2.1. Loanword phonology in New Zealand English

Loanword phonology in New Zealand English (NZE) is chosen as a test case to address the above research question. This is because NZE speakers very frequently use loanwords borrowed from te reo Māori, which is the language spoken by Polynesian people who arrived to New Zealand about 1,000 years ago. NZE has borrowed a great number of words from te reo Māori, since the 19th Century when a large number of European people started immigrating to New Zealand. Many place names in New Zealand are borrowed from te reo Māori, and many public signs are written in te reo Māori as well as English. Some common nouns such as kumara ‘sweet potato’ and whānau ‘family’ are commonly used in NZE speech. Therefore, New Zealanders are exposed to Māori loanwords every day, and they know a large number of Māori words (see Macalister, 2005; Calude, Miller, & Pagal, 2017), although most of them are not able to speak te reo Māori (Te Puni Kōkiri, 2006).

The two languages, te reo Māori and NZE, have different phonological systems. For example, the Māori phonological system includes five short vowels /i, e, a, ɔ, u/, five long vowels /i:, e:, a:, ɔ:, u:/, and ten consonants /p, t, k, f, h, m, n, ŋ, ɾ, w/, whereas the NZE phonological system includes a larger set of phonemes as in other varieties of English (see Biggs, 1961; Bauer & Warren, 2008). Hence, when words are borrowed from te reo Māori to NZE, the words may include non-native structure. For example, the tap sound [ɾ] can be regarded as non-native to NZE, because it is Māori’s rhotic phoneme, whereas NZE has a coronal approximant [ɹ] as its rhotic phoneme. In NZE speech, this non-native rhotic sound is sometimes adapted to a native rhotic sound (e.g., ko[ɹ]u and ma[ɹ]ae) and sometimes imported without modification (e.g., ko[ɾ]u and ma[ɾ]ae) (see Maclagan & King, 2005). For simplicity, the current study focuses on the non-native rhotic sound /ɾ/, in order to discuss the likelihood of choosing adapted structure and imported structure. The reason for selecting this variable under study is three-fold. First, the distinction of imported [ɾ] and adapted [ɹ] is acoustically clear, with the detail in Section 3.5. This allows us to secure the objectivity of classification. Second, this variation seems salient even for NZE speakers without linguistic knowledge. The pronunciation of /r/ sounds in Māori loanwords is sometimes discussed in daily conversation in New Zealand. Third, there is a large enough number of loanwords with /r/ that are frequently used in daily life (e.g., koru, kumara, and marae).

(2)	Variable under study in this paper

As the current study focuses on this variable, the term ‘imported structure’ refers to tap sounds [ɾ], and ‘adapted structure’ refers to approximants [ɹ] throughout this article.

2.2. Exemplar Theory

In order to deduce the prediction related to the research question in (1), we employ the usage-based theory, more specifically, Exemplar Theory (Pierrehumbert, 2001, 2002; Foulkes & Docherty, 2006). This section reviews some theoretical hypotheses about representations in the mind of a speaker, the activation of a category, and the selection of an exemplar for production.

2.2.1. Representation of exemplars and categories

Exemplar Theory assumes that linguistic knowledge is built up by representing in memory previously encountered speech with detailed phonetic information. These episodic memories are called exemplars. Exemplars with similar perceptual values cluster together, with the result categories (e.g., lexical categories, phonological categories, and social categories) formed in the cognitive system. It is assumed that exemplars and categories are connected in a complex manner. An exemplar can be associated with several categories simultaneously (see Docherty & Foulkes, 2014). For example, when we hear [pɛŋgwən] produced by a girl, the perceived token may be stored as an exemplar associated with a lexical category ‘penguin’ and social categories ‘female’ and ‘young’ simultaneously. A category can be associated with another category. Fiasson (2015) argues that allophonic categories, such as a released variant [t] and a fricated variant [F] of a coronal plosive phoneme /t/, are cognitively linked, on the basis of his finding that shadowing a short medial fricated variant [F] causes a medial released variant [t] to be shorter (see also Nielsen, 2011). In this way, a large cloud of memories is complexly represented in the cognitive system of a speaker.

