Partial reduplication often depends on a segmental melody: a sequence of segment variables present in the reduplicative form are filled by a corresponding segmental sequence in the base form (Marantz, 1982), regardless of syllable structure. That is, the segmental melody doubled in the reduplicative form may cross syllabic boundaries in the base form. For example, Agta has a reduplication process by which a CVC string is copied, regardless of syllable structure (Healey, 1960):

In contrast, Tohono 0’odham has a reduplication process for pluralization by which a CV string is copied, again regardless of syllable structure (Raimy, 2000):

Although many formal theories of reduplication make reference to syllable structure, they generally do not permit direct copying of the syllable from the base to the reduplicant (e.g., McCarthy & Prince, 1995). This aligns with the observed typological pattern: “syllable-copy reduplication” is largely unattested in natural languages. However, studies by Haugen (2003, 2014) and Haugen and Kennard (2011) provide documentation and analysis of the only known case of syllable-copy reduplication: Hiaki. In this language, the habitual verb form involves the copying of the complete initial syllable from the base. Thus, if the base-initial syllable is CV, the reduplicant will be CV; if it is CVC, the reduplicant will be CVC:

Therefore, the theoretical constraint described above – that the syllable structure of the base is not directly copied – is violated by Hiaki. A critique of existing theories is beyond the scope of this article; the crucial point is that syllable-copy reduplication, while attested, is exceedingly rare.

Intuitively, it is remarkable that there is a cross-linguistic tendency for partial reduplicants to discount – and often override – the syllable structure of the base form. After all, the syllable is a well-defined constituent in phonology, and many phonological rules and constraints make reference to it (for an overview, see Blevins, 1995). Furthermore, there is evidence from language games, or ludlings, that speakers spontaneously manipulate syllables. For example, In the Colombian version of the Spanish language game Jerigonza, /p/ and a copy of the preceding vowel are added after every syllable; for instance, /ma.es.tro/ ‘teacher’ becomes /ma.pa.es.pe.tro.po/ (Piñeros, 1998). Thus, under the reasonable assumption that the syllable is a cognitively accessible unit, syllable-copy reduplication is not more complex than string-copy reduplication. Similarly, syllable-copy reduplication does not appear to be phonetically unnatural; on the contrary, repeating a syllable is economical in terms of both planning and articulatory effort. This raises the following question: despite the apparent mismatch between the rarity of syllable-copy reduplication and the naturalness of syllable repetition in general, might the typology still be driven by an underlying learning bias?

To our knowledge, only one study has applied AGL to directly investigate cognitive mechanisms that may drive the rarity of syllable-copy reduplication: Haugen et al. (2022). This study investigated the learnability of syllable-copy reduplication using a training + feedback paradigm. In this paradigm, participants perform a forced-choice task and get feedback (correct/incorrect) on every trial. Thus, they start by making guesses; the question is to what extent and how fast they learn the pattern, based on the feedback. In their study, monolingual native English speakers were presented with written trisyllabic nonce words, the first syllable of which was either of the form CV or CVC. For each item, participants had to choose among three reduplicative options, one of which was full reduplication of the base word, one a partial reduplication of the first two segments, and one a partial reduplication of the first three segments. In half of the trials, participants had to identify the plural form; in the other half, the augmentative form. For instance, for the item ‘dog’ they were asked either ‘What is the word for dogs?’ or ‘What is the word for big dog?’. The correct plural form was always the one with full reduplication; this condition served as a baseline. The correct augmentative form depended on the participant’s group. In two string-copy groups – CV and CVC – it was determined by copying a segmental melody of the base form, regardless of syllable boundaries; in the syllable-copy group – CVX – it was determined by directly copying the first syllable of the base, similar to Hiaki.1 Thus, in the CV and CVC groups, the reduplicant always consisted of either a light or a heavy syllable, respectively, whereas in the CVX group, some reduplicants consisted of a light syllable and others, of a heavy syllable, depending on the syllable structure of the base. Examples are shown in Table 1. Note that, in case of syllable-copying, the augmentative form copies CV if the first syllable of the base is CV (nekola) and CVC if the first syllable of the base is CVC (vamseta).

