1 Introduction
Sound patterns tend to be phonetically ‘natural’: they reflect constraints on speech production and perception. For instance, many phonological rules, for example consonant assimilation or vowel harmony, increase ease of articulation; they mirror low-level gradient phonetic effects which are due to automatic processes such as coarticulation and gestural overlap. It has long been observed that, cross-linguistically, phonetically natural rules are much more prevalent than unnatural ones (Hooper Reference Hooper1976). One explanation for such typological asymmetries concerns the existence of an individual learning bias. In various phonological theories, typological asymmetries are reflected by phonetically motivated biases in the speaker's mind (Donegan & Stampe Reference Donegan, Stampe and Dinnsen1979, Archangeli & Pulleyblank Reference Archangeli and Pulleyblank1994, Hayes & Steriade Reference Hayes, Steriade, Hayes, Kirchner and Steriade2004). These grammatical biases could induce learning biases, such that phonetically natural patterns are easier to learn, and hence have an advantage in transmission across generations (Schane et al. Reference Schane, Tranel and Lane1974, Wilson Reference Wilson2006). Note that while some theories posit transmission as the sole locus of the evolution of natural patterns over unnatural ones (Ohala Reference Ohala and Jones1993, Blevins Reference Blevins2004), the learning-bias hypothesis is not strictly speaking incompatible with them (Moreton Reference Moreton2008, Beguš Reference Beguš2018). It has repeatedly been put to the test, exploring the idea of a learning bias for many different typological regularities. Some evidence has been found in favour of a learning bias, though results from such studies make it difficult to disentangle phonetic naturalness from complexity.
Indeed, there is ample evidence for a learning bias favouring simpler patterns over more complex ones (for a review, see Moreton & Pater Reference Moreton and Pater2012a). For example, White (Reference White2014) shows that participants more easily learn simple phonological alternations than saltatory ones (which are rare, but attested). Saltatory alternations involve an alternation where an intermediate segment must be ‘jumped over’ (for example, a rule that says an underlying segment /p/ surfaces as [v] between vowels, jumping over [b]). Such alternations, though, affect multiple phonological features, making them featurally more complex than those affecting only one feature (e.g. intervocalic voicing of /p/ to [b] or spirantisation of /b/ to [v]).Footnote 1 In a similar vein, Skoruppa et al. (Reference Skoruppa, Lambrechts and Peperkamp2011) show quicker and better learning of single-feature alternations than of multi-feature alternations. Likewise, Peperkamp et al. (Reference Peperkamp, Skoruppa, Dupoux, Bamman, Magnitskaia and Zaller2006) find that a rule applying to a typologically attested natural class is learned better than one applying to an arbitrary group of sounds.
Nevertheless, some evidence of a learning bias for some phonological patterns over others that does not seem to involve complexity does exist. For example, Schane et al. (Reference Schane, Tranel and Lane1974) found that a typologically attested rule of word-final consonant epenthesis before a vowel (akin to the rule of French liaison) is learned faster than a typologically unattested one of word-final consonant deletion before a vowel. Wilson (Reference Wilson2006) demonstrates asymmetrical generalisation of a newly learned velar palatalisation rule, with participants generalising the rule from before mid vowels to before high vowels, but not vice versa. This follows the typological, phonetically grounded fact that languages that palatalise velar stops before mid vowels also do so before high vowels, whereas the inverse is not necessarily true. It should be noted, though, that this study contained other experimental conditions that did not yield learning asymmetries that are predicted by typological facts. Another example of evidence for bias towards phonetic naturalness comes from Myers & Padgett (Reference Myers and Padgett2014). They show not only that the phonetically natural rule of final devoicing is learned better than the phonetically unnatural rule of final voicing, but also that the former is better generalised to utterance-medial position than the latter, in accordance with typological facts. The participants in that study, however, were native English speakers, and English is known to be subject to phonetic devoicing (e.g. Docherty Reference Docherty1992), which might have influenced the participants’ performance. Beyond segmental patterns, there is evidence of a learning bias for rules targeting suprasegmental phenomena as well. A study with English- and French-speaking participants (Carpenter Reference Carpenter2010) considered the learning of a rule where low vowels attracted stress (phonetically natural) compared to a rule where high vowels attracted stress (phonetically unnatural). Overall, the natural rule was learned better than the unnatural one. More recently, Carpenter (Reference Carpenter2016) similarly taught English- and French-speaking participants either a stress-assignment rule whereby stress was attracted to the leftmost heavy syllable (natural) or a similar rule whereby stress was attracted to the leftmost light syllable (unnatural). The results showed better learning of the former than of the latter.Footnote 2
In this article, we focus on vowel harmony. This common phonological phenomenon involves co-occurrence restrictions on vowels, such that all of the vowels within a certain domain (typically the word) must share one or more phonological features. It is often manifested in the form of morphophonological alternations. Hungarian, for example, has a restriction on the backness of vowels within a word, such that most suffixes of the language have two allomorphs: one containing a back vowel, and one containing a front vowel. The data in (1) demonstrate this restriction: (1a) contains only back vowels and (1b) only front vowels, though both have the same dative suffix.
(1)
Some experimental work concerning learning bias for vowel harmony has compared it to alternations affecting arbitrary groups of sounds (Pycha et al. Reference Pycha, Nowak, Shin and Shosted2003, Baer-Henney et al. Reference Baer-Henney, Kügler and van de Vijver2014). These studies show that vowel harmony is learned better than the arbitrary alternations. Similar studies have found better learning of vowel harmony than a vowel–consonant dependency (e.g. where the roundness of a suffix vowel depends on the voicing of a stem consonant; Baer-Henney & van de Vijver Reference Baer-Henney and van de Vijver2012). The unnatural rules in all of those studies, though, were formally more complex than the vowel-harmony rule, in that they involved multiple features.