In the case of loanword phonology in NZE, we specifically hypothesize that (a) a NZE speaker stores exemplars including adapted structure [ɹ] and imported structure [ɾ] in her mind, and the phonological categories, ‘adapted structure [ɹ]’ and ‘imported structure [ɾ],’ are formed in her mind. This is because NZE speakers are exposed to the two structures in daily speech. Next, we hypothesize that (b) the imported structure [ɾ] is closely socially associated with the social category ‘Māori,’ because imported structure is identical to the structure used in the Māori language and society and promoted in Māori community. These two theoretical hypotheses are illustrated in Figure 1. The letters inside the cloud represent individual exemplars, and the labels represent categories formed by exemplars with similar phonetic values. Note that each exemplar may also be associated with a lexical category such as koru and marae depending on which loanword the /r/ sound is produced in, and a social category such as ‘female’ and ‘young’ depending on who produced the /r/ sound.

Figure 1

Category and exemplar association.

2.3. Predictions

The aim of this section is to review previous literature related to the four sociolinguistic effects in the research question in (1), and deduce the prediction on the basis of the hypotheses presented in the preceding section.

3.1. Participants and procedure

Thirty-two NZE speakers took part in a set of two experiments, and filled in a questionnaire. The participants were recruited through advertisements on the Facebook page of the University of Canterbury Students’ Association. By specifying in the advertisements, it was ensured that all the participants are non-bilingual Pākehā New Zealand English speakers aged between 18–35. ‘Pākehā’ is an ethnicity term that usually refers to New Zealanders who are descendants of European immigrants. The reason why only Pākehā speakers were examined is that there is a different sociolinguistic variety associated with Pākehā speakers and Māori speakers (King, 1993; Szakay, 2007, among others), and Māori speakers may be exposed to te reo Māori in a different way from Pākehā speakers. Focusing on Pākehā speakers allows us to control the effect of these sociolinguistic differences. Note that three of the participants identify themselves as Māori as well as Pākehā, because their distant relatives such as great grandparents were Māori. The age requirement of 18–35 allows us to control the effect of a diachronic change. The participants were tested individually by the author in a sound booth at the University of Canterbury, and received a 15 NZD voucher at the end of the experiments.

Experiment 1 was designed to explore the topic effects on the likelihood of selecting adapted structure and imported structure, whereas Experiment 2 was designed to explore the cultural image effects. The questionnaire was used to measure speaker- and loanword-specific sociolinguistic properties. The set of these experiments was approved by Kaiārahi Māori Research and Human Ethics Committee at the University of Canterbury in 2017, and was pre-registered as #5095 using AsPredicted (https://aspredicted.org/index.php).

3.2. Experiment 1: Passage-reding task

First, the participants took part in a passage-reading task. In this task, they were asked to read 10 short passages, four of which are passages about leisure in New Zealand (hereafter referred to as neutral passages), four of which are passages about Māori culture (hereafter referred to as Māori passages), and two of which are filler passages. This manipulation of topics allows us to test the topic effects on loanword phonology. The two types of passages both include the following target loanwords, which are assembled into four groups:

Note that these words are all place names, because topics are easy to manipulate naturally. They all include word-medial rhotics /r/, of which the realizations can be classified into adapted structure [ɹ] or imported structure [ɾ] (see Section 2.1). In Experiment 1, only the word-medial rhotics are analyzed, and the word-initial rhotics in Rangiora and Rotorua are not discussed. Note that Porirua has two word-medial rhotics, which are annotated separately. These target loanwords were mentioned twice within a passage. Hence, 2,048 /r/ tokens are supposed to be collected (16 word-medial /r/ × 2 times mentioned in a passage × 2 topics × 32 participants). It was made sure that the target loanwords occur sentence-medially in the target passages. The passages employed in this task are available in Appendix D of Hashimoto (2019).

The stimuli were presented using E-prime 2.0 software (Psychology Software Tools, Pittsburg, PA). Each participant was asked to read the passages in a natural way. At the beginning of each trial, the passage appeared on the screen of a computer, and remained there for the rest of the trial. After the participants pronounced the whole passage, they pressed the space bar and the next passage appeared on the screen. They repeated this procedure until they finished reading all the ten passages.

The order of the passages was pseudo-randomized using R (R Core Team, 2016) in the following way. The passage topic alternated every two passages; that is, participants read two Māori passages after reading two neutral passages, or they read two neutral passages after reading two Māori passages. The order was counter-balanced, i.e., half the participants began with Māori topics and the other half began with neutral topics. In addition, it was ensured that the first four passages included all the loanwords belonging to the four groups. This combination was also counter-balanced, i.e., half the participants pronounced Group A and Group B loanwords in Māori passages first, and the other half pronounced Group A and Group B loanwords in neutral passages first. There were two filler passages between the first four target passages and the second four target passages, and there was a short break between the first five passages and the second five passages. The procedure of the passage-reading task is generalized in Table 1.