The results showed no difference between the full reduplication and the two string-copy patterns, which were all learned faster and better than the syllable-copy pattern. Given the crucial difference between string-copy and syllable-copy reduplication, the authors concluded that an underlying learning bias disfavors syllable-copy reduplication. In turn, they argued that this cognitive constraint against syllable-copy reduplication may drive its observed typological rarity.

A second set of AGL experiments concerning reduplication was conducted for the dissertation of Wang (2024); while these experiments were not designed to directly investigate the rarity of syllable-copy reduplication, they have bearing on the topic. The experiments focused on how native English speakers generalize an artificial reduplicative grammar to novel base structures. Stimuli were presented auditorily, and participants performed a production task. This study did not directly compare the learnability of syllable-copy reduplication to string-copy reduplication, but some of the experiments incidentally converged on results that are largely consistent with those of Haugen et al. (2022). Specifically, in Experiments 1a-c, participants were exposed to CVC reduplication (e.g., [dɔvgə] → [dɔv-dɔvgə]). During the test phase, a strong majority of participants generalized CVC reduplication to novel stimuli, including both those with an initial CVC-syllable and those with an initial CV-syllable (e.g., [dəbeɪ] → [dəb-dəbeɪ]). Only a small minority produced syllable-copy reduplication (i.e., reduplicate CVC when the initial syllable of the base form is CVC, reduplicate CV when the initial syllable of the base form is CV). Vice versa, in Experiment 2c, participants were exposed to CV reduplication (e.g., [pif] → [pipif]), and during the test phase, the majority of them produced CV-copy reduplication for novel test items, including those with an initial CVC-syllable ([dəbgɪv] → [də-dəbgɪv]); only a small minority consistently copied the first syllable of the base forms. Thus, participants across these experiments seemed to generalize string-copy reduplicative forms to the test phase, in line with both the observed typology and the results of Haugen et al. (2022): responses were suggestive of a cognitive constraint against syllable-copy reduplication.

Although the previous findings indeed suggest that the typological rarity of syllable-copy reduplication may be driven by an underlying learning bias, Wang’s (2024) experiments do not directly address this issue, and Haugen et al.’s (2022) findings are limited by a number of methodological and theoretical considerations. For example, Haugen et al. (2022) relied exclusively on written stimuli, whereas the phenomenon under consideration is a morphophonological one, not an orthographic one. Biases in the written modality do not necessarily reflect the auditory modality, particularly with regard to phonology: languages are generally acquired and transmitted via auditory (or, in the case of sign languages, visual) input. Therefore, findings observed for orthographic stimuli might not generalize to linguistic cognition more broadly – and, if so, may be insufficient to account for the typological pattern. Haugen et al. (2022) acknowledge this shortcoming. Here, we include experiments that test for the presence of a learning bias in the auditory modality.

A second important issue concerns potential differences between biases in speech perception and biases in speech production. The cognitive pressures that operate on speech perception are distinct from those that operate on speech production (e.g., minimizing articulatory effort is relevant for speakers, but not for listeners), and, indeed, some prior AGL studies in the realm of phonology have found differing patterns of results in perception versus in production (Do & Havenhill, 2021; Peperkamp et al., 2006). This raises the possibility that typological trends are selectively driven by one or the other. Haugen et al. (2022) relied on perception, while Wang (2024) relied on production; neither study compared the two. Here, we directly compare performance on a perception and a production task, using the exact same exposure.

