Glossa Psycholinguistics

1.1 Priming of Mandarin tonal variants

Most morphemes in Mandarin are one syllable long and have either a High, Rising, Low, or Falling tone. Syllables with Low tone (also called third tone), like qi^L “start”, are instead pronounced with something similar to a Rising tone (also called second tone) if they are followed by another Low tone within the same intonational unit, as in qi^R dian^L “start-point” (“starting point”);¹ this change is known as “third tone sandhi” (see, e.g., Zhang, 2010, among others). In other words, any Mandarin morpheme with underlying Low tone has multiple possible pronunciations – multiple allomorphs – depending on the phonological context, just like we saw above for English plural -s. Likewise, a surface Rising tone might be a realization of either an underlying Low tone or an underlying Rising tone.

The crucial question for us is whether hearing one Low- or Rising-tone syllable in isolation causes listeners to activate both Low and Rising underlying forms. One could imagine at least two possibilities. First, there might be no reason for a listener to activate multiple forms, because the input already unambiguously matches some canonical form in the lexicon,² and because there is no context to make the listener suspect that this surface form might have been changed from some other underlying form. On the other hand, listeners might activate multiple representations that the input corresponds to anyway; for example, if different allomorphs of each syllable are listed in the lexicon (as has been proposed by, e.g., Nixon et al., 2014, and Li & Chen, 2015) or if the lexicon itself is made up of detailed memories of many episodes in which each item was experienced (Tenpenny, 1995; Goldinger, 1998) and lexical entries are sound-meaning pairings reinforced by experience (Pulvermüller, 1999), then the recognition of one form of a morpheme may also cause activation of related forms of the same morpheme. If so, then a listener hearing a Rising tone might also activate Low-tone variants of the same syllable (because they have experience hearing syllables with the same meaning in that tone, or because they recognize that this Rising-tone input is sometimes a realization of the Low-tone form), and likewise a listener hearing a Low tone might also activate Rising-tone variants of the same syllable (because, again, they have experience hearing the same morpheme in both tones, or because they recognize that this Low-tone form can sometimes be realized with Rising tone).

Several recent studies have examined priming between Mandarin Low and Rising tones. Chien and colleagues (2016) examined the representation of Mandarin words including phonologically altered tones, like qi^R dian^L (underlyingly qi^L dian^L). They found that recognition of these words was primed when they were immediately preceded by presentation of the first morpheme in its underlying form (qi^L) but were not primed by presentation of its surface form (qi^R). This suggests that primes in isolation only activated phonologically matching lexical forms (i.e., Low tone only activated Low tone, and Rising tone only activated Rising tone), but then the presentation of the target which involved a surface Rising tone in a sandhi context involved recovery of the underlying Low tone. Thus, the Low tone recovered from the target was primed by the Low prime, and not by the Rising prime. It is, however, not clear that the recovery of the Low form from the surface Rising tone in the target is due to phonological knowledge. It is likely due instead to semantic or lexical knowledge, since the syllable with a surface Rising tone in the target is embedded within a compound word whose meaning is also related to the Low-tone prime and/or to many other words from the same cohort as the target (e.g., the target xuan^R ju^L ‘election’ is closely related to the prime xuan^L ‘choose’, the target fu^R dao^L ‘to coach/tutor’ is closely related to other words sharing the same initial morpheme such as fu^L zhu^F ‘to assist’, etc.), not to mention that the initial syllable of the target is written with a character whose canonical pronunciation is Low-tone. Therefore, the priming observed here may be mainly due to the activation of the meaning of a known compound word, rather than purely due to activation of phonological variants.

Meng and colleagues (2021, experiment 3a) used a similar manipulation as Chien and colleagues (2016) but with a cross-modal rather than purely auditory paradigm, and with a shorter interval between prime and target. They found that visually-presented two-character words in which the first syllable would undergo sandhi were primed by both Low and Rising primes. Crucially, Rising primes only facilitated targets that began with underlying Low tones when those tones were in contexts that would trigger them to be produced as Rising (e.g., kao^R primed kao^R gu^L “archaeology” – which is underlyingly kao^L gu^L – but not kao^L cha^R “investigate” [experiment 1]), whereas Low primes facilitated both types of targets. This shows that the priming of sandhi-undergoing words by both Low and Rising targets is not just a result of Low and Rising being processed the same way, but is a result of the relationship between these tones and tone sandhi. In short, for Chien et al. (2016) only Low primed sandhi targets, whereas for Meng et al. (2021) both Low and Rising did. The difference between these findings is likely due to how the stimuli were presented in these experiments; in Meng and colleagues’ study the participants could begin processing the target while the prime was likely still being processed (see also Nixon et al., 2014, on the role of stimulus-onset asynchrony in priming Mandarin tone sandhi forms). In any case, the facilitation of sandhi-undergoing targets by Low-tone primes in this study is subject to the same caveat as described above for Chien and colleagues (2016), and the facilitation by Rising-tone primes is not necessarily evidence for activation across different variants since both the prime and target in this condition share the same surface form.