3.3. Experiment 2: Word-list reading task

After Experiment 1, the participants took part in Experiment 2, in which participants were asked to read aloud Māori loanwords and native words within a carrier sentence while seeing a cultural frame. The target loanwords are listed in (9). They include word-initial and/or word-medial /r/, of which realizations can be classified into adapted structure [ɹ] or imported structure [ɾ]:

Four of the target loanwords include two /r/ sounds (Porirua, Rangiora, Rotorua, and Waimakariri), and they are annotated separately. Hence, 40 realizations of /r/ can be collected from this set of target loanwords. In addition to these target loanwords, the participants read aloud 74 filler words including Māori loanwords without /r/ and native words.

Each participant was asked to pronounce these words within a carrier sentence “Say _ please” while seeing a cultural frame. The stimuli were presented using E-prime 2.0 software (Psychology Software Tools, Pittsburgh, PA). At the start of each trial, a word written in English orthography appeared within a carrier sentence on the screen of a computer and remained there for the rest of the trial. After the participants pronounced the word within a carrier sentence, they pressed the space bar and the next word appeared on the screen. They repeated this procedure until they pronounced all the words. After they finished the task, they retook the same task again; that is, each speaker took the same task twice. Hence, 2,560 tokens (32 participants × 40 /r/ tokens × 2 times) are supposed to be collected throughout this experiment.

In order to test the effect of cultural images on the likelihood of adaptation versus importation, the cultural frames were manipulated on the screen. Every time a participant pressed the space bar, a cultural frame alternated along with the word. The frames in Figure 2 were employed. The left-hand one is supposed to be strongly associated with Māori because it includes Māori cultural expression, koru, and the right-hand frame is not associated with Māori because they look like western photo frames. When this experiment was designed, some New Zealanders were asked whether the left-hand frame looks Māori-like. All the New Zealanders said that they definitely look Māori-like. Hereafter, the former types of cultural frames are called ‘Māori cultural frames,’ and the latter types of frames are called ‘neutral cultural frames.’ Note that both the types of cultural frames were coloured in four ways (red, black, light-brown, and dark-brown); that is, there are four Māori frames and four neutral frames that are employed in this experiment (see Hashimoto, 2019, Appendix E).

Figure 2

Cultural frames: Māori cultural frame versus neutral cultural frame.

The order of the words and the frames was pseudo-randomized using R (R Core Team, 2016) in the following way. First, the loanwords were assembled into two sets, Set A and Set B. In one task, Set A loanwords appeared within Māori cultural frames and Set B loanwords appeared within neutral cultural frames. In the other task, Set A loanwords appeared within neutral cultural frames and Set B loanwords appeared within Māori cultural frames. The order was counter-balanced; that is, half the participants (Participants A) pronounced Set A loanwords within Māori cultural frames first, while the other half (Participants B) pronounced Set B loanwords within Māori cultural frames first. The other native words appeared within either Māori cultural frames or neutral cultural frames in a random way. The procedure is summarized in Table 2. It was made sure that the same cultural frames were not consecutively presented, the first cultural frame was always neutral, and the first word was always a native word.

3.4. Questionnaires

After the two experiments, the participants filled in questionnaires (see Hashimoto, 2019, Appendix C, for details). In order to test Predictions 3 and 4, the questionnaires consist of two main parts. One part is about speaker-specific properties, and the other part is about word-specific properties.

As for speaker-specific properties, the participants answered 19 questions about their relationship with Māori people, culture, and language. For example, they were asked how often they watch Māori TV programs and attend Māori cultural events. They were also asked how many Māori friends and relatives they have. They also answered eight questions about their attitudes towards Māori culture and language. For instance, they were asked whether they respect fluent speakers of te reo Māori, and whether they think that it is important to study Māori culture. Most of these questions were adapted from Te Manatū Taonga (2009) and Te Puni Kōkiri (2009). Although these national surveys use Likert scales, a linear scale was employed in this study (see Dollinger, 2015, p. 251). For example, questions about the relationship with Māori were presented as in (10). The participants were asked to mark at anywhere on a horizontal line to show their answers. The line is 100 mm long. The distance of the marked point from the left edge is regarded as a rated value, that is, the point ranges between 0–100.