In addition to these substantial extensions, we implement several smaller methodological improvements. Since Haugen et al. (2022) tested English speakers, their stimuli were written in English orthography, which is notoriously opaque. Consequently, the syllable boundaries in many of their stimuli were ambiguous – e.g., the CVC reduplicative form of hitmolu was hithitmolu, which could be interpreted as hi.thit.mo.lu or hit.hit.mo.lu. To address this, we test speakers of a language with a transparent orthography, i.e., Italian, on items that are phonotactically legal in their language. In addition, whereas Haugen et al.’s (2022) participants were undergraduate linguistics students, we draw from a more diverse sample of participants recruited online via Prolific. Furthermore, they provided participants with explicit feedback after each trial to prompt learning, which has the advantage of enabling the researchers to model learning curves. However, exposure to large quantities of explicit, negative feedback is rare in naturalistic learning settings (Chomsky, 1965, 1980); therefore, in our design, we implement separate exposure and test phases, both without feedback.

A final consideration concerns response strategies. In Haugen et al. (2022), word length was perfectly predictable: base forms were either of the form CVC.CV.CV or CV.CV.CV. Therefore, base forms with a CV initial syllable were always six letters; base forms with a CVC initial syllable were always seven letters. Correspondingly, the CVC-reduplicated form was always either ten or nine letters, and the CV-reduplicated form was always nine or eight letters, contingent on the initial syllable of the base form. Therefore, it may have been relatively easy for participants to attend to spurious cues as an effective response strategy. For instance, in the CV condition, the correct augmentative form of a six-letter base word was always the eight-letter option. We examine whether the findings of Haugen et al. (2022) would generalize to stimuli with higher variability – and, hence, whether their original results may have been influenced by a spurious response strategy.

Below, we report six online experiments. Experiments 1 and 2 use written stimuli, aiming to determine whether Haugen et al.’s (2022) results generalize when accounting for the smaller methodological issues mentioned above. Experiments 3-6 use auditory stimuli and directly contrast perception and production tasks. In all the experiments, our primary research question is whether syllable-copy reduplication is more difficult to learn than string-copy reduplication, where – like Haugen et al. (2022) – we implement the latter as CV- and CVC-reduplication for two different participant groups. Given this research question, we aim at equal sample sizes in the syllable-copy condition, on one hand, and the string-copy condition, on the other hand, even though the string-copy condition includes two subconditions. However, one may also wonder whether there is a learnability difference between these subconditions, i.e. CV-or CVC-reduplication. Even though the sample sizes for this comparison are necessarily smaller, we consider this secondary research question as well, and, hence, additionally compare the two string-copy subconditions to one another. Note, though, that we have no hypothesis regarding a potential difference between these two subconditions, and, hence, formulate no predictions – these additional analyses are exploratory.

The primary GLMM revealed that accuracy was lower in the syllable-copy condition than in the string-copy condition (β = 1.66, OR = 5.24, SE = 0.65, χ² = 6.54, p < .05). The secondary GLMM identified no significant difference between the string-copy subconditions (χ² < 1). These results (Figure 3A) are consistent with Haugen et al. (2022) and Wang (2024): participants exhibit a learning bias against syllable-copy reduplication.

In their posttest responses, participants reported relying on a wide variety of strategies in the test phase. For example, consider the following responses to (all translated from Italian) “What strategy did you use to choose between the options? Did you find a specific rule?”

These participants report pronouncing the words (either overtly or covertly); hence, they were attending to how the words sound. However, other participants reported relying on general syllabification rules:

For these participants, it is ambiguous whether they actually attended to how the words would be pronounced, as they may have referred to the syllabification of written words. Moreover, consider the following three responses:

These participants paid attention exclusively to orthography rather than inferred phonology. Presenting the stimuli auditorily may yield different results for these participants in particular. Furthermore, the relative simplicity of some of the strategies (e.g., simply repeating the first letters) highlights the concern that responses to the stimuli may be influenced by their lack of variability. Experiment 2 attempts to directly address this latter concern.