A recent study (Politzer-Ahles et al., 2022) failed to observe the sort of between-allomorph facilitation seen in the other studies described above. In an immediate priming paradigm, Mandarin-speaking participants heard a target syllable shortly after seeing either a Chinese character whose canonical pronunciation matches it or one whose pronunciation mismatches it in some way. Syllables whose tones mismatched the canonical pronunciation of the preceding character triggered some recognition difficulty (as evidenced by more negative event-related brain potentials) relative to syllables whose tones matched the preceding characters. Crucially, this effect was not modulated by the presence of a systematic phonological relationship between the heard target and the preceding character. Hearing a Rising tone when a Low tone was expected should not have caused much recognition difficulty because a Rising-tone form is actually a possible realization of an underlyingly Low-tone morpheme; however, this situation yielded just as much of a mismatch effect as hearing a Rising tone when some other wholly unrelated tone was expected. This happened regardless of whether the critical syllable was in a context that licensed the tone change. In other words, hearing an allomorph of the expected syllable was no different than hearing a phonologically unrelated syllable. This finding seems inconsistent with the others described above, as it suggests that hearing one form of a morpheme does not necessarily activate other phonological variants of this morpheme, at least when sentential context and most lexical context is removed. In that study, however, participants’ task was to judge whether the target matched the prime, which might not tap into the same mechanisms as normal word recognition (as opposed to the other studies reviewed above, which used a lexical decision task).

So, is it the case that exposure to one form of a Mandarin morpheme activates representations of its phonologically-conditioned allomorphs? While several previous results suggest that this can happen, some may not be due to the activation of phonologically-conditioned allomorphs per se, but to other factors; furthermore, one set of recent results suggests that exposure to one form of a morpheme does not always activate its allomorphs, at least not when this allomorphy is based on a suprasegmental phonological change.

Because the abovementioned studies have all used a paradigm in which the target immediately follows the prime, it is difficult to disentangle contributions of form and meaning from contributions of morphology, as well as the role of response strategies (which might not necessarily reflect natural processing) on the part of participants. In the present study we use long-lag priming, which is arguably a more suitable paradigm for examining activation of morphemes while ruling out effects of form and meaning. The goal of the study was to use long-lag priming to examine whether facilitation occurs across different phonological forms of the same morpheme – i.e., whether hearing one form of a morpheme activates its other allomorphs. Below we review the relevant properties of long-lag priming, and then introduce the details of the present study.

1.2 Overview of long-lag priming

In long-lag priming (also known as long-term priming or delayed priming), other trials intervene between a participant’s exposure to a prime and their response to a target. In other words, while a traditional paired priming study will have participants hear, e.g., hunter and then hunt and immediately respond to hunt, a long-lag priming study will have a participant hear some long series of words (and sometimes pseudowords) such as hunter, washing, blosh, table, thurl, drem, hunt…, while responding to each word. In some (particularly older) long-lag priming studies, participants explicitly “study” a prime, and later perform a task (typically some kind of metalinguistic judgment or verbal naming response) on the target. In others, participants perform the same task on both the prime and on the target, with extra trials intervening between them, as in the example above (see, e.g., Coane & Balota, 2010, and Horner & Henson, 2011, for discussion of how the tasks performed on primes and on targets may influence the priming effect). Under these conditions, participants are generally not aware that the experiment involves pairs of related stimuli – unlike in overt paired priming, where the presence of related vs. unrelated pairs is usually quite obvious and where participants may adjust their processing strategy accordingly. Furthermore, while targets that share a morpheme with (or are identical to) an earlier prime are facilitated in this paradigm, targets that share only a meaning or form relationship with an earlier prime are usually not facilitated (see, e.g., Feldman, 2000, for review); this makes the paradigm ideal for examining how different allomorphs can activate underlying morphemes while removing potential confounds due to semantic relationships or superficial form relationships.

Another paradigm with similar benefits as long-lag priming is subliminal masked priming, in which a prime immediately precedes a target but is presented for a short duration sandwiched between masks that make it unlikely for participants to be consciously aware of the prime. It also reduces the role of strategic factors, as participants are not aware of prime-target relationships (since they’re not aware of the primes at all), and it also preserves morphological and identity priming while rarely showing meaning or form priming. (In spite of these similarities, it should not be assumed that masked priming and long-lag priming measure the same process; see, e.g., Bowers, 2000, for arguments that they reflect different mechanisms.) However, it was not ideal for the present study because the present study focuses on variants of Mandarin morphemes, which necessitates presentation of spoken stimuli (as phonological alternations in Mandarin are not represented in the standard written form). While subliminal masked priming with spoken stimuli is possible (e.g., Kouider & Dupoux, 2005; Ussishkin et al., 2015), it requires substantially compressing the duration of the primes. Since duration is an important secondary cue for tone identity in Mandarin, compression may influence participants’ recognition of the primes, which could substantially complicate the present study and the interpretation of its results. Using long-lag priming avoids this complication.

To set the stage for the manipulation that will be used in the present study, below we summarize some key properties of long-lag priming that are relevant for comparing phonologically alternated forms of Mandarin syllables.