Principal component analyses were performed on the answers about the speaker-specific properties, with the result that the answers were reduced into three principal components, as listed in (11). The point of inflection was determined by a parallel analysis executing the fa.parallel() function (Revelle, 2017), and the model with the three factors can account for 50% of the total variance. Note that individual proficiency refers to the proficiency of a participant in te reo Māori, and neighbourhood proficiency refers to that of a community to which a participant belongs (see Poplack et al., 1988). As stated above, all the participants are not speakers of te reo Māori, and thus individual proficiency does not contribute to PC-culture so strongly, that is, the factor loading is lower in comparison to the other questions.

As for word-specific properties, the participants were explicitly required to answer how strongly a target loanword is deemed to be associated with Māori. The questions were designed to measure the words’ association with Māori in general (i.e., culture, people, and language) rather than their association with the Māori language. They rated all the Māori loanwords employed in Experiments 1 and 2, by answering the question as in (12). As with the speaker-specific properties, a linear scale was employed to measure the word-specific properties, and the distance of the marked point from the left edge is regarded as a rated value.

The raw data was standardized by z-scoring within each speaker, because the ranges of subjectively rated values vary in accordance with participants. A z-scored value tells us how a data point x_i deviates from the mean for a particular participant, and enables us to compare two data points that come from different participants. The z-scored values of subjectively rated loanwords’ association with Māori are fitted into the following statistical models, and they are called wordMaoriness. The peanut plots in Figure 3 represent the overall distribution of wordMaoriness in Experiments 1 and 2 respectively. Note that z-scoring was performed on solely a set of loanwords used in each experiment.

Figure 3

wordMaoriness of target loanwords in Experiments 1 (left) and 2 (right).

As another word-specific variable, we will statistically analyze word-frequency as a control variable in what follows. This is because the adaptation rate is known to be affected by word frequency (see Haugen, 1950; Poplack & Sankoff, 1984). There is not a large number of tokens of Māori loanwords in any corpora, and thus it is difficult to measure their objective word frequency. Hence, subjective word frequency is employed in this study. The participants were asked to rate how often they use each target loanword. This questionnaire also uses a linear scale as with the above questionnaires. As with words’ association with Māori, the rated values were standardized by z-scoring within each speaker.

3.5. Classification of /r/ realizations into adapted structure and imported structure

After the experiments, the realizations of /r/ in target loanwords were analyzed. This study is mainly based on the acoustic visualization to classify the realization of /r/ into imported structure [ɾ] and adapted structure [ɹ]. As stated below, about 90% of the /r/ realizations were annotated acoustically with a use of spectrograms. The acoustic classification was performed using Praat (Boersma & Weenink, 2016).

Following previous literature (German et al., 2013), voiced sounds with clear consonantal edges in sound spectrograms were annotated as taps [ɾ]; see Figure 4. Their domains were usually determined on the basis of soundwaves. In particular, domains without vocalic waves (i.e., major periodic waves) were labelled.

Figure 4

Sample spectrogram of imported structure (“Oamaru” by P3).

Voiced sounds without clear consonantal edges in sound spectrograms were classified as approximants [ɹ] if their F3 is lowered. It is well-known that retroflex approximants lower F3 (Olive, Greenwood, Colemen, 1993, Ch. 7; Ladefoged, 2006, p. 196). It is not easy to exactly determine the duration of [ɹ], as the formant structure is continuant with surrounding vowels. We made an attempt to annotate the domains including consistent F2 around the F3-lowering in most cases, because F2 is known as an indicator of consonantal domains in many phonetic studies (see Lavoie, 2001, p. 69, and her citations). However, we must admit that the annotation of the boundaries is still arbitrary to some extent. In fact, Umeda (1977) also notes “Some consonants such as /h/, /r/, /w/, /y/, and word-final /1/ were totally impossible to measure.” The sample annotation is shown in Figure 5.

Figure 5

Sample spectrogram of adapted structure (“Moeraki” by P5).

The remaining productions (e.g., voiced sounds without clear consonantal edges and without lowered F3) were classified impressionistically into [ɾ], [ɹ], and others (i.e., sounds that could not be identified acoustically and impressionistically). The classification results in the two experiments are reported in Sections 4.1.1 and 4.2.1 respectively.

