the experimental effect we are looking at may decay more quickly. Questions about the temporal decay properties of different kinds of priming have played a key role in connecting corpus and experimental work in other domains and are also linked to questions about the precise mechanisms underlying experimental priming effects. So, in Experiment 2 we tested whether the priming effect observed in Experiment 1 persists across an intervening trial.

4.1 Methods

The experimental design of Experiment 2 exactly mirrored that of Experiment 1. Two different types of response tasks were employed in a within-subjects set-up: lexical decision task for pre-primes and primes, and forced-choice categorization task for ambiguous targets. The only difference between these the two experiments was in the trial make-up. Unlike in Experiment 1 where each trial always consisted of a sequence of Pre-prime → Prime → Target, participants in Experiment 2 sometimes encountered sequences of Pre-prime → Prime → Intervener → Target. We compared listeners’ likelihood of identifying an ambiguous target as the -ing variant between the -in’-primed condition and the -ing-primed condition, both with and without a single intervening filler trial.

4.2. Analysis and results

Prior to statistical analysis, the data collected were first pre-processed following the same practice adopted in Experiment 1. Of the original 158 participants, data from 32 participants were removed from the final statistical analysis due to their poor performance on lexical decision trials (overall accuracy score of less than 80%). The final statistical analysis was therefore based on data from the remaining 125 participants. In Experiment 2, we took an additional step by excluding responses that involved interveners but were treated as ‘nonword’. In the end, the final statistical analysis was conducted based on 4313 observations (81% of the original data). The overall mean accuracy for pre-primes and primes was both 88%. For primes, the mean accuracy was 83% for -in’-suffixed items and 96% for -ing-suffixed ones. In terms of pre-primes, the mean accuracy for -in’-suffixed and -ing-suffixed ones was 79% and 97% respectively.

Statistical analysis was conducted following the same principles as detailed in Experiment 1. A generalized mixed-effects regression model was configured to predict participants’ -ing/-in’ responses when identifying ambiguous targets (0 for -in’, 1 for -ing). The fixed effects in this case are:

• Critical predictors

  • – Prime Condition: Categorical predictor referring to the within-subjects conditions that differ in which prime immediately precedes to the ambiguous target. Levels: -in’-primed vs. -ing-primed, sum-coded (-ing-primed = 0.5, -in’-primed = –0.5)

  • – IntervenerYN: Categorical predictor standing for the within-subjects conditions that differ in whether there is an intervener word between the prime and target. Levels: 1 intervener vs. 0 interveners, sum-coded (0 interveners = 0.5, 1 intervener = –0.5)

  • – The two-way interaction of Prime Condition * IntervenerYN

• Control predictors: same as Experiment 1

Following the same approach as in Experiment 1, we started a random effects structure with correlated random slopes. To resolve the singularity issue, the correlation has to be dropped. prime condition by Participant and IntervenerYN by Participant were thus included as uncorrelated (rather than correlated) random slopes and Target was included as a random intercept.

Model output, as shown below in Table 3, revealed a significant main effect of Prime Condition (β = 0.31, p< 0.001): Listeners were significantly more likely to identify the ambiguous target as -ing-containing after immediately hearing a clear -ing-suffixed word. The effect of IntervenerYN was statistically significant (β = 0.18, p = 0.03): Overall, the perceived -ing response rates increased significantly when there were no intervening items between the prime and ambiguous target. The interaction between prime condition and intervenerYN was also significant (β = 0.49, p< 0.01), suggesting that in the -ing-primed condition, participants were significantly more likely to perceive the ambiguous target as -ing-containing, compared to the global mean, when there was no intervener between the prime and ambiguous target. No other control predictors turned out to be statistically significant. The predicted effects of prime condition and its interaction with IntervenerYN are further plotted in Figure 2.

Figure 2: Predicted probability of perceived -ing responses in categorization, when all other predictors held at their average values, in probability terms.

The model as specified does not test whether there is priming within each intervener condition in isolation. Thus, post-hoc pairwise comparisons were made using the emmeans package version 1.5.5-1 (Lenth et al., 2019), with p-values corrected for multiple comparisons using the Tukey method. As shown in Table 4, these comparisons found that the difference between -ing-primed condition and -in’-primed condition was significant in the no-intervener condition (β = 0.55, p< 0.001), but not significant with an intervener (β = 0.07, p = 0.58).

