3.1 The corpus

To investigate tonal adaptation in established English loanwords in Mandarin, I carried out a new corpus study. The loanwords were drawn mostly from Dong's (Reference Dong2012) dissertation on segmental adaptation. Assembled from four dictionaries, Dong's corpus consisted of about 1200 borrowings, including place names and first names. I supplemented this corpus with words from Wu (Reference Wu2006) and with entries labelled ‘loanword’ in the online Chinese dictionary MDBG (2014).

I excluded about 65 loanwords whose adaptations I presumed were influenced by semantics (e.g. 蹦极 bèngjí ‘bungee (jumping)’, from 蹦 bèng ‘jump’ + 极 jí ‘extreme’), since the choice of a particular character with an appropriate meaning entails the assignment of a particular tone (see Miao Reference Miao2005 for further discussion). I also excluded 60 syllables subject to third tone sandhi (Chen Reference Chen2000), because it is impossible to tell whether it is their surface rising tone or their underlying low tone that reflects how tones are assigned to loanwords (e.g. the first syllable in fǎlǎo ‘pharaoh’). Additionally, I excluded about 200 tokens of the syllables a, fo, hei, le, miu, ri, se, sen, te and teng, because, due to tonotactic gaps, they can only bear one tone in Mandarin. Since the tones on these syllables are fixed, it makes little sense to include them in discussions of tonal assignment.

Another issue complicating the study of Mandarin tonal adaptation is indirect borrowing. Li (Reference Li2017) and Wang (Reference Wang2020) point out that many English loanwords, especially those borrowed before the 1950s, entered Mandarin via Cantonese or Shanghainese, which have different tonal systems (as well as phonemic inventories and phonotactics). Such loanwords do not bear on Mandarin adaptation patterns, because their pronunciations are simply the Mandarin readings of the written forms of the original Cantonese or Shanghainese adaptations, which would exhibit the segmental and tonal adaptation patterns specific to those languages. To reduce this source of noise as much as possible, I excluded about 25 loanwords that were likely borrowed into Mandarin through Cantonese (cf. Dong Reference Dong2012). Some appeared in the Cantonese loanword adaptation literature, while I suspected that others had arrived through Cantonese because of the Cantonese form's greater segmental resemblance to the English form (e.g. 威化 wēihuà ‘wafer’; Cantonese: [wɐi HH faː MM]). I further excluded 啤(酒) pí(jiŭ) ‘beer’, which Wang (Reference Wang2020: 22) identifies as a ‘notable Shanghainese loanword’.

Finally, I excluded about 750 epenthetic syllables. Epenthetic syllables (e.g. the second syllable in āsīpǐlín ‘aspirin’) are inserted in adapted forms to preserve consonants from English clusters and codas without violating Mandarin phonotactics. Since epenthetic syllables do not correspond to English syllable nuclei, I do not consider them in the analysis.

After all these exclusions, the corpus consisted of 2644 syllables, from 1287 loanwords.

3.2 Standard syllables in the corpus

Past studies have observed that loanwords in Mandarin show the influence of convention and standardisation (Mar & Park Reference Mar and Park2012, Zheng & Durvasula Reference Zheng, Durvasula, Tao, Lee, Su, Tsurumi, Wang and Yang2016, Chang Reference Chang2020, Wang Reference Wang2020). Chinese institutions like the Xinhua News Agency issue official adaptations or dictate standards for adapting foreign words, including proper names, which contributed 72% of the syllables in my corpus (Dong Reference Dong2012). Since the use of a standard character entails the assignment of a particular tone, and since the considerations that went into choosing these characters are obscure, it is necessary to control for standard characters in studies of tonal adaptation. In this section, I examine the extent to which standard characters and syllables are present in my corpus.

To do this, I consulted the English-to-Mandarin transcription chart in Names of the world's peoples (Xinhua News Agency 1993). The chart's columns correspond to onset consonants and its rows to vowel nuclei or syllable rhymes. (The chart does not take stress into account, except insofar as English vowel quality is confounded with stress.) Each cell gives the character prescribed for the adaptation of one or more English syllables (e.g. 巴 bā for /bɑː/, /bæ/ or /bʌ/).

I first established what English string each syllable corresponded to (e.g. pǐ in āsīpǐlín adapts the string /pə/ from aspirin /ˈæ.spə.ɹɪn/). I then coded each syllable for whether it used the standard character for the English string it adapted. In this first pass, I erred on the side of allowing standard character use to account for as much of the corpus as possible. While the chart maps English pronunciations to Mandarin syllables, English orthography often appeared to influence which cell a standard character was chosen from. For example, while the standard character for /bə/ is 伯 bó, the string /bə/ in baroque /bəˈɹɒk/ is adapted with 巴 bā, the standard character for /bɑː/. It seems likely that this is because the syllable /bə/ is spelled <ba>, so I coded this syllable, and others like it, as using the standard character. Additionally, if an English string was adapted with the standard character for a different but phonetically very similar string (e.g. /ən/ for /ɪn/), I also coded the loanword syllable as using the standard character.