4.1. Experiment 1: Topic effects

4.1.3. Logistic regression analyses

The 1,874 tokens of /r/-sounds were hand-fitted into a mixed-effects logistic regression model with bobyqa optimizer using the glmer function in the lme4 library (Bates, Maechler, Bolker, & Walker, 2015) implemented in R (R Core Team, 2016). We started with a model with all the variables and two random intercepts for speaker and item without any interactions. Then, some variables were removed one-by-one through pairwise comparisons of models with and without each variable. As for these variables, their interactions with the other variables were not examined, because the model including their interactions did not converge well and their p-values were not significant at all. Then, backward elimination was run manually through pairwise model comparisons using ANOVA tests by taking into consideration all the two-way interactions of the remaining variables. The elimination was based on p-value; that is, either a single effect or an interaction with the highest p-value was eliminated one-by-one. If a model comparison showed no significance (p > .05), then the smaller model was adopted; otherwise, the larger model was implemented. After the elimination of the non-significant variables, by-speaker random slopes for wordMaoriness and Session and by-item random slopes for PC-attitude were added to the model, as they improve the model. Finally, the model in Table 4 was selected as the best-fitted model. A Variance Inflation Factor (VIF) test was performed on the model, and all VIF scores were below 4, which suggests that there is no multicollinearity in the model. The marginal R squared value is 0.16, and the conditional R squared value is 0.88, according to r.squaredGLMM function in the MuMIn package (Barton, 2018). Note that the reference level is set as Topic Neutral, Session First, and firstTopic Neutral.

Table 4

Best-fitted model to predict the likelihood of choosing adapted structure [ɹ] versus imported structure [ɾ] in Experiment 1.

	Estimate	SE	z value	Pr(>|z|)
(Intercept)	–2.0172	0.8502	–2.373	0.017660	*
PC-attitude	1.7165	0.4956	3.464	0.000533	***
Topic Māori	2.5085	0.3218	7.794	6.48e-15	***
Session Second	1.4084	0.2909	4.841	1.29e-06	***
firstTopic Māori	0.1478	1.0805	0.137	0.891177
Topic Māori:Session Second	–0.9698	0.3554	–2.729	0.006357	**
Topic Māori:firstTopic Māori	–1.7942	0.3410	–5.262	1.42e-07	***

The response variable (adapted structure [ɹ] versus imported structure [ɾ]) is converted to a so-called dummy variable (i.e., 0 is given to the adapted structure [ɹ] and 1 is given to the imported structure [ɾ]) in the above model. Hence, positive slopes indicate higher likelihood of choosing imported structure, and negative slopes indicate higher likelihood of choosing adapted structure. According to the model, the fitted logit value for the intercept is –2.017. This corresponds to a fitted proportion of choosing imported structure of 0.117. That is, this model predicts that 11.7% of /r/ realizations are imported tap sounds when PC-attitude value is 0, Topic is Neutral, Session is First, and firstTopic is Neutral. The fitted logit value for the effect of Topic is 2.508. This suggests that the fitted logit value becomes 0.491 (i.e., –2.017 + 2.508), when Topic shifts from Neutral to Māori as long as PC-attitude is 0, Session is First, and firstTopic is Neutral. The fitted logit value is equal to a fitted proportion of importation of 0.62. That is, the importation rate increases by 50.3%, when topic shifts from Neutral to Māori and the two variables Session and firstTopic are First and Neutral respectively.

As two significant interactions of Topic with other variables are observed, let us consider the interactions. The interaction between Topic and Session indicates that the magnitude of the topic effect is significantly smaller in the second session than in the first session (β = –0.96, z = –2.72, p < 0.01). This interaction is illustrated in Figure 6. The left-hand figure illustrates the raw probability, and the right-hand figure illustrates the probability predicted by the best-fitted model. This is why the subsets of data including tokens produced in the first session and those produced in the second session were also explored. Then, the effect of Topic was still found to be significant in both the first session (β = 2.75, z = 7.1, p < 0.001) and the second session (β = 1.08, z = 3.11, p < 0.01).

Figure 6

Interaction between Topic and Session: Raw probability (left) and model prediction (right).

The other interaction between Topic and firstTopic suggests that the effect of Topic is stronger when participants begin with neutral passages than when participants begin with Māori passages. Once again, this interaction is illustrated by Figure 7. The subset analysis was performed, and it was found that Topic is significant when speakers begin with neutral passages (β = 2.92, z = 7.5, p < 0.001) while Topic is non-significant when speakers begin with Māori passages (β = 0.27, z = 0.9, p = 0.36).

Figure 7

Interaction between Topic and firstTopic: Raw probability (left) and model prediction (right).