4.3. Experiment 2 discussion

The goal of Experiment 2 was to investigate whether the variant priming effect from Experiment 1 is robust to the presence of an intervening trial between the prime and target. In the no-intervener condition, we replicated the finding of variant priming from Experiment 1. However, with an intervening trial consisting of a lexical decision to a single monosyllabic word, the phonological variant priming was no longer detectable. Phonological variant priming appears to be short-lived.

While the main effect of variant priming was replicated in Experiment 2, notably, the trend of global convergence towards the model talker’s rates was absent. Given that the instructions were identical across both experiments–which should have resulted in a comparable salience of the -in’ form–this is an unexpected result. We propose that future research could systematically vary the model talker’s -ing rates to investigate this discrepancy further.

This short-lived phonological variant priming effect differs notably from the temporal properties of phonological variant persistence in conversational speech in the same variable. In corpus data, ING persistence has been found to last for over two minutes (Tamminga, 2016a). The fact that phonological variant priming of ING in a controlled experimental setting does not survive even a single intervening word thus suggests that variant priming may not be the only cause of the long-lived persistence effect of ING in conversational speech. However, this conclusion needs to be treated with some caution, as we will discuss. The rapid decay observed here may be exacerbated by the presence of the pre-prime (the opposite variant) in the experimental design–a factor we cannot isolate in this study. While the prime was evidently strong enough to overcome any hold-over interference from the pre-prime, the rapid decay after one intervener may reflect the complex interplay between the pre-prime and the prime.

5.1. Is corpus repetitiveness really a result of priming?

The mixed empirical results mentioned above naturally lead us to question the extent to which we should view them as strong evidence for or against the idea that conversational variant repetitiveness has its source in priming. If one follows the line of thinking from Clark (2018)—that parallels between the empirical properties of priming and corpus repetitiveness can be taken to suggest a shared underlying mechanism—one might be tempted to conclude that any divergence in the empirical properties of these phenomena should conversely be interpreted as suggesting that they do not have a common mechanism. We do think the results of Experiment 2 should raise doubts as to whether conversational repetitiveness in phonological variation can be straightforwardly attributed to priming. However, this is not the same as concluding that we now have good evidence against this kind of explanation. There are any number of reasons why we might see divergences between the behavior of phonological variation in our experimental task and conversational data, even if they did indeed arise from a common mechanism. We consider several of these caveats here, and leave open the possibility that there may be others.

The first type of caveat involves perception-production asymmetries. Our experiments target perception-to-perception priming, whereas in corpus data we are observing either repetition from perception to production (between-speaker convergence), or repetition from production to production (within-speaker persistence). Furthermore, the majority of the literature based on corpus data consistently finds stronger repetitiveness when speakers prime their own choices in production (Li & Tamminga, 2021; Li et al., 2023). The specific ING repetitiveness decay profile we have appealed to here comes from prior work specifically on ING persistence (Tamminga, 2016a), where all observations involved within-speaker variant repetition in production (with no intervening speech from the interlocutor, see Tamminga, 2014, for methodological details). The variant priming observed here nevertheless was based on the listeners’ responses to ambiguous targets produced by a model talker, more akin to cross-speaker repetitiveness. On top of this, priming effects may not always behave uniformly across comprehension and production tasks, with prior work documenting notable discrepancies between the two (see Pickering and Ferreira, 2008, for a detailed discussion). Therefore, in the case of priming evidence derived from an individual’s perceptual behavior, the results should not be translated directly into a straightforward understanding of production repetitiveness.

Second, even though a controlled context may help tease apart possible mechanisms that give rise to repetitiveness, the paradigm adopted here with words in isolation differs in many ways from natural conversations that deal with connected speech with rich contextual information. For instance, connected speech often licenses the variation that is stylistically unexpected in the citation form of words. The listener responses elicited under the context of words in isolation thus may differ from those elicited under connected speech.

A third caveat concerns a range of other potential factors in the rich interactional context of conversational speech that could make a difference. It is possible that in corpus repetitiveness, while priming may be at play, its effect may be too subtle to be detected due to the co-influence of other factors such as interspeaker convergence and style-shifting. This being said, it remains unclear why these factors would make priming effects last longer, if they indeed play a role.

With these possibilities kept in mind, it is unclear how to interpret the fact that the rapid decay of experimental ING variant priming observed in our Experiment 2 is strikingly divergent from the long-lasting ING repetitiveness observed in prior corpus work. One conclusion is that we should not jump too quickly to any interpretation of corpus phonological repetitiveness as priming, since there are reasons to doubt this connection. The difference in decay properties for the same variable suggests that priming alone is not a straightforwardly sufficient explanation for corpus persistence. It may well play a role, but we also need to understand how it interacts with other factors to generate the observed patterns of conversational repetitiveness. This uncertainty highlights that major gaps remain in our understanding of the relationship between priming and conversational repetitiveness and, more broadly, between basic processing mechanisms and the production of sociolinguistic variation in real-world contexts, inviting a more detailed consideration of these factors in future research.