Applying this coding scheme, I found that 1852, or about 70%, of the syllables in the corpus used standard characters. An additional 99 syllables used a character homophonous with the standard character for the English string they adapted (e.g. the second syllable in toffee /ˈtɒ.fi/ is adapted with 妃 fēi, homophonous with the standard character 菲 fēi). Another 362 syllables were adapted with some standard syllable from the chart, though not the one prescribed for the given English string. For instance, the second syllable in the adaptation of Martina /ˌmɑː.ˈtiː.nə/ is 丁 dīng, rather than the 廷 tíng prescribed for English /tin/. However, dīng is a standard syllable; it is used to adapt English /din/. In the end, only 331 syllables (13%) did not match some standard syllable, showing how pervasive the influence of standardisation was in the corpus.

The use of standard syllables, regardless of Chinese characters or English strings, presents a problem for uncovering patterns of tonal adaptation in the corpus. In loanwords like toffee, with, as noted above, 妃 fēi instead of the standard 菲 fēi, adapters may have been aware of the standard loanword syllable fēi, and used a different character with the same pronunciation, either due to ignorance of which character was officially prescribed or for other reasons.Footnote ⁷ The adaptation has nevertheless been shaped by the standard syllable fēi. In loanwords like Martina, even though adapters used a different segmental syllable than is standard for the given English string, they may still have been guided by their knowledge of the repertoire of standard syllables for loanwords. That is, even if they chose dīng instead of the expected tíng for Martina, this does not mean the adaptation is free of the influence of standardisation: they may have opted for dīng because it is also a standard syllable. Since the focus of this study is tonal adaptation, I do not seek to explain the segmental adaptation of the loanwords. If segmental adaptation happens first (Chang Reference Chang2020), then any time the subsequent step of tonal assignment produces a standard syllable, there is no reason to conclude that the choice of tone indicates anything beyond adapters’ awareness of these standard syllables. All standard syllables in the corpus therefore potentially have tones that principally reflect official standards, and reveal little about possible phonological determinants of tone.

The pervasiveness of standard syllables in the loanword corpus constitutes a major confound that threatens to distort any analysis of tonal adaptation patterns. It has also likely contributed to the difficulty that past studies have had in identifying anything more than general trends in the data and in accounting for all the variation (cf. Wang Reference Wang2020). One way to avoid this confound would be to analyse only the 331 non-standard syllables. Excluding this much data drastically reduces the size of the corpus, though, and having fewer data points would make it harder to test for a wide range of potential phonological influences on tonal assignment.

Another possible pitfall of excluding all the standard syllables is that, despite undoubtedly reflecting official standards, they may still, in the aggregate, reflect some phonological patterning as well. For instance, suppose there were a tendency for adaptations of English stressed syllables to receive high tone. One standard syllable with high tone is dīng. The segmental syllable ding appears in the corpus as the adaptation of a number of English strings. It could be the case that, beyond ding's ‘baseline’ tendency to have high tone because the standard syllable is dīng, it is even more likely to have high tone when the English syllable it corresponds to is stressed. If standard syllables are excluded entirely, patterning like this would be inaccessible to the analysis. Phonological determinants of tone might go undetected, or their strength could be over- or underestimated. Ultimately, I kept the standard syllables in the corpus and carried out an analysis using maximum entropy Harmonic Grammar that controlled for lexical effects and the use of standard syllables.

3.3 MaxEnt analysis of the corpus

Maximum entropy Harmonic Grammar (MaxEnt; Goldwater & Johnson Reference Goldwater, Johnson, Spenader, Eriksson and Dahl2003, Hayes & Wilson Reference Hayes and Wilson2008) is a probabilistic constraint-based framework of phonology that can handle variation in the output for a given input. A MaxEnt grammar has weighted constraints, and candidates for each input receive harmony scores, which are the sums of their weighted constraint violations. The harmony scores are used to generate probabilities of occurrence for each candidate.Footnote ⁸ The most harmonic candidate (i.e. the one with the lowest harmony score) will be the most frequent output, the next most harmonic candidate will be the next most frequent output, and so on. A successful MaxEnt model generates candidate probabilities that closely resemble candidates’ observed frequencies. Given a set of inputs, candidates, candidate frequencies, constraints and constraint violations, a learning algorithm can determine the constraint weights that maximise the probability of the observed outputs (i.e. that yield the best match between the candidate frequencies predicted by the MaxEnt grammar and the actual candidate frequencies). Importantly, the MaxEnt models I present are intended not as grammatical models of Mandarin tonal adaptation, but as statistical analyses that allow us to control for lexical tone frequencies and the use of standard syllables, and to test for phonological effects in tonal assignment. Potential phonological determinants of tone are implemented as constraints in a MaxEnt grammar, and the importance accorded to these constraints by the learning algorithm reveals how much influence various phonological factors exert.