In addition, the speakers’ potential association with Māori is also found to affect the likelihood of importation to some extent. The PC-attitude coefficient is positive and highly significant (β = 1.71, z = 3.46, p < 0.001). As PC-attitude means “positive attitudes towards Māori culture and language,” this coefficient indicates that imported structure is more likely to be produced by speakers with more positive attitudes towards Māori culture and language. The PC-culture was removed from the best-fitted model, as its p-value is 0.08. Note that PC-culture was significant (p < 0.05) unless by-item random slopes for PC-attitude were added to the model. The PC-culture did not achieve sufficiently significant effects, but the slope is positive and it might imply that speakers strongly related with Māori culture and language tend to use imported structure more often. On the other hand, PC-people was not significant at all (p = 0.31), and thus it was removed from the model. The variable speakerPlace was non-significant but does trend (p = 0.07). The direction indicates that North Islanders are more likely to use imported structure than South Islanders.

Next, words’ association with Māori is not found to be significant (p = 0.11), although the direction was as predicted, i.e., words strongly associated with Māori are more likely to be produced with imported structure. The variable wordPlace was non-significant but shows a trend (p = 0.07). The direction suggests that place names in the North Island are pronounced with imported structure more often. As these two variables do not reach significant levels, they were removed from the best-fitted model as with other non-significant variables.

As for other control variables, only Session shows a significant effect (p < 0.001). As the reference level is the first session and the slope is positive, this effect suggests that speakers are more likely to produce imported structure after a short break, that is, as the experiment progresses.

As a whole, the variables showing significant effects or showing a trend are as follows. Variables showing trends are shaded in Table 5.

Table 5

Summary of finding in Experiment 1.

Topic	When a passage is about Māori, speakers are more likely to produce imported structure (p < 0.001). However, speakers beginning with neutral passages do not show this topic effect (p = 0.36).
PC-attitude	Speakers with more positive attitudes towards Māori are more likely to produce imported structure (p < 0.05).
PC-culture	Speakers strongly related with Māori culture and language are more likely to produce imported structure (p = 0.08).
speakerPlace	North Islanders are more likely to produce imported structure (p = 0.07).
wordPlace	Loanwords referring to places in the North Island are more likely to be produced with imported structure (p = 0.07).
Session	Speakers are more likely to produce imported structure in the second half of the experiment than in the first half of the experiment (p < 0.001).

4.2. Experiment 2: Cultural image effects

As stated in Section 3.3, 2,560 tokens of /r/ were supposed to be collected in Experiment 2. Fifty-nine tokens were excluded because of mispronunciation or disfluency. One hundred thirty-four tokens were removed, because participants did not know some loanwords. Table 6 shows the total set of the remaining tokens.

4.2.2. Logistic regression analyses

The 2,312 tokens of /r/-sounds were hand-fitted into a mixed-effects logistic regression model with bobyqa optimizer using the glmer function in the lme4 library (Bates et al., 2015) implemented in R (R Core Team, 2016). The model selection was performed through pairwise model comparisons using ANOVA tests in the same way as the statistical analyses performed on the results of Experiment 1. As for random slopes, by-item slopes for PC-attitude, speakerPlace, and Session, and by-speaker slopes for Position and wordMaoriness were added to the model. The best-fitted model is shown in Table 7. A Variance Inflation Factor (VIF) test was performed on the model, and all VIF scores were below 4, suggesting that the model has no multicollinearity problem. The marginal R squared value is 0.35, and the conditional R squared value is 0.84, according to r.squaredGLMM function in the MuMIn package (Barton, 2018). Note that the reference levels for the following summary are speakerPlace North and Session First.

Table 7

Best-fitted model to predict the likelihood of choosing adapted structure [ɹ] versus imported structure [ɾ] in Experiment 2.

	Estimate	SE	z Value	Pr(>|z|)
(Intercept)	3.7243	0.9606	3.877	0.000106	***
PC-attitude	1.6865	0.4535	3.719	0.000200	***
wordMaoriness	0.2568	0.1281	2.004	0.045024	*
speakerPlace South	–3.4659	1.1668	–2.970	0.002974	**
Session Second	0.2939	0.1715	1.714	0.086542	.
Session Second:PC-attitude	–0.2570	0.1265	–2.031	0.042236	*