1.3 The present study

The present study examines whether hearing one form of a Mandarin morpheme activates phonologically-derived variants of the same morpheme. As described above, a Mandarin syllable that has Low tone underlyingly is instead pronounced with something like a Rising tone, in certain phonological contexts. Thus we wanted to see whether hearing shi^L (for example) activates shi^R, and vice versa. Unlike all the previous studies on priming of phonologically-conditioned tonal variants, we use long-lag priming, which allows us to rule out potential contributions of meaning and form in order to identify priming that is purely due to repeated access of the same lexical representation. Note that, even though the change between Low and Rising tone only occurs in phonological contexts that license it, we tested single-syllable stimuli in isolation in order to avoid the complication of having potential priming from both syllables of compound words, not to mention the additional complication that it would not be possible to include Low tones in the same context as Rising tones (since Low tones would not survive in a context that triggers tone sandhi); the consequences of the use of single-syllable stimuli will be revisited in Section 5. Below we summarize the specifics of the three experiments in this study, and preview their findings.

In the first experiment, we compare the priming yielded by phonological variants of the same morpheme (e.g., shi^L and shi^R, which are both possible pronunciations of a morpheme whose underlying form is shi^L) to priming yielded by form-related primes with no morphological relationship to the target. For example, shi^H shares the same segments with shi^L but has a different tone; thus, these words differ by one phoneme. Crucially, there is no systematic phonological relationship between Low and High tone in Mandarin (i.e., there is no phonological alternation that changes Low to High or vice versa), so shi^H is not a possible pronunciation of shi^L, nor vice versa. These two stimuli would thus be phonetically, but not morphologically, related. As discussed above, previous long-lag priming experiments in other languages suggest that there should not be priming for form-only overlap; however, we do not yet know if that is the case for tones, as all previous long-lag priming studies on form have used segmental or orthographic manipulations. Thus, in order to conclude that we are observing morphological priming, we should find facilitation of shi^L by the prime shi^R, or vice versa, but no facilitation by a morphologically unrelated prime like shi^H.

This first experiment surprisingly failed to find priming in either condition. Thus, in order to make sure that long-lag priming works at all for Mandarin monosyllables, we follow it up with a second experiment including exact repetition priming (e.g., shi^L primed by shi^L), where facilitation should be expected under any account. In this experiment we successfully observe repetition priming – ruling out the possibility that this paradigm just doesn’t work with Mandarin, or that some problem with our experiment design prevented us from being able to detect priming effects – but still no morphological priming.

Finally, in a third experiment we test whether the repetition priming was due entirely to episodic priming, with no contribution from lexical access to the same lexical entry. We do this by testing “repetition” priming with the same word spoken in different voices, following Kouider and Dupoux (2009). Surprisingly, however, while repetition priming still holds (and with a similar magnitude as it did with identical voices in Experiment 2), morphological priming is still absent. Taken together, the results suggest that lexical repetition priming – beyond mere episodic priming – does occur in Mandarin monosyllables, but that there is no morphological priming of monosyllable allomorphs removed from any larger context. These results raise questions about what the underlying mechanisms for morphological priming are.

Materials (including recorded stimuli and experiment control scripts), data, and analysis scripts for all experiments are available at https://osf.io/pt8qv.

2.2 Results

As shown in Figure 1, accuracy among most participants was high, although three participants had very low (near chance) accuracy. As we did not pre-register any exclusions based on participant-wise accuracy, and the sample is large enough that the results are unlikely to be skewed by a few participants’ performance, we did not exclude these participants.

Figure 1: Beeswarm plot (i.e., a univariate scatterplot with the dots shifted horizontally to avoid overlapping, with the added effect that the plot roughly simulates a violin plot) showing the distribution of accuracy scores across participants. Each dot represents the accuracy (across all trials, including primes, fillers and pseudowords) for one participant. The vertical axis represents participant accuracy, ranging from 0% to 100%. The majority of dots are clustered high on the graph, between 80% and 100%, with only a few participants below this range (particularly, three participants with accuracies near or below 60% stand out from the rest of the distribution).

Figure 2 shows the priming effects to the critical targets. The left side shows the priming effects for the morphologically-related condition, and the right side the form-related condition. Priming effect size (unrelated minus morphologically-related, or unrelated minus form-related) is represented on the vertical axis, such that facilitative priming effects should be above zero. Each dot represents one participant’s or item’s priming effect in the given condition, with a line connecting a given participant’s/item’s dot in the morphological condition to the corresponding participant’s/item’s dot in the form condition. These dots show the general distribution of the priming effects in each condition; in fact, for each condition, the dots are roughly evenly scattered above and below zero (rather than mostly above or mostly below zero), suggesting that there is not a robust priming effect for either condition. The lines show the general distribution of the relationship between the conditions; since there is a roughly even mixture of upward- and downward-sloping lines (rather than most of the lines sloping upward or most of the lines sloping downward), this suggests that there is not a robust difference between the morphological and form priming effects. For ease of interpretation, gray bars are also shown underneath the clouds of dots; the heights of the bars indicate the mean priming effect for each condition. Each bar is below zero; that is to say, both morphologically-related and form-related targets elicited slower response times than totally unrelated targets. This suggests that, contrary to our hypothesis, morphological relatedness did not elicit a facilitative priming effect. Since neither condition elicited priming (and if anything, the morphological primes facilitated responses less than the form primes did), the confound with phonetic similarity (i.e., the fact that the morphologically-related primes are more phonetically similar to their targets than the form-related primes are) is moot.

Figure 2: Priming effects in Experiment 1. See main text for details.