The primary GLMM identified no significant difference between the string-copy and syllable-copy conditions (χ² < 1), and the secondary GLMM identified no significant difference between the string-copy subconditions (β = 1.00, OR = 2.72, SE = 0.60, χ² = 2.71, p = .099). These results (Figure 3B) contradict both Experiment 1 and Haugen et al (2022): participants did not exhibit a learning bias against syllable-copy reduplication. In particular, the numeric difference between mean accuracy in the string-copy and syllable-copy conditions was negligibly small in this experiment (<1%), despite a larger sample size and higher power than Experiment 1. This suggests that the failure to replicate the finding in this experiment is due to the difference between designs rather than, e.g., noise in significance-based comparisons (Vasishth et al., 2018). Moreover, whereas the difference between the two string-copy subconditions was not significant, the magnitude of the effect (as quantified by the odds ratio) was moderate (Rosenthal, 1995); for subsequent exploratory analyses, see 8.2.

At first glance, these results seem to suggest that there is a complex interdependence between typology, learning biases, and syllable-copy reduplication: the learning bias against syllable-copy reduplication applied to the highly simple items in Experiment 1, but it did not generalize to the more variable ones in this experiment. This suggests that previous findings have been the spurious result of participants adopting metalinguistic strategies that are easier to apply to simple stimuli, casting doubt on prior findings about the relationship between reduplicative typology and cognitive learning biases. However, there is a crucial factor that neither Experiment 1 nor Experiment 2 accounted for: modality. What results do we observe when we use audio recordings, rather than written words? A compelling prediction is that, given more variable stimuli, the underlying learning bias is masked by the limitations of written stimuli – and will reemerge with auditory input.

The results (Figure 4A) are consistent with Experiment 2 rather than Experiment 1: accuracy was not significantly lower in the syllable-copy condition than in the string-copy condition (χ² < 1), and there was no significant difference between string-copy subconditions (β = 0.76, OR = 2.15, SE = 0.46, χ² = 2.65, p = .104). Notably, the effect sizes of both comparisons were comparable to those in Experiment 2 (negligible for the former, and moderate for the latter). There were also no significant differences between the two stress conditions (χ² < 1). This suggests that the results of Experiment 2 generalize to the auditory modality. More broadly, they corroborate the conclusion that the perceptual learning bias against syllable-copy reduplication fails to replicate, given more variable stimuli.

However, this seems to introduce an empirical contradiction: Wang (2024) found results consistent with Haugen et al. (2022), despite using relatively variable stimuli; why are our results inconsistent with their convergent findings? One possibility is that Wang’s (2024) alignment with Haugen et al. (2022) was spurious. In line with this, recall that she did not directly test for a learning advantage of string-copy reduplication over syllable-copy reduplication; this finding was secondary to her primary empirical targets. Another possibility lies in the fact that she used a speech production task rather than a perception task. As discussed in 1.3, the cognitive mechanisms underpinning speech perception and production are distinct, and AGL studies have revealed disparities between phonological learning biases as a function of task modality (Do & Havenhill, 2021; Peperkamp et al., 2006). Thus, before concluding that there is no learning bias that can account for the extreme cross-linguistic rarity of syllable-copy reduplication, we examine in the next experiment if such a bias can be observed in a speech production task.

Among the data from participants included in the experiment, 2.65% of transcriptions were coded as “other” (i.e., failing to produce any reduplicative form or producing a reduplicative form other than CV/CVC). These trials (automatically counted as incorrect) were included in the analysis; excluding them does not yield any differences in model outcome direction or significance.

Accuracy was significantly lower in the syllable-copy condition than in the string-copy condition (β = 1.64, OR = 5.17, SE = 0.28, χ² = 31.13, p < .001), but there was no significant difference between string-copy subconditions (χ² < 1) (Figure 4B). There was also no significant difference between stress conditions (β = 0.36, OR = 1.43, SE = 0.32, χ² = 1.25, p = .263). This provides a striking contrast to Experiment 3: the learnability of the syllable-copy condition was strongly reduced, relative to both of the string-copy subconditions (direct comparisons between these two experiments are described in 8.1). Crucially, these two experiments shared identical exposure phases, so the disparity is most likely due to the methodological difference in the test phases: production versus perception. This suggests that the typological rarity of syllable-copy reduplication is driven by an underlying cognitive learning bias, but that bias selectively operates in speech production. In turn, this provides a clear resolution to the seemingly conflicting findings between our previous three experiments and the previous two studies: when accounting for stimulus variability, the finding originally attested by Haugen et al. (2022) does not replicate in perception tasks, but does replicate in production tasks, in line with Wang (2024) – a surprising finding, that is, however, reminiscent of two other studies that compared perception and production in phonological rule learning. First, Peperkamp et al. (2006) observed better learning of a phonetically natural rule, compared to a matched unnatural one, when using a production task, but not when using a perception task. Second, using only a perception task, Do and Havenhill (2021) found better learning of a phonetically natural rule only when participants had to repeat items out loud during the exposure phase.