As in Section 4.1.3, the response variable (adapted structure [ɹ] versus imported structure [ɾ]) is converted to a so-called dummy variable (i.e., 0 is given to the adapted structure [ɹ] and 1 is given to the imported structure [ɾ]) in the above model. That is, positive slopes indicate higher likelihood of selecting imported structure, whereas negative slopes indicate higher likelihood of selecting adapted structure. Let us discuss key variables first. PC-attitude and wordMaoriness show significant effects in the predicted direction. The coefficient of PC-attitude is positive (β = 1.68, z = 3.71, p < 0.001), and it suggests that NZE speakers with positive attitudes towards Māori culture and language are more likely to import non-native structure. The coefficient of wordMaoriness is also positive (β = 0.25, z = 0.12, p < 0.05). This suggests that loanwords strongly associated with Māori are more likely to be produced with imported structure. The interaction between PC-attitude and Session is significant, and it was found that the effect of PC-attitude is slightly weaker in the second session, as shown in Figure 8. The subsets of data including tokens produced in the first session and in the second session were explored. It was found that the effect of PC-attitude is significant in both the first session (β = 1.55, z = 3.23, p < 0.01) and the second session (β = 1.34, z = 2.85, p < 0.01).

Figure 8

Interaction between PC-attitude and Session.

As for the other three key variables, they do not show significant effects in our data set: Frame (p = 0.09), PC-culture (p = 0.52), and PC-people (p = 0.81). Note that Frame does trend in the predicted direction, that is, imported structure is slightly more likely to be produced when Māori cultural images are presented on screen. This is illustrated in Figure 9. The importation probability is 2.2% higher given a Māori frame (65.4%) than given a neutral frame (63.2%).

Figure 9

Rate of importation according to presented cultural frame.

As for control variables, most variables are not significant, and they were removed through pair-wise comparisons. Only speakerPlace shows a significant effect (β = –3.46, z = –2.97, p < 0.01), and it was found that South Islanders are more likely to use adapted structure than North Islanders. Note that Session shows a trend, and it was found that imported structure is more likely to be produced at the second session (β = 0.29, z = 1.71, p = 0.08).

As a whole, the variables showing significant effects and a definable trend can be summarized in the following way. Variables doing trend are shadowed in Table 8.

Table 8

Summary of finding in Experiment 2.

Frame	Speakers are more likely to produce imported structure while seeing Māori cultural images (p = 0.09).
PC-attitude	Speakers with positive attitudes towards Māori are more likely to produce imported structure (p < 0.001). This effect is more robust at the first session than at the second session (p < 0.05).
wordMaoriness	Loanwords strongly associated with Māori are more likely to be produced with imported structure (p < 0.05).
speakerPlace	North Islanders are more likely to produce imported structure (p < 0.01).
Session	Imported structure is more likely to be produced at the second session than at the first session (p = 0.08).

5.1. Topic in speech

It was predicted that, speaking about Māori activates the social concept ‘Māori’ and a representation of imported structure, the result of which is that imported structure is more likely to be produced (see Prediction 1 in Section 2.3.1). Our result reported in Section 4.1 shows a robust topic effect on the likelihood of adaptation versus importation, and the direction is as predicted. Although topic effects have been reported in literature on sociolinguistic variation, such as style shifts and dialect shifts, they were yet to be explored with regards to loanword phonology. This study has extended the effects to the domain of loanword phonology, which has rarely been discussed in relation to sociolinguistic factors.

As for topic effects, there are two significant interaction effects. One interaction is with Session. This interaction indicates that topic effects are more robust earlier in the experiment in comparison to later in the experiment. This finding seems to be in line with Exemplar Theory. The exemplar-based account is as follows. The representations of the social concept ‘Māori’ and imported structure are not especially activated at the beginning of the experiment, that is, the likelihood of activating adapted structure and imported structure depends on the potential strength of each structure at this stage. As is assumed in Exemplar Theory, the potential strength is a function of the frequency of each structure, and the strength of each structure may differ in accordance with speakers and loanwords. As the experiment proceeds, participants have opportunities to read Māori passages, which further raise the activation of the social concept ‘Māori.’ Once the social concept is activated further, the extra-activation continues to some extent, and imported structure is more likely to be activated in general via the socio-indexical link between imported structure and the social concept ‘Māori.’ Consequently, topic effects become less robust in the second half, as the social concept is already activated to large extent and less likely to be activated further by a Māori topic in speech. This exemplar-based account may also be supported by the observation that the importation rate is in general higher in the second half of the experiment in comparison with in the first half (see Figure 6). It could be speculated that the extra-activation of the social concept ‘Māori’ in the first half of the experiment continues, and thus imported structure is more likely to be produced during the second half of the experiment.