I carried out the MaxEnt analysis in three steps. The first step established a model that predicted the tones of loanword syllables based solely on tone distributions in the lexicon. The second step added constraints that captured the additional effects of standard syllables on tonal assignment. In the third step, I examined the role of phonological factors involved in tonal adaptation in the corpus. This three-step analysis represents the most conservative approach to identifying phonological effects in loanword tones. By first modelling the roles of the lexicon and of standard syllables, I attribute as much of the variation in tonal assignment as possible to these factors, so that the third step only uncovers phonological effects that could not be otherwise explained by lexical statistics or reliance on standard syllables with potentially incidental tones.

I conducted the MaxEnt analysis on the full corpus of 2644 syllables. The input to the learning algorithm consisted of a large tableau composed of sub-tableaux, each of which corresponded to one Mandarin segmental syllable associated with various properties of the English input. The English inputs were not fully specified English syllables, but rather abstract units that preserved information about English stress, position and onset properties. I distinguished stressed and unstressed syllables, and final and non-final syllables.Footnote ⁹ There were seven possible onset types: voiceless aspirated stop/affricate, voiceless unaspirated stop/affricate, voiced stop/affricate, voiceless fricative, voiced fricative, sonorant and zero. Sub-tableaux were organised by Mandarin segmental syllable, however, because I assume, following Chang (Reference Chang2020), that tonal adaptation follows segmental adaptation.Footnote ¹⁰ The same English syllable can be adapted with different Mandarin segmental syllables: for example, final unstressed /kəʊ/ is rendered as kě in rococo, but as kòu in one version of cocoa. I am interested not in how /kəʊ/ came to be adapted as ke or kou, but in how kě came to bear low tone and kòu falling tone. Thus, despite corresponding to the same English syllable, kě and kòu are data points in two different sub-tableaux, one for ke and one for kou.

Additionally, adaptations of different English syllables may be data points in the same sub-tableau if the English syllables are identical after abstraction and adapted with the same Mandarin segmental syllable. For example, non-final stressed syllables (or strings) such as /ˈdæ/ in Daphne /ˈdæf.ni/, /ˈdɑː/ in Darwin /ˈdɑː.wɪn/ and /ˈdʌ/ in dozen /ˈdʌ.zμ/, all of which are adapted with Mandarin da, do not differ on any phonological factor that I consider. They are therefore all included in a single da sub-tableau. This sample sub-tableau, representing 13 tokens from the corpus, is given in (3). The input is the Mandarin segmental syllable da, with associated information about the English form (i.e. /ˈdæ/, /ˈdɑː/, /ˈdʌ/, etc.), while the candidates are da, with each of the four possible lexical tones. The observed counts are taken from the corpus. Other elements of the sub-tableau are explained below.

1. (3)

   

Finally, there can be multiple sub-tableaux for the same Mandarin segmental syllable. For instance, while the tokens of da in (3) adapt non-final stressed syllables with voiced onsets, other tokens of da adapt non-final stressed syllables with voiceless onsets (e.g. the second syllable in Montana /ˌmɒn.ˈtæ.nə/). These tokens were entered into a separate da sub-tableau, to allow testing of the effects of distinct English onset properties on tonal assignment.

In the first step of the MaxEnt analysis, I established a model that predicted what tones the syllables in the corpus would have if they were projected directly from the lexical tone frequencies for each segmental syllable. I used Chang's (Reference Chang2020) tone probabilities, which were calculated from Da's (Reference Da, Zhang, Xie and Xu2004) character frequencies. The lexical tone probabilities for da are shown in (3). The most probable tone for da is falling tone: based on all the characters with the pronunciation da and their frequencies in Da's corpus of Modern Chinese texts, the probability that a given token of da will have falling tone is 0.653.

To establish the baseline lexical model, I created four constraints for each segmental syllable, each of which penalised the occurrence of that syllable with one of the four tones. Each candidate for a given syllable violated the constraint that penalised its tone. For instance, in (3), the candidate dā incurs one violation of *da1. I used the MaxEnt Grammar Tool (Hayes et al. Reference Hayes, Wilson and George2009) to fit the constraint weights, not to the observed tone counts in the corpus but to the lexical tone probabilities. In the case of da, the constraint *da1 should have the highest weight, because dā has the lowest lexical probability, and therefore needs to be suppressed the most by the constraint that penalises it. I used positive weights for the lexical constraints, so as the weight of *da1 increases, the harmony score (i.e. badness) of dā increases, thereby reducing its predicted probability. The fitted weights of the lexical constraints for da are given in (3), where *da1 has the highest weight.