While the visualization in Figure 2 strongly suggests that the results do not support our hypothesis, we also ran a mixed-effects model on the reaction time data to account for the influence of nuisance covariates that may not be reflected in this visualization. The results of the model (see 2.1.4 for details of the model specification) were consistent with the description above, and are summarized below.

The nuisance covariates yielded reasonable coefficients in the model, which can also function as a quality check to ensure the data are behaving as auditory lexical decision data ought to (note that the parameter estimates below have been transformed from z-scored units back into original units): the lag between prime and target was not significantly associated with reaction time (b = –0.41, t = –0.21), but longer stimuli were associated with substantially longer reaction times (b = 56.60, t = 13.82), and later trials in the experiment were responded to more quickly (b = –15.09, t = –1.87).

Most importantly, contrary to our hypothesis, the form-related condition was responded to 13ms more quickly than the morphologically-related condition (b = –13.35, t = –2.75), and the unrelated condition 21ms more quickly than the morphologically-related condition (b = –20.74, t = –4.22). As we only had predictions about facilitative priming for the morphologically-related condition, and the effect observed here is in the opposite direction, we do not make any claims about statistical significance of this unexpected inhibition pattern; rather, we test it in a confirmatory replication in Experiment 2.

The other effect of interest is the interaction between prime type and target tone. The interaction was barely significant in an omnibus test using model comparison (χ²(2) = 6.40, p = .041). While the effect of prime type, as discussed above, was opposite the expected direction, this pattern was stronger in Rising-tone targets than in Low-tone targets. Specifically, for Rising targets, the form-related condition was 21ms faster than the morphological condition (b = –20.51, t = –2.96) and the unrelated condition 33 ms faster (b = –32.82, t = –4.74); for Low targets, though, the form-related condition was only 6ms faster than morphological (b = –6.19, t = –0.91) and the unrelated condition only 9ms faster (b = –8.66, t = –1.27).

2.3 Discussion

Contrary to our expectation, morphologically-related prime-target monosyllable pairs in Mandarin did not trigger facilitative priming; in fact, morphologically-related targets were responded to more slowly than unrelated targets. This is different than what has been observed for long-lag morphological priming in other experiments.

It would be premature, however, to conclude from this experiment that morphological relationships based on Mandarin tone work differently than morphological differences based on segmental alternations in other languages. Instead, it is possible that the long-lag priming paradigm simply does not work with Mandarin monosyllables, for whatever reason – for instance, maybe the experiment only works with multisyllabic stimuli like those used by Kouider and Dupoux (2009), or maybe there are some other characteristics of Mandarin that we had not considered but which cause this paradigm to not work. Therefore, to test whether the paradigm works at all, it was necessary to run the experiment with an identity priming condition as a manipulation check (e.g., shi^L primed by shi^L, the exact same stimulus). If the long-lag priming paradigm works with Mandarin monosyllables, this condition should elicit faster reaction times than the unrelated condition, regardless of the psychological status of Mandarin tone alternation. On the other hand, if our failure to observe morphological priming in the above experiment was because long-lag priming just doesn’t work in Mandarin, then this identity priming condition should also not elicit faster reaction times than the unrelated condition.

The second reason for the new experiment was to test the unexpected inhibition effect. As we had predicted a facilitative effect, we are not in a position to make conclusions about the reality of the inhibition for morphologically-related pairs observed in Experiment 1; that is to say, it is possible that this pattern was just noise and that claiming there is inhibition would be a Type I error. We can only say that we did not find evidence for facilitative priming, not that we did find strong evidence for inhibition. Therefore, in Experiment 2 we again included the comparison between morphologically-related and completely unrelated, in order to test this pattern in a confirmatory way now that it could be explicitly predicted.

3.2 Results

As in Experiment 1, accuracy was fairly high for most participants, as shown in Figure 3.

Figure 3: Beeswarm plot showing the distribution of accuracy scores across participants. As in Experiment 1, the majority of participants are between 80% and 100%, with only a few participants below this range.

The critical priming effects are shown in Figure 4, using the same type of graph used for the previous experiment. In this case the left side shows the priming effects for the identity condition, and the right side the morphologically-related condition. In this case we can see that for the identity condition, most of the dots are above zero, suggesting that there is a fairly reliable priming effect (i.e., that unrelated is slower than identity for most participants and items); on the other hand, for the morphological condition, the dots are roughly evenly distributed above and below zero, as in Experiment 1, suggesting that there is little priming effect. Furthermore, the majority of the lines appear to be sloping downward, suggesting that the identity priming effect is bigger than the morphological priming effect for most participants and items. Finally, the bars show that the identity priming effect is fairly large (about 60 ms), whereas the morphological effect is a small inhibition effect.

Figure 4: Priming effects in Experiment 2. See main text for details.

The statistical model provided further support for this pattern. Targets preceded by identical primes were responded to 64ms faster than targets preceded by unrelated primes (b = –63.90, t = –13.51), whereas targets preceded by morphological primes were responded to 10 ms more slowly than targets preceded by unrelated primes (b = 10.32, t = 1.88); the latter effect is significant in a one-tailed test, based on our a priori expectation that there would be inhibition like there was in Experiment 1.