Beyond a simple analysis of accuracy, phonological production data allow for a richer analysis of potential biases. Specifically, it is possible that errors in the syllable-copy condition are directionally biased: individual participants may tend to rely on either a CV-copy or a CVC-copy strategy, resulting in a mix of incidentally accurate responses (e.g., when a participant using a CV-copy strategy encounters a CV-initial base form) and inaccurate responses (e.g., when a participant using a CV-copy strategy encounters a CVC-initial base form). If there is indeed a systematic bias in favor of string-copy reduplication over syllable-copy reduplication, we would predict that, for a given participant in the syllable-copy condition, errors are disproportionately biased towards either CV or CVC forms (rather than a mix of the two). This leads to a predicted relationship between accuracy and response bias in the syllable-copy condition: for participants with high accuracy, there will be an even mix of CV and CVC responses (by necessity, as the correct CVX reduplicated forms require a mix of the two); as accuracy decreases, responses diverge towards mostly CV responses and mostly CVC responses.

In order to test this, we ran a non-preregistered exploratory analysis: We computed for each participant in the syllable-copy condition a copy strategy bias index (CSBI), defined as the number of trials in which participants erroneously produced a CVC reduplicated form minus the number of trials in which participants erroneously produced a CV reduplicated form. For example, a maximum score of 10 indicates that a participant erroneously produced CVC forms for all 10 trials where CV forms would have been correct, but correctly produced CVC forms for all 10 trials where CVC forms were correct – in other words, always produced CVC. The minimum score of –10 indicates the opposite (i.e., that participants always produced CV). A score of 0 indicates that a participant either made no errors or made an even mix of CVC and CV errors. The expected pattern is evident in a visualization of the CSBI data: as accuracy decreases, CSBI diverges towards 10 – reflecting a CVC-copy strategy – and –10 – reflecting a CV-copy strategy (Figure 5). This also leads to a specific quantitative prediction: the absolute value of CSBI should be inversely correlated with accuracy. This prediction was borne out in a Pearson correlation analysis (r = –0.90, SE = 0.06, t(48) = 14.08, p < .001). Therefore, beyond our main results, we even see a bias towards string-copy reduplication strategies within the syllable-copy condition.5

The learning bias against syllable-copy reduplication revealed by this experiment calls for a reconsideration of data from our earlier orthography-based experiments. Because the stimuli from those experiments were written words, some participants may have either “mouthed” the words silently or pronounced the words out loud – thereby producing the plural forms. If this is the case, we would predict that these participants may have selectively been subject to the learning bias observed here, even if it was absent in the experiment as a whole. One of the posttest strategy questions posed to participants in Experiment 16 has bearing on this possibility (translated from Italian): “During the first stage, did you say the words (silently or aloud)?” The response options were: “yes, always” (selected by 36% of participants); “yes, only for singular” (0%); “yes, only for plural” (12%); “yes, sometimes” (32%); and “no” (20%).