Since there was a constraint for every possible syllable–tone combination, the model was able to achieve essentially a perfect fit to the native tone frequencies: in (3), the predicted probabilities of the candidates exactly match da's lexical tone probabilities. I converted the model's predicted probabilities to predicted counts for the tones in the corpus. For example, the lexical model predicts that eight of the 13 da tokens in this sub-tableau should bear falling tone. In fact, none of them did.

The square of the correlation (r ²) between the observed corpus counts and the predicted counts of the baseline lexical model was 0.585. This goodness-of-fit measure is a way of gauging how much of the variation in tonal assignment the model can explain. Lexical tone probabilities evidently influence what tones loanword syllables receive, but they cannot fully account for tonal assignment in the corpus.

In the second step of the MaxEnt analysis, I added standard syllable constraints to the model. For instance, for the segmental syllable da there exists the standard syllable dá, so I added a constraint da2, which, as I will explain, rewards the candidate dá for matching a standard syllable. In some cases, there was no standard syllable for a segmental syllable in the corpus (e.g. zhi), and in other cases there were multiple standard syllables (e.g. di has both dí, for /diː/ and /dɪ/, and dì, for /tiː/ and /tɪ/), giving rise to multiple constraints.

The standard syllable constraints were added to the baseline lexical model, whose already fitted lexical constraint weights were held constant. Thus the learning algorithm could assign weight to the standard syllable constraints in order to achieve a better fit to the observed tones in the corpus. In other words, adding the standard syllable constraints and determining their weights allowed me to quantify the role of standard syllables in causing assigned tones to deviate from the lexical tone distributions. (4) shows how the sub-tableau in (3) is updated in the standard syllables model with the addition of the standard syllable constraint da2. The candidate that ‘violates’ da2 is dá; however, I fitted the standard syllable constraints with negative weights, so that a ‘violation’ of a standard syllable constraint reduces (i.e. improves) the violating candidate's harmony score, thereby increasing its predicted probability. I used negative weights so that a constraint like da2, which specifies that a da token should have rising tone, could simply reward dá by reducing its harmony score, rather than having to penalise dā, dǎ and dà by increasing their harmony scores by the same amount. Negative weights allow for a very straightforward interpretation.

1. (4)

   

In the standard syllables model, the lexical constraints retained their previously fitted weights. They were active in the model insofar as their weights contributed to the candidates’ harmony scores, but their weights were fixed, modelling the influence of the lexicon. I used Excel Solver to fit negative weights just to the standard syllable constraints, in such a way that the probability of the observed tones from the corpus was maximised. The predicted probabilities of the standard syllables model were generated from the weights of both the lexical constraints (the role of the lexicon) and the standard syllable constraints (the role of standardisation). (4) shows the new predicted probabilities and counts for da. Based on all the da sub-tableaux in the model, da2 received a relatively large negative weight, increasing the predicted incidence of dá in this sub-tableau from two tokens in the lexical model to eleven in the standard syllables model. This is closer to the actual number of observed dá tokens, so the influence of standardisation on tonal assignment appears to be strong for da.

The fit between the observed and predicted counts increased greatly in the standard syllables model (r ² = 0.952, up from 0.585 in the lexical model). Factoring in the effect of standard syllables on top of the tone distributions projected from the lexicon accounts for many more of the assigned tones in the corpus. In fact, it appears that lexical tone probabilities and the use of standard syllables explain nearly all tonal adaptation in established loanwords.

In the third step of the MaxEnt analysis, I investigated what further role, if any, phonological factors played in tonal assignment in the corpus. I retained the lexical and standard syllable constraints, with their fixed weights, from the previous model, and added a host of phonological constraints. By fitting weights to these constraints, I could gauge the presence and strength of phonological effects in the corpus.

I devised a fairly exhaustive set of potential phonological effects, based on findings in the literature. For each phonological property, there were four constraints, one for the association of that property with each tone. Just as a standard syllable constraint like da2 specifies that tokens of da should bear rising tone, a phonological constraint like Stressed1 specifies that loanword syllables that adapt English stressed syllables should bear high tone. Since past studies reported stress-to-tone effects, I included the constraints Stressed1–4 and Untressed1–4. I also included Final1–4 and Non-final1–4, which applied to loanword syllables adapting word-final and non-word-final English syllables, because of Wu's (Reference Wu2006) finding that monosyllables (and thus perhaps all stressed final syllables) were adapted with falling tone. Moreover, I included constraints that could capture interactions of stress and position: StressedNon-final1–4, StressedFinal1–4, UntressedNon-final1–4 and UnstressedFinal1–4.