On the other hand, it is worth mentioning that the nuisance covariates did not show the same effects as observed in Experiment 1. Longer stimulus duration was associated with a small and non-significant increase in reaction time (b = 9.08, t = 0.63), while later trials were associated with a non-significant decrease in reaction time (b = –13.05, t = –0.92). Finally, targets with more lag between themselves and their primes were associated with marginally slower reaction times (b = 3.37, t = 1.78). These results are roughly opposite those of Experiment 1, where lag was not a significant predictor of reaction time but stimulus duration and trial number were.

The interaction between priming condition and target tone was again significant (χ²(2) = 7.40, p = .025). Specifically, while the size of the identity priming effect did not significantly differ between Low- and Rising-tone targets (b = –6.42, t = –0.67), the size of the morphological priming effect was 30 ms smaller with Low targets than it was with Rising targets (b = –29.90, t = –2.72). While Rising-tone targets were significantly inhibited by Low-tone morphological primes (25ms slower than unrelated: b = 25.27, t = 3.22), Low-tone targets were not significantly inhibited by Rising-tone morphological primes, and in fact were numerically, but not significantly, faster than Low-tone targets with unrelated primes (b = –4.63, t = –0.60). This mirrors the pattern observed in Experiment 1, in which Rising targets were also inhibited by morphological primes but Low targets were not.

3.3 Discussion

This experiment demonstrated that the long-lag priming paradigm does indeed work with Mandarin monosyllables: identity priming, the most basic manipulation check, does occur (targets preceded by identical primes are responded to faster than targets preceded by unrelated primes). Thus, the failure to observe morphological priming in Experiment 1 was not because the paradigm overall does not work; it must be due to other reasons.

This experiment also confirmed the unexpected morphological inhibition pattern found in Experiment 1; while this pattern was an unpredicted surprise in that experiment, here we observed it when it had been explicitly predicted. Together, these experiments suggest that this particular morphological relationship causes inhibition, rather than facilitative priming, in Mandarin long-lag priming with auditory lexical decision.

This pattern of results leaves us with a conundrum. In some models of morphological representation, a morphological relationship should be the same as an identity relationship, as both involve accessing the same lexical representation twice (Stockall & Marantz, 2006; Kouider and Dupoux, 2009). Why, then, do we not observe morphological priming when identity priming is possible? One potential explanation, of course, is that this assumption about morphological representation is wrong (i.e., morphological priming is not the same thing as identity priming). Another, however, is that the priming observed here was not linguistic priming at all (i.e., not based on activation of a lexical or phonological representation), either identity or morphological priming, but rather was episodic priming: recognition of the same physical event that happened recently. This recognition could occur without accessing linguistic representations at all. Kouider and Dupoux (2009) demonstrated such a pattern in French. With relatively short lags between prime and target, and when identical primes and targets were the same physical stimuli, identity priming was larger than morphological priming; on the other hand, with a longer lag between prime and target, and when the “identical” prime and target were the same word spoken by different speakers (i.e., different physical stimuli), then identity priming was reduced to be the same size as morphological priming. Kouider and Dupoux argue that this occurred because the priming effect observed in the former situation (short lag, physically identical stimuli) was a sum of two distinct priming effects: repeated activation of a lexical entry or phonological representation (linguistic priming), and recognition of a familiar event (episodic priming). In the latter condition (long lag, physically different stimuli), the contribution of episodic priming was removed, leaving only the linguistic/morphological contribution, which was of the same magnitude as it was in the identity priming condition. Under this understanding of contributions to long-lag priming, it is possible that our Experiment 2 still did not demonstrate that long-lag linguistic priming works in Mandarin monosyllables. Rather, it is possible that the identity priming effect we observed was wholly episodic in nature, rather than linguistic. In short, maybe there really is no linguistic long-lag priming effect for Mandarin monosyllables.

To test this, we performed a third experiment attempting to remove the contribution of episodic memory.

4.2 Results

Accuracy was fairly high again, as shown in Figure 5. Two participants had accuracy near zero, which is likely due to these participants mixing up which key to press for words and which to press for pseudowords; therefore, we switched the correct/incorrect labels in these participants’ data for the analysis (recall that trials with incorrect responses are removed in our analysis).

Figure 5: Beeswarm plot showing the distribution of accuracy scores across participants.

The critical priming effects are shown in Figure 6, using the same type of graph used for the previous experiments. The results from this experiment look almost identical to those of Experiment 2: there is large priming for the identity condition and small inhibition for the morphologically-related condition, and the difference between these patterns is fairly robust across participants and items. In other words, using different voices for the primes and targets had no appreciable impact on the results; the present experiment using different voices showed the same pattern as Experiment 2 using identical voices across primes and targets.

Figure 6: Priming effects in Experiment 3. See main text for details.

The statistical model confirmed what can be seen in the figure. There was a significant 59ms identity priming effect (b = –59.08, t = –13.15). Morphological inhibition, while not significant this time, was in the same direction as the previous experiments, with targets preceded by morphological primes responded to 7ms more slowly than targets preceded by unrelated primes (b = 7.26, t = 1.40).

The nuisance covariates followed the pattern of Experiment 1 and not that of Experiment 2: later trials were significantly faster (b = –19.01, t = –2.09), longer stimuli elicited longer reaction times (b = 35.32, t = 7.71), and prime-target lag did not substantially impact reaction times (b = –0.87, t = –0.47).