Based on their responses, participants can be divided into two groups: one group which consistently produced the plural form (encompassing “yes, always,” and “yes, only for plural”; N = 24) and one group which did not, or only inconsistently (encompassing “no” and “yes, sometimes”; N = 26). We carried out two separate exploratory analyses, one for each group: a logistic regression (binary logit model) with fixed factor Main Type (string-copy vs. syllable-copy). The regression for the “no/yes, sometimes” group included random intercepts for subject and item; the random intercept for item was removed from the “yes, always/yes, only for plural” group, in order to avoid a convergence failure that could not be resolved by adding the BOBYQA optimizer. These models (Figure 6) revealed a significant bias against syllable-copy reduplication in the “yes, always/yes, only for plural” group (β = 3.70, OR = 40.40, SE = 1.19, χ² = 9.59, p < .01), but not in the “no/yes, sometimes” group (χ² < 1). This suggests that the production-dependent learning bias revealed by the present experiment is observable even in our first experiment, but was originally concealed by the inconsistency of strategies used by participants in its orthography-based paradigm.

Our experiments so far are indicative of the presence of a selectively production-driven learning bias in favor of string-copy over syllable-copy reduplication. However, beyond the production-perception distinction, there is one methodological concern that remains unaddressed: sampling bias. Despite the inclusion of a total of 20 base forms and 20 reduplicative forms in the exposure phase, a remarkable number of participants (44) were excluded and replaced, due to failing to produce reduplication in the majority of trials. In the previous experiments, the two-alternative forced-choice task did not allow us to identify participants who may not have learned any sort of reduplication in the exposure phase. Therefore, it is possible that in the present experiment, we ended up with a more “conscientious” sample of participants than in the previous ones, leading to an asymmetric comparison. In order to address this issue, it is necessary to create an easier production experiment, such that fewer participants will fail to produce a substantial amount of reduplicated forms, as well as a matched perception experiment to allow for an appropriate comparison between perception and production. To this end, Experiments 5 and 6 are matched perception and production experiments that differ from Experiments 3 and 4 in the exposure phase.

The results (Figure 4C) are consistent with Experiment 3: accuracy was not significantly lower in the syllable-copy condition than in the string-copy condition (β = 0.57, OR = 1.77, SE = 0.30, χ² = 3.48, p = .062), and there was no significant difference between string-copy subconditions (β = 0.74, OR = 2.09, SE = 0.44, χ² = 2.63, p = .105). This suggests that the results of Experiment 3 generalize when the relevant grammatical feature – reduplication – is made more explicit in the exposure phase.

The results (Figure 4D) are consistent with Experiment 4: accuracy was significantly lower in the syllable-copy condition than in the string-copy condition (β = 1.86, OR = 6.40, SE = 0.35, χ² = 29.41, p < .001), and there was no significant difference between string-copy subconditions (β = 1.39, OR = 4.00, SE = 0.75, χ² = 3.65, p = .056). This suggests that the findings of Experiment 4 generalize when highlighting the salience of reduplication in the exposure phase. Crucially, the present experiment – in tandem with Experiment 5 – was not subject to the same asymmetric exclusion volumes as Experiments 3 and 4. Only nine participants in the present experiment were excluded due to a failure to produce reduplication in the majority of trials; this suggests that our manipulation of the exposure phase was successful in making reduplication more explicit as the relevant grammatical feature. Therefore, this experiment has a comparable sample to Experiment 5, in turn corroborating the pattern observed in Experiments 3 and 4: we reliably identify a selectively production-driven learning bias against syllable-copy reduplication. In fact, this experiment yielded a numerically stronger effect size for the string-copy vs. syllable-copy comparison than Experiment 4 (6.40 vs. 5.17), confirming that the prior finding was not inflated by the higher exclusion rate. The present experiment also exhibited the same pattern of error bias as Experiment 4: in the syllable-copy condition, participants systematically diverge towards CVC-copy and CV-copy reduplicative strategies as accuracy decreases (Figure 7). This is supported by a Pearson correlation analysis, identifying an inverse relationship between accuracy and absolute CSBI (r = –0.90, SE = 0.06, t(48) = 14.68, p < .001). As in Experiment 4, this further illustrates the robustness of the underlying learning bias: even in the syllable-copy condition, participants still gravitate towards string-copy reduplicative strategies.