However, there are a number of typological tendencies concerning vowel harmony that do not confound phonetic naturalness with complexity, many of which have been the subject of previous research into learning bias. For instance, Finley & Badecker (Reference Finley and Badecker2008) consider the fact that vowel harmony is always directional in nature. This means that vowel features spread to other segments either to the left or to the right of a trigger. There are no rules in the typology based on a majority-count rule, where the feature that occurs the most times in a word spreads to vowels without that feature. Accordingly, when exposed to input that is compatible with both a directionality-based and a majority-based harmony rule, participants overwhelmingly infer the former. Similar results are reported in an additional study (Finley & Badecker Reference Finley and Badecker2009a). The same authors also considered the fact that vowel harmony causes agreement of subsegmental features (e.g. [back]), and showed that participants do indeed base generalisations on such features rather than on individual segments (Finley & Badecker Reference Finley and Badecker2009b), though insofar as segments are themselves composed of multiple features, it is unclear if this result truly disentangles complexity (one vs. many features) from naturalness (features vs. segments).
Some types of harmony show asymmetries in the typology. In rounding harmony systems, mid vowels may trigger harmony for all vowel types while high vowels do not, but the converse is not true. That is, if a language has a rounding harmony system with unrestricted high vowel triggers, it also tends to have unrestricted mid vowel triggers. This asymmetry is hypothesised to enhance the perceptual salience of harmony patterns and boost the perception of the non-high vowels, which tend to have weaker phonetic cues to rounding. (Thus, having all the vowels of the word agree in this feature means the listener is more likely to correctly identify the presence of rounding.) Finley (Reference Finley2012) shows better learning of rounding harmony patterns triggered by mid vowels than those triggered by high vowels, in line with phonetic naturalness and the typology. Taking this one step further, and in a similar vein to Wilson (Reference Wilson2006), Kimper (Reference Kimper, Hansson, Farris-Trimble, McMullin and Pulleyblank2016) looks at extrapolation in cases of rounding harmony. He taught listeners rounding harmony patterns, and shows that learners actively extrapolated the harmony pattern from high to mid vowels, but not from mid to high vowels, again in line with the typology and the phonetic grounding.
We focus here on a further typological tendency concerning vowel harmony: the logically equivalent rule of vowel disharmony, whereby suffixes must disagree along some phonological dimension with vowels in the root, is virtually unattested cross-linguistically. Vowel harmony has been proposed to be born out of vowel-to-vowel coarticulation (Ohala Reference Ohala1994),Footnote 3 which could explain why disharmony patterns would be less likely to emerge. But if speakers possess phonetic knowledge, could this not bias them towards the learning of a phonetically natural harmony pattern over an unnatural disharmony one? Then, in addition to unnatural rules like vowel disharmony being less likely to arise, they would be less likely to survive repeated transmission, as learners are biased against them (or rather, towards their natural counterparts). This would heavily disadvantage unnatural rules over time, and could explain in part why they are so rare in the typology.
Looking for a naturalness bias in the case of harmony vs. disharmony presents a clear design advantage in comparison to any of the studies examining such a bias reviewed above: the test items in the two conditions are exactly the same. Hence, any observed effect between conditions must be due to the experimental manipulation, without any possible confound from the properties of the items. It should be noted also that a learning bias favouring harmony over disharmony is difficult to explain simply in terms of complexity, since they both involve a single abstract feature. Depending on the formalisation, it could be argued that disharmony is more complex than harmony, because its description requires the use of a negative operator. Work from developmental psychology has shown that the concept ‘different’ may be harder to learn than the concept ‘same’, since the presence of negation makes ‘different’ more complex than ‘same’ (Hochmann et al. Reference Hochmann, Carey and Mehler2018). From a constraint-based perspective, though, it is the harmony pattern, not the disharmony pattern, that is formalised with the negative operator (e.g. harmony: *[αF][―αF]; disharmony: *[αF][αF]). We thus remain agnostic with respect to the impact of negation on the pattern, and consider vowel harmony and disharmony to be on a par, as they both involve a single feature. These single-feature patterns make a good test case for the naturalness learning bias hypothesis.
Two previous studies have failed to show evidence for a learning bias in favour of vowel harmony compared to vowel disharmony. In one of them (Pycha et al. Reference Pycha, Nowak, Shin and Shosted2003), American English listeners were exposed in an artificial language learning paradigm to singular/plural alternations that were either harmonic or disharmonic. The numeric pattern of results suggested an advantage for the harmony rule, but this difference was not statistically significant. As noted by the authors themselves, with a sample size of only ten participants per group, this study might have been underpowered. However, similar results were obtained in another study, with 30 participants per group: French listeners who were exposed to short stories in an artificial accent of their native language which followed a systematic rule of harmony or disharmony showed equivalent learning of both rules (Skoruppa & Peperkamp Reference Skoruppa and Peperkamp2011). It is worth noting that both studies also contained a condition in which the participants were exposed to harmony for some vowels and disharmony for others. Interestingly, and in line with the work cited above, in both studies performance on this more complex rule was significantly worse than on either the harmony or the disharmony rule. This suggests that if there is a bias favouring natural rules at all, it is not as strong as the one favouring structurally less complex rules (Moreton & Pater Reference Moreton and Pater2012b).