I included constraints for onset properties covering English sonority, English voicing and English and Mandarin aspiration. Obs1–4 and Son1–4 applied to loanword syllables whose onsets corresponded to English obstruents and sonorants respectively. Voiced1–4 and Voiceless1–4 applied analogously, with respect to the phonemic voicing of the English consonant. Asp1–4 and Unasp1–4 applied to loanword syllables whose onsets correspond to English aspirated and unaspirated stops and affricates respectively, while MandarinAsp1–4 and MandarinUnasp1–4 applied to loanword syllables with aspirated and unaspirated stop and affricate onsets in Mandarin. To account for the possibility that English voiced obstruents might have a phonological effect distinct from that of all obstruents or of all voiced consonants, including sonorants, I also included the constraints VoicedObs1–4. Wu (Reference Wu2006) observes that loanword syllables with aspirated onsets adapting English voiced stops have rising tone. I therefore included the constraints Voiced & MandarinAsp1–4 to check for this interaction of English voicing and Mandarin aspiration. Finally, to test for an across-the-board preference for a particular tone in loanword syllables independent of lexical and standard syllable influences (for example, one could imagine high tone serving as a default tone in loanwords), I added the constraints Tone1–4, each of which simply specifies that loanword syllables should have that tone.

Like the weights fitted to the standard syllable constraints, the weights fitted to the phonological constraints were negative. Thus, for example, a constraint like Stressed1 rewarded a candidate like dā in (4) for being the adaptation of an English stressed syllable and having high tone. The larger the negative weight of Stressed1, the more the harmony score of dā would decrease, resulting in a higher predicted probability of dā. If the predicted probability of dā did not need to increase to better match the tones observed in the corpus, then Stressed1 would not accrue a large negative weight. In this way, the fitted weights of the phonological constraints, as well as the degree to which they contribute significantly to the MaxEnt model, reveal which phonological properties are most influential in tonal assignment.

I again used Excel Solver to determine the weights of the 76 phonological constraints that maximised the probability of the actual tones in the corpus. In this phonological model, the predicted tone probabilities are generated from the lexical, standard syllable and phonological constraints. After finding the weights of the phonological constraints, I removed 15 constraints with weights of zero. Next, I pruned the phonological constraint set by successively removing the constraint that contributed least to the model, as determined by a likelihood ratio test, until removing a single constraint made the model's fit significantly worse at the 0.05 level. The final pruned model contained the 23 phonological constraints in Table I.

Table I Phonological constraints in the corpus MaxEnt analysis.

In the pruned phonological model, the fit between the observed and predicted tone counts improved slightly (r ² = 0.967 vs. 0.952 in the standard syllables model). A likelihood ratio test showed that the addition of the 23 phonological constraints significantly improved the model's performance in predicting the observed tones, justifying their inclusion (χ ²(23) = 199.517, p < 0.001). Thus, although the fit between observed and predicted tones increased only modestly from the standard syllables model, there were real phonological effects in the corpus.

Table I lists the 23 phonological constraints in the final MaxEnt model and their weights in descending order of Δ log likelihood (i.e. the decrease in log likelihood if the given constraint is removed from the model; the larger its Δ log likelihood, the more a constraint is contributing to the model). Note that constraint weights and Δ log likelihoods do not go hand in hand: a constraint may have a relatively large weight but a relatively small Δ log likelihood, meaning its contribution to the model is relatively small. This can happen if a constraint's effect is limited to only a few sub-tableaux in the model. For each constraint, Table I also gives the rate at which the relevant tone was assigned to loanword syllables with the relevant phonological property by the standard syllables model (before the addition of phonological constraints), by the present phonological model and in the corpus. For instance, in the case of VoicedObs2, the standard syllables model predicts that 35.7% of loanword syllables whose onsets correspond to English voiced obstruents will have rising tone; the addition of the phonological constraints increases that rate to 44.4%. The observed rate is 44.9%.

The top three constraints by Δ log likelihood are VoicedObs2, Voiced2 and Voiceless1, all of which involve the voicing and/or sonority of the English consonant to which a loanword syllable's onset corresponds. VoicedObs2 has a Δ log likelihood that is nearly an order of magnitude larger than any other phonological constraint, as well as by far the largest negative weight. This means that the strongest phonological effect in the corpus was an increased likelihood of rising tone on loanword syllables whose onsets corresponded to English voiced obstruents, even when lexical tone frequencies and the use of standard syllables were controlled for. The Δ log likelihood of the second most important constraint, Voiced2, also sets it apart from the other 21 phonological constraints, since it is twice that of the next constraint. The strength of Voiced2 reveals a more general association between English voicedness, across obstruents and sonorants, and rising tone assignment. The third most important constraint, Voiceless1, reflects the flip side of the voicing effect: loanword syllables whose onsets correspond to English voiceless obstruents are more likely to receive high tone.