In order to compare performance in perception vs. in production, we constructed two GLMMs (logistic regressions), one over the combined results of Experiments 3 and 4, and one over those of Experiments 5 and 6. The models had Task (perception vs. production), Main Type (string-copy vs. syllable-copy), and their interaction as fixed factors, and included random intercepts for Item and Participant. For Experiments 3 and 4, the model revealed a main effect of Task (β = 1.58, OR = 4.85, SE = 0.28, χ² = 21.13, p < .001), with higher accuracy in production, and a significant interaction between Task and Main Type (β = 1.39, OR = 4.00, SE = 0.38, χ² = 13.14, p < .001). Post-hoc analyses carried out with the emmeans package (Lenth et al., 2025) showed that, for string-copy reduplication, accuracy was higher in production than in perception (EMM = 1.58, SE = 0.28, z = 5.71, p < .001), while for syllable-copy reduplication, there was no difference between tasks (EMM = 0.19, SE = 0.26, z = 0.76, p = .871). The analysis of Experiments 5 and 6 likewise yielded a main effect of Task (β = 1.16, OR = 3.18, SE = 0.34, χ² = 5.55, p < .05), again with better performance in production, and a significant interaction between Task and Main Type (β = 1.25, OR = 3.50, SE = 0.45, χ² = 7.78, p < .01). While effect sizes were marginally smaller for these comparisons than the corresponding comparisons for Experiments 3 and 4, this may be due to higher accuracy across conditions in Experiments 5 and 6 (i.e., a ceiling effect – although both the main effect and interaction are still significant). Once more, for string-copy reduplication, accuracy was higher in production (EMM = 1.16, SE = 0.34, z = 3.44, p < .01), while there was no difference for syllable-copy reduplication (EMM = 0.10, SE = 0.30, z = 0.32, p = .988).

The fact that performance was better in production than in perception is remarkable, as generating a correct form is intuitively more difficult than merely recognizing one. And, indeed, in a study of phonological rule learning mentioned earlier, higher performance was found in a perception task, compared to a production task, despite identical exposure (Peperkamp et al., 2006). Why we found the opposite pattern is unclear. The large number of participant exclusions in Experiment 4 may have yielded a more conscientious group than in Experiment 3, which could account for the higher overall accuracies in the former. However, this explanation cannot account for the same difference between Experiments 5 and 6, which were designed to ensure comparable samples. Therefore, it is possible that some aspect of reduplication itself facilitates production over perception – e.g., perhaps being forced to produce the plural forms (or anticipating, during the exposure phase, that the test phase will involve production) draws attention to the phonological make-up of the words that is more likely to be ignored in a purely perceptual task. However, this explanation is purely speculative, and it is unclear how it may account for the additional finding that the difference between production and perception is selectively driven by the string-copy condition.

None of our experiments identified a significant difference between the string-copy subconditions (CV and CVC). Yet the two perception experiments (3 and 5) followed a consistent trend: participants in the CVC condition had numerically higher accuracy than participants in the CV condition (Figure 8, left panel). Moreover, the effect size for this comparison was moderate in both of these experiments – suggesting that the difference between the two subconditions was not detected due to insufficient power. Indeed, a logistic regression over the data of the string-copy subconditions from these two experiments taken together does show significantly higher accuracy in the CVC condition, compared to the CV condition (β = 0.77, OR = 2.16, SE = 0.34, χ² = 5.00, p < .05). In the two production experiments, no similar trend is apparent, and the corresponding analysis for Experiments 4 and 6 (Figure 8, right panel) shows no significant difference between the two string copy conditions (β = 0.75, OR = 2.12, SE = 0.47, χ² = 2.68, p = .102).

Furthermore, recall the CSBI analyses in these experiments: participants tended to either systematically produce a CV or a CVC reduplicated form in the syllable-copy condition, rather than an even mix of the two. However, there is no evidence that these participants were more biased to produce CVC than CV: when combining the data from the syllable-copy conditions of Experiments 4 and 6, the median CSBI is exactly 0 (i.e., neither in the direction of a CV copy bias nor a CVC copy bias). The reasons for this perception-selective pattern are unclear: speculative explanations will be considered in Section 9.