We re-examine the question of a learning bias favouring harmony over disharmony. Testing two rules that are equivalent with regards to their complexity, we address the issues of relative power highlighted in previous work. We additionally investigate a novel factor that might influence the learning of phonological rules, namely sleep. Sleep is known to enhance the learning process by way of memory consolidation (Walker & Stickgold Reference Walker and Stickgold2004). Newly acquired knowledge consolidates overnight, yielding improved performance on the following day. In the domain of language, sleep-dependent memory consolidation has been shown in adults for perceptual adaptation to synthetic speech (Fenn et al. Reference Fenn, Nusbaum and Margoliash2003), as well as the learning of non-native sounds (Earle & Myers Reference Earle and Myers2015), phonotactic constraints (Gaskell et al. Reference Gaskell, Warker, Lindsay, Frost, Guest, Snowdon and Stackhouse2014), words (Davis et al. Reference Davis, Di Betta, Macdonald and Gareth Gaskell2009, Dumay & Gaskell Reference Dumay and Gaskell2007, Havas et al. Reference Havas, Taylor, Vaquero, de Diego-Balaguer, Rodríguez-Fornells and Davis2018) and morphosyntactic rules (Batterink et al. Reference Batterink, Oudiette, Reber and Paller2014). We examine whether sleep enhances the learning of phonological rules, and if so, whether it does so differentially for phonetically natural vs. unnatural rules. In particular, if there is only a small learning bias favouring phonetically natural rules, but an additional effect of sleep, with more consolidation for these rules compared to their unnatural counterparts, this would add to the evidence of a role for learning biases in shaping the typology.
We use an artificial language learning experiment administered in two sessions (test and retest), separated by twelve hours, either with or without an intervening period of sleep. As in Pycha et al. (Reference Pycha, Nowak, Shin and Shosted2003), our test case is palatal harmony, a rule whereby vowels within the domain of the word must share the same value along the front–back dimension. This long-distance dependency between vowels is well attested in the typology, but its converse is not. We know of only one language that has been reported to have a possible case of productive palatal disharmony, Ainu (Krämer Reference Krämer1999).
Given the logistical difficulties of testing participants in the morning and the evening with a twelve-hour interval in the lab, we opted for an online experimental set-up, allowing participants to take part in the study from home. We recruited (self-reported) North American English-speaking participants on Mechanical Turk, and tested them on stimuli produced by a native speaker of Northern Metropolitan French from a previous, unpublished, study.Footnote 4 Mechanical Turk has been successfully exploited previously using artificial language learning to test learning biases for syntactic typological universals (Culbertson & Adger Reference Culbertson and Adger2014, Martin et al. Reference Martin, Abels, Adger, Culbertson, Goel, Seifert and Freksa2019). Additionally, Steele et al. (Reference Steele, Denby, Chan and Goldrick2015) use this platform to compare implicit phonotactic learning with native vs. non-native stimuli; they found that American English participants were able to learn implicit rules with either English or French stimuli similarly well. Our set-up has two main advantages. First, as Steele et al. point out, online testing allows for samples of large size to be recruited quickly, and requires relatively few resources on the part of the experimenter (although for the present case with a precisely timed second test session these advantages were less clear, as we will see in §2.1.3). Second, the use of non-native stimuli reduces the likelihood that participants will rely on a metalinguistic strategy to perform the task. As the sounds of one's native language have fixed mappings to orthographic symbols, participants in artificial language learning experiments might encode the stimuli in terms of orthographic rather than phonological categories. Non-native stimuli encourage more phonetic listening, as naive listeners are less likely to have fixed mappings between non-native stimuli and graphemes.
2 Experiment
2.1 Methods
2.1.1 Stimuli
96 CVCV items containing French phonemes were used, each composed of two different consonants and two different vowels. Half of the items contained two front vowels, drawn from the set /i e ε̃/, while the other half contained two back vowels, drawn from the set /u o ͻ̃/;Footnote 5 consonants were drawn from the set /b d g p t k v z ʒ f s ʃ n ʁ/. Each of the twelve possible vowel combinations (/i-e/, /e-i/, /i-ε̃/, /ε̃-i/, /e-ε̃/ and /ε̃-e/, and likewise for the back vowels) occurred equally often (i.e. in eight items). The 96 stimuli were divided into three sets, A, B and C, each containing 16 front stems and 16 back stems. These sets were used variously as exposure or test sets, as detailed below.
For each of these items, which were to be used as ‘singulars’, two ‘plurals’ were created, one containing a front vowel, and the other a back vowel: CVCV-/tɛl/ and CVCV-/tɔl/ respectively. Thus half of the plural forms were harmonic (i.e. CVCV with front vowels + /tɛl/, e.g. /pegitɛl/, or CVCV with back vowels + /tɔl/, e.g. /gͻ̃dutɔl/), and half of them were disharmonic (i.e. CVCV with front vowels + /tɔl/, e.g. /pegitɔl/, or CVCV with back vowels + /tɛl/, e.g. /gͻ̃dutɛl/).
All items (singular and plural) were recorded in a soundproof booth by a female native speaker of Northern Metropolitan French, using an M-Audio Micro Track II digital recorder and an M-Audio DMP3 pre-amplifier in 16-bit mono, at a sampling rate of 44.1 kHz.
2.1.2 Procedure
The experiment was run online from a server based at the Département d’Études Cognitives of the École Normale Supérieure. Participants were therefore not present in the lab, and all interfacing with them was done via e-mail. Upon logging into our website for the first time, they were forewarned that the experiment consisted of two parts, and that they would need to complete the second part around 12 to 13 hours after completing the first part; they were asked only to accept the invitation to participate in the experiment if they were sure they would be able to do so. Those who accepted were then asked to provide their e-mail address, so that they could be contacted when it was time to take part in the second session.
During the first session, participants received instructions regarding the exposure phase. They were told that they would hear words from an invented language, and that words would be presented in their singular and plural forms. They were also told that the language had two forms of the plural suffix: TEL and TOL.
Participants were first exposed to two repetitions of one of the sets of 32 unique stems, A, B and C (randomly assigned). During exposure, a stem was played, followed after 500 ms by its plural form (either harmonic or disharmonic, depending on the condition, which was also randomly assigned). Immediately following the auditory presentation of the stimuli, two boxes appeared on screen in random locations, each containing one of the plural suffixes (TEL or TOL). Participants were requested to click on the box corresponding to the plural form they had heard. If they provided an incorrect response, the trial was repeated, to ensure that they had correctly heard the singular and plural forms of all exposure stimuli. This task was added during the exposure phase to ensure that participants were paying attention to the stimuli, and had their speakers turned on. The random button position manipulation was chosen so that participants were required to actively seek out their response, and could not repeatedly press the same button. We recorded the number of errors committed by participants, and used this information to exclude those who were either not paying attention or did not have proper audio equipment, as described in §2.1.3 below.