What about the phonological constraints with progressively smaller Δ log likelihoods? The next two constraints, MandarinUnasp1 and Unasp4, are not far behind Voiceless1 in contributing to the model. The MaxEnt model is a very complex system, in which many intersecting constraints interact to adjust the predicted probabilities of the candidates in each sub-tableau. Consequently, a relatively important constraint does not necessarily reflect a true phonological effect. A heuristic means of checking whether a constraint is likely to represent a real phonological effect is to inspect the improvement in predicted tonal assignment from the standard syllables model to the phonological model. VoicedObs2, Voiced2 and Voiceless1 all push up the assignment rate of the given tone for the relevant syllables, and achieve a closer fit to the observed data, suggesting they represent genuine phonological effects. MandarinUnasp1 and Unasp4 are different. Despite the relatively large Δ log likelihoods (and negative weights) of these constraints, the phonological model predicts that numerically fewer syllables with the relevant onset type will bear the relevant tone than the standard syllables model does. While these reductions do bring the predicted rates closer to the observed rates, MandarinUnasp1 and Unasp4 do not seem to represent straightforward phonological effects whereby, for example, loanword syllables with unaspirated onsets in Mandarin are more likely to bear high tone.

Finally, the stress- and position-related constraints Stressed1, StressedNon-final1 and Final4 remain in the model after pruning, and appear to improve predicted rates of tonal assignment, but they have smaller weights and much smaller Δ log likelihoods than the voicing-related constraints.

4.2 Standard syllables in the experiment

As in the loanword corpus, there was evidence for the use of standard syllables in the experiment. I first identified oral responses that matched the pronunciation of the character prescribed for the given English stimuli syllable in the Xinhua News Agency's English-to-Mandarin transcription chart. For stressed syllables, I looked up the syllable in the chart. If an oral response matched the pronunciation of the character prescribed for that syllable, I coded it as using the standard syllable (e.g. 巴 bā is prescribed for /bɑː/, so any oral response bā for the stimulus [ˈbɑ] was considered standard). For unstressed syllables, I categorised more possible characters as standard, to account for the possibility of orthographic influence or misperception. While subjects never saw written forms of the stimuli, their knowledge of English may have caused them to make inferences about how the words they were hearing were spelled. Additionally, unstressed [ɪ] and [ə] are relatively confusable, and voiceless unaspirated onsets may have been misperceived as voiced. Consequently, I looked up characters prescribed for possible spellings or misperceptions of stimuli syllables, and coded oral responses as standard if they used the pronunciations of any of these characters. The stimulus syllable [ʃəɹ] does not have an adaptation in the chart, because [əɹ] is not one of the listed rhymes. Subjects usually adapted this syllable as shi or she. The chart includes standard characters with the pronunciations shí, shě and shè, so I coded such oral responses for [ʃəɹ] as standard.

In all, I identified 827 syllables (37% of oral responses) as corresponding to standard adaptations of the English stimuli syllables. Another 230 adapted syllables corresponded to some standard syllable in the Xinhua News Agency chart, even if it was not the standard syllable for the given English stimulus. The remaining 1197 syllables (53%) did not belong to the set of standard syllables.

While standard syllables were less ubiquitous in the experiment than in the corpus, they were nevertheless prevalent. Moreover, there may be a greater need to control for standard syllables in the experiment, because the range of English syllables was so narrow, increasing the likelihood of spurious phonological effects. For example, ignoring vowel reduction, there were only four English syllables with voiceless stop or affricate onsets ([ʧʰɑ], [pʰi], [ʧʰi] and [pʰɪn]), and the standard adaptations of all four bear rising tone (chá, pí, qí and píng). In the following section, I first examine the non-standard syllables, in order to gain an initial idea of what phonological effects are at play in the experiment. I then include the standard syllables in the MaxEnt analysis, for the same reason as in the corpus analysis: while they reflect official standards, they may also participate in phonological patterning, and this information would be lost if they were excluded.

4.3 Initial results and MaxEnt analysis of the experiment

I first take a general look at tonal assignment in the 1197 non-standard syllables in the experiment. Figure 1 shows the proportion of syllables adapted with each of the five tones by stress and position in English. The graph is inevitably noisy, since there is no accounting for the lexical tone probabilities of the individual segmental syllables. Nevertheless, the figure suggests certain patterns. Stressed initial syllables overwhelmingly prefer high tone (0.71 vs. <0.13 for any other tone), while stressed final syllables strongly prefer falling tone (0.47 vs. <0.20 for any other tone). This looks like stress-to-tone: high tone captures the H* pitch accent aligned with an initial stressed syllable, while falling tone captures the H* L-L% contour realised on a final stressed syllable. Thus it appears plausible that English intonation, here operationalised as stress and position, influenced tonal adaptation in the experiment.

Figure 1 Tonal assignment by stress and position in non-standard syllables, averaged across participants.