Unlike in Experiments 1 and 2, the interaction between priming condition and target tone was not significant this time (χ²(2) = 2.96, p = .227). The size of the identity priming effect did not significantly differ between Low and Rising targets (b = –7.11, t = –0.79). The size of the morphological priming effect was 17ms smaller with Low targets than it was with Rising targets (b = –17.73, t = –1.72), but this is less of a difference than what was observed in the previous experiments. Specifically, Rising-tone targets were significantly inhibited by Low-tone morphological primes (16ms slower than unrelated: b = 16.12, t = 2.18), and Low-tone targets were not significantly inhibited by Rising-tone morphological primes (b = –1.61, t = –0.22). While this pattern did not yield a significant interaction, it is in the same direction as the pattern observed in Experiments 1 and 2.

4.3 Discussion

We again failed to find morphological priming, replicating the findings of Experiments 1 and 2. On the other hand, identity priming persisted, even when the primes and targets were presented in different voices. This suggests that the identity priming effect was not merely an episodic memory effect, but was a genuine lexical effect based on repeated access of the same lexical or phonological representation. The results, therefore, suggest that participants hearing, e.g., shi^L activate its lexical entry (as evidenced by the identity priming for physically different tokens of shi^L-shi^L), but do not activate a lexical entry associated with its phonological variant shi^R (as evidenced by the lack of morphological priming for shi^L-shi^R).

An alternative explanation could be that our identity priming effect was indeed episodic and the lag between prime and target was simply not enough to wipe out episodic priming. 18 to 52 trials intervened between prime and target in this experiment; by comparison, in Kouider and Dupoux’s (2009) experiment which successfully eliminated episodic priming, there were 96–192 intervening trials. We are not aware of any study testing different-voice identity and morphological priming with a lag of about 18–52 trials, so we cannot be certain that this lag should be enough to eliminate episodic priming. Importantly, however, Kouider and Dupoux also found that same-voice identity priming diminished with increasing lags (82 ms at short lags, 62 ms at medium lags, and 47 ms at long lags with different voices), whereas morphological priming remained constant across lags and voices. In the present experiment, different-voice identity priming (60 ms) was similar in magnitude to same-voice identity priming in the previous experiment (64 ms). Our lag of 18–52 is also fairly typical, if not longer than average (albeit admittedly shorter than that of Kouider and Dupoux’s different-voice experiment), as long-lag priming experiments go; for example, in a study by Wilder and colleagues (2019), a lag of just five intervening trials was enough to eliminate most of the difference between identity and morphological priming. Finally, given that the prime and target are physically different stimuli, it is unclear how the priming effect observed here could be episodic without some sort of abstraction – indeed, episodic priming should be ruled out of the present experiment by design. For these reasons, we find it more likely that the identity priming in the present experiment reflects genuine lexical access, rather than episodic memory.

5.1 Comparison to previous studies of priming across allophonic variants

While we have described the present study’s failure to observe morphological priming as surprising, it is arguably not unprecedented. In particular, the results of the present study are somewhat similar those of at least two previous studies.

The first is that of McLennan and colleagues (2003), who examined priming by allophonic variants based on American English flapping. In this variety of English, intervocalic alveolar stops preceding unstressed vowels, such as the /d/ in raider or the /t/ in rater, are often pronounced as alveolar flaps [ɾ]. McLennan and colleagues found that, under limited circumstances, tokens pronounced with the flap (i.e. [ɹeɪɾɚ]) fail to prime hyperarticulated tokens pronounced with full stop consonants [ɹeɪdɚ] or [ɹeɪtɚ], and vice versa. In both this study and our study, we see that one form of a word does not prime another form of the same word. Beyond this similarity, though, there are several important differences between these studies.

Most importantly, the conditions in which we did not find priming across different forms were precisely the conditions in which McLennan and colleagues predict it would be found. Specifically, they did not find across-form priming when participants’ task in response to primes and targets was “easy lexical decision” (lexical decision in which the nonwords were not very word-like), whereas they did find across-form priming when participants performed shadowing or “hard lexical decision” (lexical decision in which the nonwords were more word-like) for the primes and/or the targets. They explain this pattern by proposing that abstract effects (i.e., priming across different forms of the same word) take time to emerge. Therefore, faster and easier lexical decision tasks tap into form-specific encoding that occurs before abstract underlying forms have been accessed (or, more specifically, before those underlying forms have established resonances with other possible surface forms of the same word), whereas more abstract encoding is tapped into by slower and harder lexical decision tasks, and by speech shadowing, which they argue requires deeper phonological processing than easy lexical decision does. If our study’s lexical decision task was easy enough to allow fast processing that taps into more specific surface forms as opposed to abstract underlying forms, then we should have also found little to no priming when the same word was presented in different voices; to the contrary, though, we found substantial (and not substantially reduced) priming in that situation (see also Luce & Lyons, 1998, who also find this pattern). In fact, as our nonword stimuli included a mix of more and less word-like stimuli (but even the least word-like stimuli were still phonological neighbours of real words), our task was arguably a “harder” lexical decision task and thus would be expected to have yielded abstract priming across different allophonic forms under their account.