The first test phase began immediately after this exposure phase. Test trials consisted of the presentation of a stem followed by both possible plural forms (i.e. CVCV-/tɛl/ and CVCV-/tɔl/), with an ISI of 500 ms. The order of presentation of the two plural forms was randomised across trials. After all three stimuli were played, two buttons appeared at random locations on the screen (exactly as during the exposure phase), labelled with the two plural suffixes. Participants were asked to select which of the two plural forms they thought was correct. They were first tested on the exposure set (i.e. the same items that they had just heard) and then on another randomly assigned set of 32 stimuli that they had not previously heard (for example, they might have been exposed to and tested on set B, and then tested on set A, with set C being reserved for the retest). They were not told anything about the two types of items they were tested on.
Approximately twelve hours after having completed the first session, participants received an e-mail inviting them to log back into the website (the server refreshed every ten minutes, so some participants received the e-mail after 12 hours and 9 minutes, for example). They had exactly one hour to do so. If participants attempted to log in before the twelve hours had passed, they were instructed to return later. If they attempted to log in after the one-hour grace period, they were blocked from continuing, and were excluded from the study. When they logged back in, they immediately began the retest. They were first tested on the exposure set again, and then on the third set of stimuli (i.e. the set of items they had heard neither during exposure nor during the first test). Participants were therefore tested on the exposure items and different sets of novel items in both test (session 1) and retest (session 2).
At the end of the retest, participants were asked to fill out a questionnaire concerning some basic personal information, as well as their sleep habits and strategies during the task.
On average, the first session (exposure + test) lasted around 20 minutes and the second session (test + questionnaire) 10 minutes.
2.1.3 Participants
Participants were North American English speakers recruited on Amazon's Mechanical Turk platform. ‘Wake’ batches were launched at 16:00 CET (10:00 EST), while ‘sleep’ batches were launched at 04:00 CET (22:00 EST). This meant that half of the participants were recruited in the morning, returning twelve hours later, at the end of their day. The other half were recruited in the evening, completing the first session before going to bed, and then participating in the second session the following morning.
Once recruited, participants were redirected from the Mechanical Turk website to our own website. Upon logging into the website for the first time, they were randomly assigned to the harmony or the disharmony condition. After completing the first session, they received compensation (US$2), and if they completed the second session, they were given a bonus payment (US$3). A total of 723 participants were recruited, but 146 did not complete the first session, 297 did not return for the second sessionFootnote 6 and a further ten completed either too few or too many trials (for instance, by doing the first session twice). Of the 245 participants who correctly completed both sessions, 72 were excluded from data analysis for one or more of the following reasons: they made at least ten errors in the 64 trials during the exposure phase (N = 39), used only one response (i.e. either TEL or TOL) throughout an entire test phase (N = 1), did not fill out the questionnaire (N = 16), napped during the day (N = 11) or did not respond to the question about napping (N = 5), or took notes during exposure (N = 16) or did not respond to the question about note taking (N = 16).Footnote 7 Note that some excluded participants fall into multiple categories (e.g. they took notes during exposure and did not answer the question about napping).
A total of 173 participants, aged 20 to 67 (mean 37; SD 10.0), were included in the data analysis, distributed among the harmony/disharmony and wake/sleep conditions as shown in Table I. Because of the difficulty in recruiting participants using Mechanical Turk who had to return after a specific period of time, we ended up launching a great many more ‘sleep’ batches, so that the total numbers of participants in the sleep groups were larger than those in the wake groups.
Distribution of participants by experimental conditions.
We compared the participants in the wake and sleep groups in a number of ways, based on their responses to the questionnaire, as shown in Table II. For most measures, the two groups did not differ. There was an expected asymmetry between the levels of fatigue at test and retest for the wake and sleep groups. The sleep group (initial test at the end of the day) reported an average higher level of fatigue at initial test than the wake group (initial test in the morning). Likewise, the wake group reported an average higher level of fatigue at their end of the day test (i.e. at retest) than the sleep group. Unsurprisingly, participants were thus more tired during the session they took part in at the end of the day.
Average participant responses to questionnaire by group, and the corresponding test statistics comparing the wake and sleep groups.
2.2 Results
As a sanity check before beginning our planned analyses, we compared the number of TEL responses to the number of TOL responses. Recall that in the input, each suffix was used an equal amount of the time. On average, participants had no preference for one or the other ending (meanTEL 50.2%, SD 8.6).
We calculated the accuracy of responses during the test phase for all participants. These data were analysed using logistic mixed-effects models in R (Bates et al. Reference Bates, Maechler, Bolker and Walker2014), run with bound optimisation by quadratic approximation (BOBYQA; Powell Reference Powell2009) whenever a model did not converge with maximal random-effects structure. Of course, what was considered a ‘correct’ response depended on the rule that participants were exposed to, such that harmonic and disharmonic responses were considered correct and incorrect respectively in the harmonic condition, and vice versa in the disharmonic condition. All factors in all of the models detailed below were defined using contrast coding, and significance was assessed through model comparison following the procedure described in Levy (Reference Levy2014), which entails the removal of factors and interactions in comparison to the full model. Results for exposure items and novel items were analysed separately, and are reported here sequentially. Good performance on the former can be achieved either by learning the rule or by memorising the suffix for each individual item; by contrast, good performance on the latter is an unambiguous indicator of rule learning.