5.2. Is phonological variant priming merely phonological overlap priming?

The variant priming effect we found in Experiment 1 raises a new question: Might this priming effect be the same as phonological overlap priming in contexts not involving variation? In other words, does the effect we observed necessarily involve phonological variants, or does it reflect a more general process of phonological facilitation? The question is salient in light of the rapid decay of ING variant priming observed in Experiment 2, because there are independent reasons to believe that phonological facilitation is short-lived. In spoken word recognition, listeners have been found to become faster at responding to a target word after they have encountered a preceding prime that shares some phonological overlap with the target, especially if the overlap is in the syllable rhyme (e.g., dough → snow). Like the variant priming effect we observed, phonological priming in word recognition seems to be short-lived, significant only at immediate distances (Dufour et al., 2023; Radeau et al., 1995; Wilder et al., 2019).

Although a role for phonological overlap in facilitating lexical access may not be the same as an increased likelihood of identifying one variant or the other in a categorization task, the possible links are worth considering. Insofar as the literature on variation in word recognition evidence is relevant, it provides us with some preliminary evidence against the idea that variant priming can be fully reduced to phonological overlap priming (White et al., 2024). In particular, White (2021) finds that recognition of targets ending in the -in’ variant, like thinkin’, is faster when preceded by a different word containing the -in’ variant (such as jumpin’), but not when preceded by a word ending in the non-variable phonological string /In/, like dolphin. On the corpus side, Tamminga (2014) finds that there is no detectable influence of a recent instance of /ŋ/ on the use of the -ing variant in conversational data. So, while we might think of phonological variant priming as a type of priming within phonology that may therefore have some of the same general qualities as phonological overlap priming, including very limited durability, there is reason to suspect that there is some independent facilitation process here, distinct from priming of phonological strings. This possibility also speaks to the theoretically-interesting question of whether phonological variants have the same representational and/or grammatical status as obligatory phonological phenomena. Future work is thus needed to examine to what extent phonological variant priming resembles or differs from phonological priming.

This study shows that phonological variant identification in perception is subject to psycholinguistic priming. All else being equal, hearing a clear -ing makes listeners more likely to choose -ing again, given an ambiguous target for categorization. We also show that this phonological variant priming effect decays rapidly over time, after only one monosyllabic word, suggesting that phonological variant priming may not be the underlying mechanism that drives long-lasting phonological variant repetitiveness in conversational speech, at least in the case of ING variation. We thus argue that when it comes to detecting and interpreting processing phenomena in conversational speech, the results could be strengthened by the development of more precise experimental analogues, as opposed to cross-referencing assumptions from different domains.

Demographic information of participants in Experiment 1

The demographic information about the participants in Experiment 1 including their age, gender and race/ethnicity is further summarized below (Number of participants in each experimental list: List 1 N = 28, List 2 N = 22, List 3 N = 25, List 4 N = 27):

• Age distribution: N. of 17–20y = 64; N. of 21–25y = 38;

• Race/ethnicity distribution: N. of Asian = 28; N. of Black/African = 11; N. of Caucasian/White = 48; N. of Hispanic/Latinx = 6; N. of Mixed race = 4; N. of Native American = 0; N. of Other = 5; N. of Pacific Islander = 0;

• Gender distribution: N. of woman = 59; N. of Man = 41; N. of Other = 2;

Demographic information of participants in Experiment 2

The demographic information of participants in Experiment 2 is further summarized below (Number of participants in each experimental list: List 1 = 14, List 2 = 16, List 3 = 18, List 4 = 19, List 5 = 14 , List 6 = 14, List 7 = 15 , List 8 = 15):

• Distribution of gender: N. of Man = 76, N. of Woman = 80, N. of Other = 2;

• Distribution of race/ethnicity: N. of Asian = 2, N. of Black/African = 22, N. of Caucasian/White = 119, N. of Hispanic/Latinx = 5, N. of Mixed Race = 6, N. of Native American = 1; N. of Other = 0; N. of Pacific Islander = 0;

• Distribution of age: N. of 17–20y = 3, N. of 21–25y = 41, N. of 31+y = 113;