The other situation in which McLennan and colleagues (2003) failed to find priming across different forms was when the same word was pronounced in different ways that do not trigger allophonic variation or ambiguity. Pronouncing raider casually as [ɹeɪɾɚ] leads to ambiguity (the [ɾ] could correspond to either a /d/ or a /t/), but pronouncing, e.g., bacon casually does not lead to any ambiguity (a fast, casual pronunciation of the /k/ in bacon still does not correspond to any phoneme other than /k/). While McLennan and colleagues (2003) found that casual, flapped tokens of raider prime hyperarticulated tokens of raider and vice versa, they also found that casual tokens of bacon did not prime carefully produced tokens of bacon or vice versa (see also McLennan et al., 2005). They explained this as resulting from ambiguity: specifically, they argue that hearing an ambiguous token causes listeners to activate multiple possible underlying forms and to subsequently “restore” their possible surface forms, whereas hearing an unambiguous form just activates the corresponding underlying form and then resonance between that underlying form and the initially perceived surface form reinforces activation of that same surface form (along with all its specific acoustic characteristics), without any need to activate other forms. For that reason, ambiguous forms trigger abstract priming whereas unambiguous forms trigger form-specific priming. This cannot explain the present study’s pattern of results, however, because our primes and targets were highly ambiguous (see further discussion of this point below) and the surface tone itself is also ambiguous in half of them (as a surface Tone 2 can correspond to underlying Tone 2 or Tone 3) and yet did not trigger abstract priming across different allophonic forms of the same morpheme, or across different allotones of the same underlying Tone 3. The main difference between the types of ambiguity in these two studies is that in our study the ambiguous tones correspond to the canonical form of one representation (Tone 2 is the canonical form of underlying Tone 2 and is the sandhi form of underlying Tone 3), whereas in McLennan and colleagues’ (2003, 2005) studies the flap is arguably not the canonical form of either /d/ or /t/, under traditional phonological theory. Perhaps, then, these studies’ differing results could be reconciled by positing that the speech comprehension algorithm first compares input to its canonical underlying form, and successful matching with a canonical underlying form precludes any attempt to match with other underlying forms (see also Sumner & Samuel, 2005, who argue that the lexical decision task “bolsters a preference for the canonical form” [p. 333]) – if so, then between-form priming might occur for other forms of tone alternation that yield a surface form which does not closely match any underlying form. This possibility remains to be tested.

All in all, the fact that our priming effect was not specific to individual speakers (i.e., priming persisted across indexical variability) but was specific to tone categories (i.e., priming did not persist across allophonic variants of the same morpheme) is difficult to reconcile with McLennan and colleagues’ (2003) account; furthermore, the task which failed to elicit abstract priming in our experiments is arguably most similar to one of the tasks that did elicit abstract priming in theirs, and the sort of variation tested in our experiment is arguably more “abstract” and less “indexical” than the sort tested in theirs. Therefore, we cannot conclude that the lack of morphological priming observed in our study is demonstrating the same phenomenon, or has the same locus, as the lack of abstract priming they describe. For the same reason, we cannot assume that our lack of priming is predicted by the lack of priming in their study. Instead, we consider the lack of priming in our study to be unexpected based on existing theories of priming. The question that remains is why we failed to find priming under these circumstances.

Another study reporting results that appear similar to ours is one by Sumner and Samuel (2005), who found that word-final variants of /t/ in American English all appear to contact the same underlying representation in immediate priming (e.g., when there is no lag, flute primes music regardless of which allophone of /t/ is produced in flute) but not in long-lag priming. In long-lag priming, they argue that the carefully articulated canonical form of flute primes itself, but other forms do not prime themselves (e.g., flute with an unreleased [t̚ ] does not prime itself) and under no circumstances does one form prime another form. While our results are somewhat different, in that both variants prime themselves in our study (Rising-Rising identity priming was 59 ms and 53 ms faster than control in Experiments 2 and 3, respectively, and Low-Low identity priming was 60 and 62 ms faster than control), our results are similar to theirs in that we also failed to find priming across different variants at long lags. Sumner and Samuel’s (2005) long-lag paradigm, however, is not comparable to ours, because they did not include unrelated controls for their critical targets; while they showed that different variants prime each other less than identity priming (e.g., the reaction times for a canonical [t] word primed by an unreleased [t̚ ] word are slower than those for a canonical [t] word primed by another canonical [t] word), they did not show that these variants don’t prime each other at all (i.e., they did not compare [t̚ ]-[t] priming to unrelated priming within the same items; they did include a separate set of items with unrelated control primes, but this makes the crucial comparison between related and unrelated primes a between-items comparison). Therefore, it is difficult to be sure that this study did not observe across-variant priming whatsoever; if it did, however, that would be consistent with our own finding.

On the other hand, in a more recent study, Sumner (2013) found evidence that different variants are activated in long-lag priming. Specifically, careful and casual articulations of center, i.e. with or without a pronounced [t], both make it hard for participants to later reject a pseudoword senner. If the reason this occurs is because both forms activate center, then these results would be evidence for priming across different variants, contra the present study and the other studies discussed above that failed to find such priming.