We began by analysing the initial test session only, so as to compare our results with those reported in the literature. Data from the initial test session are qualitatively comparable to results from Pycha et al. (Reference Pycha, Nowak, Shin and Shosted2003), in that they represent performance just after exposure. Mean accuracy from the initial test is displayed in Fig. 1. We designed models with fixed factors for Rule (harmony vs. disharmony), Interval (wake vs. sleep) and their interaction, and random intercepts for Participant and Stem (one model for exposure items, and another for novel items). These models were compared to simpler models that excluded the factor Rule. Unlike in Pycha et al. (Reference Pycha, Nowak, Shin and Shosted2003), performance in the first test phase was better for harmony than for disharmony (exposure items: β = 0.16, SE = 0.06, χ 2(1) = 6.44, p < 0.02; novel items: β = 0.21, SE = 0.07, χ 2(1) = 9.58, p < 0.002). Thus, just after exposure, participants had learned the phonetically natural rule better than the unnatural one. However, no significant difference was observed in the full model for the factor Interval or the interaction between Rule and Interval (both z < 1). In other words, whether participants were tested in the morning or in the evening did not influence the results in any way. The better performance on the harmony rule compared to the disharmony rule is also reflected in the percentage of participants who reached above-chance performance: according to a binomial test, the threshold for individual above-chance performance is 21 correct responses on the 32 exposure items (65.6%) and 39 on the 64 novel items (60.9%).Footnote 8 For harmony, 42% and 37% of participants were at or above threshold on the exposure and novel items respectively; for disharmony, this was the case for only 27% and 12% of participants.
Boxplots showing mean accuracy scores at initial test on exposure and test items as a function of rule.
We then analysed the full dataset, including both sessions; we first consider performance on the exposure items. Mean accuracy across the different manipulations for these items are shown in Fig. 2. A full model was designed that included the following fixed factors: Rule (harmony or disharmony), Interval (wake or sleep), Session (test or retest) and all two- and three-way interactions. The random structure included intercepts for Participant and Stem, which both included random slopes for Session. This model was compared to simpler models that excluded one of the factors or interactions. The full model was found to explain significantly more variance than a model that excluded Rule (β = 0.18, SE = 0.07, χ 2(1) = 7.11, p < 0.008), and than one that excluded Session (β = 0.06, SE = 0.03, χ 2(1) = 4.15, p < 0.05), but no such difference was observed for Interval or any of the interactions (all z < 1). Thus performance was generally better for participants exposed to harmony than for those exposed to disharmony, and there was a general decrease in performance over the course of twelve hours, regardless of whether this period contained a night of sleep or not.Footnote 9
Boxplots showing mean accuracy on exposure items for the wake and sleep groups as a function of rule and test session.
We further examined the extent to which participants’ performance on individual exposure items was stable across the two test sessions. We designed a logistic mixed-effects model with performance on exposure items in Session 2 (retest) as the dependent variable, performance in Session 1 (test) as a fixed predictor and a random intercept for Participant. Participants tended to respond correctly to the same exposure items at test and retest (β = 0.86, SE = 0.06, χ 2(1) = 198.57, p < 0.001), indicating that they were likely relying on having memorised these items rather than extending the phonological rule to them, as they would need to do for novel items.
We next considered the results for novel items. Mean accuracy across the different manipulations for novel items is shown in Fig. 3. Models were designed identically to those for the exposure items, except that there were no slopes for session under Stem in the random structures, since the novel stems at test and retest were not the same (see §2.1.2). The full model was found to explain significantly more variance than a model which excluded Rule (β = 0.23, SE = 0.07, χ 2(1) = 9.63, p < 0.002), but not than a model which excluded Session (β = 0.03, SE = 0.02, χ 2(1) = 1.30, p > 0.1). For the factor Interval and all interactions, no significant effects were observed in the full model (all z < 1).Footnote 10 This indicates that harmony was generally learned better than disharmony, and that unlike for the exposure items, performance for neither rule decreased over the course of twelve hours, regardless of whether this period contained a night's sleep.
Boxplots showing mean accuracy on novel items for the wake and sleep groups as a function of rule and test session.
Overall performance was low, especially for the disharmony groups and even for the exposure items. The most obvious way to evaluate differences in learning over time, however, is to start from a base where all participants demonstrate individual learning and measure how their performance changes. We therefore ran a series of post hoc analyses with the full dataset, but including only those participants who achieved above-chance performance on exposure items in the initial test session (according to a binomial test; see above), thus demonstrating that they had at least learned how the rule applied to stems they had seen before. There were 59 such participants, aged 20 to 59 (mean 39, SD 10.0), and distributed as follows: wake/harmony: N = 13; wake/disharmony: N = 9; sleep/harmony: N = 22; sleep/disharmony: N = 15.
Model design was identical to that used to analyse the data from all participants. As before, we first consider performance on the exposure items. Mean accuracy across the different manipulations for these items is shown in Fig. 4; note that as we only included participants who achieved above-chance performance in the initial test session, the dark grey boxes are all above chance level.
Boxplots showing mean accuracy on exposure items for the wake and sleep groups as a function of rule and test session, for participants with above-chance performance on exposure items at initial test.
The analyses revealed an effect of Session (β = 0.15, SE = 0.06, χ 2(1) = 4.77, p < 0.03), with better performance in test than in retest, and an interaction between Rule and Session (β = 0.12, SE = 0.06, χ 2(1) = 4.28, p < 0.04), but no other main effects or interactions (Rule: β = 0.23, SE = 0.15, χ 2(1) = 2.25, p > 0.1; Rule × Interval: β = 0.19, SE = 0.15, χ 2(1) = 1.65, p > 0.1; all others: z < 1). The interaction between Rule and Session was due to the fact that participants exposed to harmony showed a smaller decrease in performance between test and retest than those exposed to disharmony.