The MaxEnt analysis of the experiment was identical to that of the corpus, except for minor differences related to the design of the experiment. From the 2254 oral responses, including standard syllables, I excluded the 72 syllables produced with neutral tone. Based on Fig. 1, neutral tone is used almost exclusively for unstressed final syllables. Adaptation with neutral tone is appropriate for unstressed final syllables, because Mandarin syllables with neutral tone are always unstressed and word-final (Duanmu Reference Duanmu2007). It seems unlikely that any additional phonological factors (e.g. segmental properties) would favour or disfavour adaptation with neutral tone, since it is not a lexical tone with a fixed pitch level or contour, but rather varies in realisation, depending on the preceding tone. Since the corpus data consisted of written Mandarin forms, it contained no identifiable instances of neutral tone, so excluding neutral tone syllables in the experiment makes the two studies more comparable. I therefore conducted the analysis on 2182 oral responses. The English inputs were again abstract units distinguishing stress (stressed vs. unstressed), position (initial vs. final) and four onset types: sonorant, voiced stop/affricate, voiceless stop/affricate and voiceless fricative. Sub-tableaux were again organised by Mandarin segmental syllable.

In the first step of the analysis, I established the lexical model, which predicted what tones subjects’ oral responses would have had if they had assigned tones based purely on the lexical tone probabilities of the segmental syllables they had chosen. The fit (r ²) between the observed counts in the study and the counts predicted by the lexical model was 0.574. Just as native tone frequencies could not fully account for tonal adaptation in the corpus, so they cannot fully explain which tones subjects used in the experiment.

In the second step of the analysis, I added standard syllable constraints to the MaxEnt model and determined their weights. In the standard syllables model, the fit between the observed and predicted counts improved (r ² = 0.722). This shows that the tones of standard syllables did influence subjects’ tonal adaptations, but unlike in the corpus, it is not the case that lexical effects and standardisation can explain nearly all tonal assignment.

The third step of the analysis investigated the additional role of phonological determinants of tone. I used the same 76 phonological constraints as in the corpus MaxEnt analysis (replacing Non-final with Initial wherever it occurred, since all non-final syllables were initial in the experiment). After determining their weights, I removed 13 constraints that were assigned weights of zero. I then pruned the phonological model, using the same procedure as for the corpus. The final pruned model contained 28 phonological constraints. In this model, the fit between the observed and predicted tone counts improved further (r ² = 0.837), and the 28 phonological constraints collectively contributed significantly to the model (χ ²(28) = 572.592, p < 0.001). Table III lists the 28 phonological constraints with their weights, in descending order of Δ log likelihood.

Table III Phonological constraints in the experiment MaxEnt analysis.

Ignoring Unasp2, which I discuss below, the top three constraints are Initial1, StressedInitial1 and StressedFinal4, suggesting that English stress and position were the strongest phonological determinants of tone. Specifically, the initial syllables in English disyllables were overall more likely to receive high tone. On top of this effect, there was a particular preference for adapting stressed initial syllables with high tone. Finally, stressed final syllables were more likely to be adapted with falling tone. Relative to the standard syllables model, the addition of the 28 phonological constraints does in fact increase the predicted counts of initial syllables with high tone, stressed initial syllables with high tone and stressed final syllables with falling tone, bringing these counts closer to the observed counts. Two other relatively important constraints, Unstressed3 and Stressed1, also behave straightforwardly in the model: relative to the standard syllables model, the phonological model predicts more unstressed syllables to have low tone and more stressed syllables to have high tone, in both cases achieving a closer match to the observed tones.

Among the top tiers of phonological constraints, there are also several that refer to onset properties. The importance of Unasp4 seems to lie in complex and opaque interactions with the rest of the model (cf. MandarinUnasp1 in the corpus MaxEnt analysis), because the predicted rate of falling tone assignment in adapted syllables whose onsets correspond to English unaspirated onsets is not greater in the phonological model than in the standard syllables model. Rather, it is in fact smaller, coming closer to the observed rate. Thus Unasp4 does not correspond to a clear phonological effect in the way that the stress and position constraints above do.

The case of Unasp2 is different. The addition of the phonological constraints slightly increases the predicted rate of rising tone assignment in syllables whose onsets adapt unaspirated stops and affricates, and this increase leads to a better fit to the observed rate. In examining the model more closely, though, I discovered that what seemed to be driving the large negative weight and large Δ log likelihood of Unasp2 were standard syllables with rising tone whose associated standard syllable constraints had received quite low or zero weights in the standard syllables model. Additionally, many of these syllables’ onsets corresponded to voiceless unaspirated stops and affricates, which only occurred in unstressed final position. To give the most dramatic example, which is representative of the wider phenomenon, the standard syllables model predicted that only 19 tokens of pi adapting English (unstressed final) [pi] would receive rising tone, while the phonological model, with Unasp2, predicted 29 rising tone tokens. The actual number was 30. Although pí is a standard syllable, the constraint Pi2 had the 39th smallest weight out of the 45 standard syllable constraints that received any weight. The incidence of pí in stressed and/or initial syllables, where stronger stress-to-tone effects may have reduced standard syllable use, may not have justified assigning Pi2 a larger weight. Consequently, the standard syllables model underestimated standard pí specifically in unstressed final syllables (the only English syllables with voiceless unaspirated onsets), which led to Unasp2 having to pick up the slack in the phonological model. In short, I contend that the importance of Unasp2 does not represent a genuine phonological effect, but is rather an artefact of how English aspiration is confounded with stress and position, the fact that standard syllables may have been used more where stress-to-tone effects were weaker, and the tones of the standard syllables for the particular stimuli in the experiment. There is evidence that the importance of VoicedObs1 and Obs1, two constraints with Δ log likelihoods similar to those of Unstressed3 and Stressed1, can be explained in the same way.