5.2 Possible explanations for the lack of morphological priming

5.3 A directional asymmetry between Low and Rising tones

Not only did morphologically-related primes unexpectedly inhibit activation of targets across all three experiments, but this pattern was driven by the Rising targets (preceded by Low primes) in all three experiments. Specifically, in Experiments 1, 2, and 3, Rising targets preceded by Low primes were responded to 33, 30, and 16 ms more slowly than Rising targets preceded by unrelated primes. On the other hand, Low targets preceded by Rising primes were responded to 9, –5, and –2 ms more slowly than the same targets preceded by unrelated primes in Experiments 1, 2, and 3, respectively. (Note that the manipulation of the target’s tone was within items in Experiment 1, but between items in Experiments 2 and 3.)

Identity priming did not show an asymmetry like this; the identity priming effects in Experiments 2 and 3 were 59 and 56 ms for Rising targets, and 60 and 63 ms for Low targets. The directional asymmetry thus seems to be limited to how Low and Rising forms activate their variant forms, not their canonical forms. There was, however, a comparable asymmetry in form-only conditions: across experiments, Rising targets were inhibited by 13 ms if they were preceded by High- or Falling-tone primes (compared to completely unrelated primes), whereas Low tones were primed/inhibited by 0 ms when preceded by High- or Falling-tone primes. Unfortunately our critical stimuli did not include form-related priming conditions allowing us to compare the effects of Low and Rising primes,⁵ which means that, while we can claim that Low and Rising targets were subject to different amounts of form priming, we cannot be sure whether Low and Rising primes caused different amounts of form priming.

We did not expect such asymmetries, and are not aware of previous priming studies on allophonic variants reporting comparable asymmetries. There is, however, some precedent for Low-Rising tone asymmetries outside of the priming literature, specifically in electrophysiological studies examining habituation effects in an oddball paradigm. Li and Chen (2015) and Politzer-Ahles and colleagues (2016) both found that an unexpected Low tone after a series of Rising tones is treated as a bigger mismatch than an unexpected Rising tone after a series of Low tones. We are not yet aware of any satisfactory explanation for this pattern. Li and Chen (2015) argue that this pattern happens because Low-tone morphemes have multiple stored variants (one which is Low and one which is similar to a Rising tone), and hearing Low tone activates both, so that the unexpected Rising tone later does not strongly mismatch with the Rising-like variant that has already been activated. This can successfully explain their own results but does not on its own explain the results of Politzer-Ahles et al. (2016), who found that series of Low tones also were not strongly mismatched by other unexpected tones either. Politzer-Ahles et al. (2016) argued that Low tone may involve a phonologically underspecified feature, but this is just a claim they put forth to account for their results and there is no independent phonological evidence for it (see also Meng et al., 2021, for criticism of the argument that Low tone is underspecified). Meng et al. (2021) suggest that this asymmetry may happen because a series of multiple Low tones is phonologically illegal, but this is probably not the case for the habituation paradigms in these studies, where the stimuli are presented with pauses in between (400 ms in Li and Chen, 2015, and 500 ms in Politzer-Ahles et al., 2016), which means that they are presumably not in the same intonational phrase and thus not in a context that would license tone sandhi. Given that there is not yet any well-supported explanation for what mechanism causes these asymmetries in electrophysiological studies, it is also not yet clear whether the asymmetry in the current study is caused by the same mechanism or something else.

Might ambiguity play some role in this asymmetrical effect? Rising-tone syllables are potentially ambiguous, because they could be variants of underlyingly Rising or underlyingly Low tones. On the other hand, Low-tone syllables are unambiguous, because only underlyingly Low syllables can surface as Low. Crucially, Low (unambiguous) primes pretty reliably inhibited Rising targets in our experiments, whereas Rising (ambiguous) primes had much less inhibition on Low targets.

It would be tempting to explain this pattern by positing that primes activate all underlying forms that they might be surface forms of, and cause inhibition for any similar forms which are not activated (or, as discussed in the previous section, interfere with the lexical decision task carried out on similar forms). Under this explanation, Low primes activate Low representations and suppress Rising representations; Rising primes, on the other hand, activate both. This could explain why we observed inhibition for Rising targets after Low primes (which, under this account, do not activate Rising variants) but not for Low targets after Rising primes (which, under this account, do activate both Low and Rising variants). There are, however, several problems with this account. Most seriously, the difference between our identity priming results and morphological priming results is not predicted by this account: Rising primes facilitated Rising targets in identity priming, but they didn’t facilitate Low targets in morphological priming, they merely failed to inhibit them. Some additional mechanism would be needed to explain why Rising primes activate Low representations enough to keep there from being inhibition in the Rising-Low priming condition, but not enough for there to actually be facilitation. Secondly, this explanation relies on the assumption that form-only relationships cause inhibition in long-lag priming, which has generally not been observed in extant research.⁶ Finally, this explanation also cannot explain why a similar asymmetry emerged for Low and Rising targets preceded by purely form-related High and Falling primes in Experiment 1, where Rising targets had a 12 ms form inhibition effect whereas Low targets only had a 3 ms form inhibition effect. That suggests that the explanation for the asymmetry should be based on the processing of the target, rather than the processing of the prime, since the targets are what is in common across all these comparisons.

Overall, while this Low-Rising asymmetry is tantalizing and its robustness across experiments suggests that there is something important causing it, we cannot yet offer any satisfactory explanation for it and can only conclude that future research is needed to investigate what causes this asymmetry.