Next, we consider performance on the novel items. Mean accuracy across the different manipulations for these items are shown in Fig. 5. The analyses showed a marginal effect of Rule (Rule: β = 0.38, SE = 0.19, χ 2(1) = 3.73, p = 0.053), and no other main effects or interactions (Rule × Interval: β = 0.23, SE = 0.19, χ 2(1) = 1.33, p > 0.1; Session × Interval: β = 0.07, SE = 0.05, χ 2(1) = 2.33, p > 0.1; all others: z < 1).
Boxplots showing mean accuracy on novel items for the wake and sleep groups as a function of rule and test session, for participants with above-chance performance on exposure items at initial test.
Overall, the analyses on this restricted set of participants who reached above-chance performance on the exposure items in the initial test session confirm a bias in favour of harmony as opposed to disharmony. For the exposure items, they additionally show a larger decrease in performance between test and retest in the disharmony condition than in the harmony condition. Of course, these post hoc analyses entailed a substantial reduction in sample size – to a third of the full sample – but they serve to confirm what our full analysis already shows. That is, while we have no evidence of sleep-related consolidation in our experiment, a consistent bias favouring harmony over disharmony is present in all of our analyses.
3 Discussion
This study has considered the hypothesis that phonetically natural rules might benefit from a bias during learning, such that they are more likely to survive the repeated transmission process than phonetically unnatural rules. Previous examinations of this hypothesis have indeed shown easier or better learning of phonetically natural rules compared to unnatural ones, though the natural rules are almost always less formally complex than the unnatural ones (for a review, see Moreton & Pater Reference Moreton and Pater2012a). In the few studies that tested rules that are logically equivalent in terms of complexity, evidence for a learning bias favouring phonetically natural rules is weak, in particular because of methodological concerns (e.g. predisposition to the natural rule because of L1 experience, low statistical power). The present study focuses on two phonological rules, one natural and typologically recurrent (vowel harmony), one unnatural and exceedingly rare (vowel disharmony), matched in complexity. As weak learning biases are not easy to demonstrate in the lab, and given that at least two previous studies that also considered these rules did not show a difference in learning patterns (Pycha et al. Reference Pycha, Nowak, Shin and Shosted2003, Skoruppa & Peperkamp Reference Skoruppa and Peperkamp2011), we further examined the role of sleep-related memory consolidation. We did not necessarily expect to observe a clear asymmetry between harmony and disharmony at initial test. Rather, we expected that we might observe a numerical trend (like Pycha et al.), and that this small bias might be reinforced by differential memory consolidation after sleep, yielding a boost for the phonetically natural rule but not the unnatural one. Our findings do not entirely align with these predictions, but rather highlight the existence of a phonetically motivated learning bias. Contrary to previous work, we found that the phonetically natural and typologically common rule of vowel harmony was learned better than the phonetically unnatural and exceedingly rare rule of vowel disharmony. However, the extent to which the rule was learned and applied to novel items remained stable over a period of twelve hours, regardless of the absence vs. presence of a night of sleep. We will discuss these findings in turn.
Better learning of vowel harmony in the first test session, i.e. before any potential sleep, contrasts with the two previous studies just mentioned. Here, we observed better performance for the harmony rule than for the disharmony rule during the initial test session for both exposure and novel items. We further observed a clear effect of harmony vs. disharmony when considering the full dataset (i.e. including the retest session), again both for exposure and for novel items. Participants exposed to the phonetically natural harmony rule performed better than those exposed to the unnatural disharmony rule. This, then, is the first solid evidence that harmony is easier to learn than disharmony.
Our sample was considerably larger than those in previous studies, which might have been too underpowered to detect a learning bias. Another methodological difference worth mentioning is that our participants were native speakers of English, tested on French stimuli. Previous research has shown that artificial language learning experiments focusing on phonology can be run with non-native stimuli (Steele et al. Reference Steele, Denby, Chan and Goldrick2015). It could be, though, that such stimuli are actually more appropriate for exploring phonological learning asymmetries, given the hypothesis that their processing is both more phonetic and less likely to be influenced by orthographic knowledge. We leave this to future work, but note the use of non-native stimuli as a potentially crucial manipulation in the present article, allowing us to uncover a learning bias.
Yet, before uncorking the champagne and proposing a toast to phonetic substance in phonological grammar, it is important to consider a plausible alternative explanation for the bias favouring harmony over disharmony. Although English does not have a vowel-harmony rule, it is possible that the English lexicon happens to be organised in such a way that words are more likely to be harmonic than disharmonic. If that is true, participants in our experiment may use their knowledge of English (which biases them towards harmony) and extend that knowledge to the artificial language. At first sight, one way to test for this bias would be by conducting a control experiment without an exposure phase. Participants would be tested on their relative preference for harmony vs. disharmony, rather than on their relative capacity to learn the rules. Note, though, that finding a preference for harmony over disharmony prior to exposure to the pattern in an artificial language would have no consequences for our findings or conclusions. Indeed, general phonetic knowledge (of the kind we hypothesise to be present in phonological grammar) could lead participants to prefer harmonic words over disharmonic ones, and be precisely the driving force that renders the learning of a harmonic pattern easier than the learning of a disharmonic pattern. In other words, a simple preference experiment would not allow us to establish whether the learning bias we found stems from participants’ experience with their native language. Instead, to assess this specific possibility, we conducted a series of analyses aimed at calculating the amount of evidence for harmony compared to disharmony present in the English lexicon. This allows us to measure how much the bias found in our experiment could plausibly be due to knowledge of English. We considered two possible sources of a lexical harmony bias: (i) a relatively high number of harmonic words (i.e. words containing only front or only back vowels), and (ii) a number of harmonic words that is larger than would be expected by chance, given the frequency of the vowels of English.