4.4 Comparing the corpus and experiment analyses

The phonological models of the corpus (Table I) and the experiment (Table III) reveal sharp differences in the phonological effects that are active in each set of adaptations. The three constraints that contributed most significantly to the corpus model, VoicedObs2, Voiced2 and Voiceless1, were not retained in the experiment model after pruning (i.e. they did not contribute significantly at the 0.05 level), indicating that these onset voicing effects played no meaningful role in participants’ online adaptations. Meanwhile, of the top three stress- and position-related constraints in the experiment model, Initial1, StressedInitial1 and StressedFinal4, only StressedInitial1 (= StressedNon-final1) remained in the corpus model after pruning, where its contribution was less significant (χ ²(1) = 5.830, p = 0.016 vs. χ ²(1) = 93.270, p < 0.001 in the experiment model). Similarly, Unstressed3 was not retained in the corpus model, while Stressed1's contribution was less significant than in the experiment model (χ ²(1) = 7.522, p = 0.006 vs. χ ²(1) = 50.146, p < 0.001). The stress-to-tone effects that prevailed in online adaptations (see below) were thus either absent or weaker (especially compared to other phonological factors) in integrated loanwords.

4.5 Discussion

As in established loanwords, lexical tone frequencies and the influence of standard syllables were major determinants of tone in the experiment, though their combined role was not as overwhelmingly dominant as in the corpus. In considering the phonological determinants of tone, I return to the hypotheses put forward at the beginning of §4. First, the voicing effect observed in the corpus did not emerge in the experiment. Adapted syllables whose onsets corresponded to English sonorants or voiced obstruents were not more likely to receive rising tone, and adapted syllables whose onsets corresponded to English voiceless obstruents were not more likely to receive high tone. The prediction that stressed syllables would be more likely to receive high tone was borne out, but the picture is more nuanced, with both stress and position implicated. Being initial in an English disyllable increased the likelihood of high tone assignment. Additionally, there were interactions of stress and position, such that high tone was more likely specifically in stressed initial syllables, and falling tone more likely specifically in stressed final syllables. Two lesser effects were an increased likelihood of low tone on all unstressed syllables and of high tone on all stressed syllables.

Together, these patterns suggest that English intonation was the principal phonological determinant of tone in online adaptations. This can be accounted for by the perceptual mapping model. In choosing tones, subjects sought to make their adaptations phonetically similar to the stimuli by reproducing the English stress-aligned pitch contours as best they could. The pitch accent on stressed initial syllables translated to high tone, while the pitch accent and subsequent fall on stressed final syllables translated to falling tone. The effects of Initial1, Unstressed3 and Stressed1 can be interpreted as grosser generalisations, perhaps reflecting a preference for simpler mappings. High tone could adapt the pitch accent of a stressed initial syllable or the unspecified pitch of a pretonic unstressed initial syllable. Low tone could adapt the relatively low pitch of unstressed syllables, whether they are initial, and therefore less prominent than the following stressed syllable, or final, realising the L-L% of the English intonation pattern. High tone could also be used generally for all stressed syllables, which bear the pitch accent, even if stressed final syllables also contain a pitch fall. There is almost no evidence that subtler pitch perturbations following different types of onsets shaped subjects’ tonal assignments. Instead, it was the overall intonation of the English forms that mattered.

In the corpus MaxEnt analysis, adding the phonological constraints increased the fit (r ²) between predicted and observed tones only from 0.952 to 0.967, but in the experiment analysis, the fit increased from 0.722 to 0.837. This suggests that phonological effects played a bigger role in the online adaptations than in established loanwords. At the same time, the final model accounted for more tonal assignment in the corpus than in the experiment, where more variation remains unexplained. As in the corpus, some of the remaining variation may be due to subjects avoiding characters (even though they did not have to give responses in Chinese characters) or avoiding the production of real Mandarin words. Other possible sources of variation in the experiment include subjects’ individual tonal adaptation preferences and a potential self-priming effect whereby subjects tend to adapt the next word with the same tones used on the previous word. Because the input data to the MaxEnt analysis were not coded for subject or trial number, any such patterning would have been inaccessible to the model, impeding its ability to explain all tonal assignment.