We extracted all polysyllabic lemma forms from the CMU pronouncing dictionary (2008). (Monosyllabic words are not informative with regards to vowel harmony, and were therefore excluded from the present analysis.) For each word, a harmony score was calculated, where each vowel within the word was assigned a score, 0 or 1, depending on whether it was front or back. Schwas were not considered in the analysis, as it is unclear whether they should be considered as front or back. The harmony score of a word was calculated as the variance of backness of its vowels, such that a word with three front vowels (0, 0, 0) or three back vowels (1, 1, 1) has a harmony score of 0 (no variance in backness), while a word with one front and two back vowels (0, 1, 1) has a score of 0.22. This gives a distribution of harmony scores bounded between 0 and 0.25 (regardless of word length), with a word containing as many front as back vowels having a score of 0.25.
The distribution for words in English can be seen in Fig 6a. Figure 6 shows that only around 35% of English polysyllabic words are harmonic.Footnote 11 To ensure that this is not driven by systematic pressure that participants might be able to pick up on, we generated a random version of English by extracting all vowels from all words, shuffling them and reinserting them into the consonant frames. A word like moodiness /mudinɛs/ could therefore become /mædunɔs/ or /mɪdʌnis/, etc. We then recalculated the distribution of harmony scores for this random English lexicon and compared it to the real English lexicon. As can be seen in Fig. 6b, the random distribution is almost completely identical to the real distribution; in fact, only 1.74% of the real distribution is not overlapping with the random one.
Distributions of palatal harmony scores for real and random lexicons in English (no vowel harmony rule) and Hungarian (productive vowel-harmony rule). Each bar represents the percentage of the lexicon that falls into that bin. A score of 0 means a word is fully harmonic and a score of 0.25 means a word contains as many front as back vowels.
To ensure that our measure properly captures explicit harmony, we also performed the same analysis on the lexicon of a language with productive vowel harmony: Hungarian. The real Hungarian lexicon should contain more harmonic words than a random lexicon made by shuffling the vowels. Indeed, Hungarian has grammaticalised a pressure for words to contain vowels of only one kind (productive vowel harmony). Typically, a word can contain only front or only back vowels. This leads to morphophonological alternations as in (1) above, but also affects roots (both roots in that example are themselves harmonic). We used a Hungarian pronunciation dictionary (see Grimes Reference Grimes2010) to extract the phonological transcription of words. We followed the traditional division of Hungarian vowels, with /y yː ɛ eː ø øː/ considered front, /u uː o oː ɒ aː/ back and /i iː/ neutral. As can be seen in Fig. 6c, the Hungarian lexicon is overwhelmingly harmonic, with over 70% of words containing only front or back vowels. We then tested whether this distribution is a result of an explicit pressure by creating a random version of the Hungarian lexicon, just as we did for English above. As can be seen in Fig. 6d, a random Hungarian lexicon shows a great many more disharmonic words, and far fewer harmonic words; the bar on the left is reduced by nearly a half, with only 36% of the random words containing only front or only back vowels, compared to over 70% in the real lexicon. Our method therefore appears able to distinguish between chance harmony (due to the statistical distribution of vowels) and systematic harmony (due to an active process in the language).Footnote 12
Our reasoning has considered two ideas: the presence of fully harmonic words (made up of only front or only back vowels), and a pressure for the vowels to cluster in a certain way above what would be expected by chance, given their individual frequencies. While it may be enough to consider that the number of harmonic words in English does not provide speakers with enough evidence of harmony to explain our experimental results (only 35% of English words are harmonic), our lexical analyses further considered the possibility that, despite the relatively few harmonic words in English, there are still more harmonic words than would be predicted by chance. This, too, does not seem to be the case in the English lexicon (though it very clearly is in the Hungarian lexicon). All in all, we conclude that the experimental results reported above cannot simply be explained by a language-specific bias whereby learners fare better on whatever pattern is most frequent in their native lexicon.
Our results strongly suggest, then, that, compared to vowel disharmony, vowel harmony has a preferential status in learning. We argue that this is due to the fact that vowel harmony is phonetically natural (grounded in phonetic substance), whereas vowel disharmony is not. This type of phonetically based bias could play a role in explaining why phonetically unnatural patterns like vowel disharmony are rare in the world's languages. Such a bias could come into play, for example, in the case of ambiguous input, with multiple plausible representations. In such cases, a learner could use their phonetic knowledge to determine that one representation is more likely than another. This type of substantive bias could complement what Blevins (Reference Blevins2004) refers to as chance, where one representation is randomly chosen over another. In cases where competing representations are equally likely (i.e. phonetically plausible), a learner might indeed assume one as frequently as another. However, in cases where one representation is phonetically more natural than a competing one, a phonetically based learning bias, encoded in synchronic grammar, might lead a learner to favour the former. This would in part explain why such phonetically grounded (natural) patterns are typologically more frequent than their unnatural counterparts. While diachronic explanations for sound patterns have a great deal of explanatory power, dismissing synchronic explanations entirely is overly restrictive, especially in light of our experimental results.
A further point to consider, given previous work highlighting the importance of simplicity bias in language (e.g. Moreton & Pater Reference Moreton and Pater2012a, Culbertson & Kirby Reference Culbertson and Kirby2016), is how we have considered complexity. The rules we tested were of equal complexity (i.e. they were both defined by a single feature), but could asymmetries amongst specific phonological features also affect the perceived complexity (and learnability) of patterns in general? A number of previous studies have shown that a difference in one featural dimension is not necessarily perceived as equivalent to a difference in another featural dimension, and that this can be the result of both universal and language-specific phonetic properties (Cole et al. Reference Cole, Jakimik and Cooper1978, Ernestus & Mak Reference Ernestus and Mak2004, White Reference White2014, Martin & Peperkamp Reference Martin and Peperkamp2017). In our design, this is not of great import, since our harmony and disharmony patterns involve the same features; only the direction of the alternation changes. But it is interesting to consider whether a phonological pattern involving a perceptually distinct feature would be more difficult to learn than one involving a feature with less perceptually distinctive phonetic correlates. This is precisely the argumentation provided by White (Reference White2014) to explain why participants in his study more readily learned a voicing alternation than a manner of articulation alternation (though both involved only one feature). If learners use their phonetic knowledge in learning phonological alternations, such differences are to be expected, and could even depend on modality. That is, an alternation that is more distinct articulatorily might be harder to learn if the task requires the participants to produce the items.
Turning finally to the learning we did and did not observe, the design we used allows us to highlight the fact that performance on exposure items and on novel items likely involves different underlying mechanisms: memorisation on the one hand and rule learning on the other. Indeed, performance on individual exposure items in the first and the second test sessions showed a correlation; that is, the items that participants performed well on during retest tended to be ones they had also performed well on during the initial test. Moreover, even though we did not perform any statistical analyses to compare performance on exposure and novel items, it is worth noting that for the former performance generally worsened between test and retest, whereas for the latter it neither improved nor deteriorated.
The finding that harmony is learned better than disharmony made it more difficult to study our second question (i.e. whether the absence vs. presence of sleep would differentially affect learning), since baseline performance before the twelve-hour interval differed between participants exposed to harmony and those exposed to disharmony. Moreover, the number of participants overall who reached above-chance performance during the first session was low. Yet our post hoc analyses on the subset of participants with above-chance performance on exposure items in the initial session showed the same results as analyses on the entire sample: for both the harmony and disharmony conditions, performance on novel items was stable across the two test sessions. If consolidation generally helped, without differentially benefitting one of the two rules we tested, we should have seen an interaction between Session and Interval, but no such interaction was observed. We also did not observe a triple interaction, with sleep benefitting one of the rules more than the other. Thus we observed no evidence of consolidation at all. However, the fact that performance on novel items was stable across the two test sessions is a robust result, and the first to show that phonological rule learning in an artificial language learning paradigm is not entirely ephemeral. Of course, it would be interesting to see how long the learning effect survives.
Although we found no evidence for consolidation, it would be premature to claim that phonological rule learning cannot benefit from sleep-related memory consolidation, especially given evidence for such consolidation in other cases of linguistic learning, such as the learning of novel phonotactic constraints (Gaskell et al. Reference Gaskell, Warker, Lindsay, Frost, Guest, Snowdon and Stackhouse2014) or morphosyntactic rules (Batterink et al. Reference Batterink, Oudiette, Reber and Paller2014).Footnote 13 In light of our use of non-native stimuli, it is also worth noting that in the domain of word learning, sleep-related consolidation has been shown to play a special role for words containing non-native sounds (Havas et al. Reference Havas, Taylor, Vaquero, de Diego-Balaguer, Rodríguez-Fornells and Davis2018). The absence of consolidation in our study is, like any null effect, particularly hard to interpret. There are at least two aspects of our manipulation that might have affected our ability to observe consolidation. First, the population we tested was on average older than the typical undergraduate population (mean = 37), and memory consolidation tends to decline after young adulthood (Scullin & Bliwise Reference Scullin and Bliwise2015). Second, performance was generally fairly low (even though above chance level in the first session for all groups). It is worth noting that both Gaskell et al. (Reference Gaskell, Warker, Lindsay, Frost, Guest, Snowdon and Stackhouse2014) and Batterink et al. (Reference Batterink, Oudiette, Reber and Paller2014) tested young adults with tasks on which the participants reached high accuracy. It would thus be appropriate for future research to lengthen the exposure phase and/or use a phonological rule that is more easily learnable (while still being difficult enough to observe variation in the population), as well as to select participants in their early twenties.
If in future experiments we do see a difference in performance according to the presence vs. absence of sleep between testing sessions, the causal role of sleep would still need to be confirmed. Indeed, rather than the presence of sleep per se, it might be the absence (or the reduced amount) of language processing during sleep that accounts for such difference. A reliable way to demonstrate the role of sleep involves recording EEG (electroencephalography) during sleep and carrying out individual correlation analyses; that is, sleep-dependent consolidation is revealed by a positive correlation between the amount of rapid eye movement (REM) sleep and/or slow-wave sleep on the one hand and the improvement in performance following a period of sleep on the other hand (Walker & Stickgold Reference Walker and Stickgold2004, Nishida & Walker Reference Nishida and Walker2007, Tamminen et al. Reference Tamminen, Payne, Stickgold, Wamsley and Gareth Gaskell2010, Batterink et al. Reference Batterink, Oudiette, Reber and Paller2014, Gaskell et al. Reference Gaskell, Warker, Lindsay, Frost, Guest, Snowdon and Stackhouse2014). Thus, while for practical reasons we opted for online testing, it would eventually be necessary to resort to a lab-based study.
A final methodological note is warranted. While we had hoped that testing participants online would avoid the difficulties of bringing participants to the lab at precise twelve-hour intervals, we found that our strict test/retest set-up was not ideal for Mechanical Turk. We wound up needing to recruit an excessively large number of participants, simply because many did not return within the specified window. This was despite explicit instruction, monetary incentive (the second session paid a bonus that was larger than the original payment) and precisely timed automated reminder e-mails. We thus urge caution when using this platform in cases where participation in multiple test sessions depends on precise timing.
4 Conclusion
We investigated whether phonetically natural rules benefit from a learning bias, and, if so, whether this bias stems from, or is enhanced by, consolidation during sleep. The results provide clear evidence in favour of a learning bias, but no evidence for a role for sleep. In fact, sleep did not appear to enhance learning at all, regardless of whether the rule to be learned was phonetically natural or unnatural. The question of whether or not sleep-dependent consolidation plays a role in shaping sound patterns in human language thus remains open. However, our results clearly demonstrate that a phonetically natural rule is learned better than a phonetically unnatural rule. For the first time, we have clear evidence that vowel harmony (phonetically natural) is easier to learn than vowel disharmony (phonetically unnatural). This in turn may well affect the way such systems evolve during the repeated transmission process, influencing the asymmetrical distribution of phonetically natural rules compared to unnatural ones in the typology.