The other significant interaction is that with firstTopic. Recall that some participants began with Māori passages whereas the other participants began with neutral passages. This interaction suggests that only speakers beginning with neutral passages changed their importation rates in accordance with topics of reading passages. This finding can also be captured as with the interaction of Topic and Session. When speakers begin with Māori passages, the social concept ‘Māori’ receives further activation at the very beginning of the experiment. The extra-activation may continue throughout the rest of the experiment, and the topic alternation does not affect the activation of the social category and imported structure anymore. On the other hand, when speakers begin with neutral passages, the representation of the social category ‘Māori’ is not especially activated at the beginning of the experiment. That is, the likelihood of adaptation versus importation purely depends on the potential strength of adapted structure and imported structure at the initial stage of the experiment. Once passages shift to those about Māori, the representation of the social concept ‘Māori’ starts to be further activated, and imported structure becomes more likely to be initiated via the socio-indexical link, that is, imported structure becomes more likely to be chosen in production. This is why topic effects are robust for speakers beginning with neutral passages. This exemplar-based account may be supported by the observation that participants beginning with Māori passages are more likely to produce imported structure in general (see Figure 7).

5.2. Cultural image

On the basis of previous literature, it was predicted in Section 2.3.2 that, when a cultural image associated with Māori is presented in speech, NZE speakers may be more likely to produce imported structure in speech (see Prediction 2). Our result reported in Section 4.2 indicates that this effect is tentative. As predicted, the rate of importation is slightly higher when a Māori cultural frame is presented in comparison with when a neutral cultural frame is presented (see Figure 9). However, this difference is statistically non-significant (p = 0.09). Although it was demonstrated by Hay and Drager (2010) that speech perception is influenced by cultural images as reviewed in Section 2.3.2, our result might indicate that speech production is less likely to be affected by cultural images.

Before concluding this section, we would like to point out four possible reasons which may account for the non-significant effect of cultural images on production of a linguistic variant. One reason is that participants took part in this experiment after the passage-reading task (Experiment 1). As was discussed above, the representation of a social category ‘Māori’ may be already activated to the point of saturation during Experiment 1, and this category activation may have resided until after the experiment. This is why cultural images could not further activate the representations of a social concept ‘Māori’ and imported structure, and they do not change the likelihood of adaptation versus importation significantly.

Another possibility is that participants could not pay enough attention to cultural images due to the experimental design. In this experiment, a cultural frame and a word both shifted together, when a participant read aloud a word and pressed a space bar (see Section 3.3). Although the time was not formally measured, a cultural frame may have shifted every few seconds. As a frame shifts very quickly, speakers may not have looked at frames. Besides, participants focused on a word shown at the centre of a screen, so they may not have looked at a frame shown at the edge of a screen. As a result, their speech production may not be significantly affected by a cultural frame significantly.

The null result might also be due to the nature of a word-list reading task. In Experiment 2, the effects of cultural images were explored using a word-list reading task, which may cause speech to be more careful (Labov, 1972). Due to the careful manner, the pronunciation of Māori loanwords becomes less natural and the likelihood of producing imported structure becomes higher in general, as noted above, the result of which is that the effects of cultural images may be concealed.

The other possibility is due to the degree of consciousness. Alternating images may be less explicit in comparison to alternating topics in speech, because the exposure to a cultural image is subtle and it is not exactly a part of a speech act. Our speculation is as follows: Because of the different degrees of consciousness, Māori topics may strongly raise the activation of the social category ‘Māori,’ the result of which is that the linguistic category called ‘imported structure’ is activated sequentially; Māori images may raise the activation of the social category weakly, and the activation is not strong enough to activate imported structure. Hence, topics could change the importation rate significantly, but images could not, as illustrated in Figure 11.

Figure 11

Difference in social category activation between topic and image.

If this speculation is true, then one might wonder why Hay and Drager (2010) observed the image effects on dialect perception (see Section 2.3.2). The reason might be that they explored gradient difference between vocalic variants (e.g., higher [ɪ] ⇔ lower [ɪ]), while the categories discussed in the current study have a discrete nature (i.e., tap [ɾ] versus approximant [ɹ]). It could be speculated that shifts between discrete categories require stronger activation of categories in comparison to gradient shifts, because exemplars belonging to different discrete categories may not be stored so closely as those belonging to a gradient continuum. This may be why alternating images could not significantly influence the likelihood of choosing a variant in the current study. Since the current study could not replicate cultural image effects in relation to speech production of categorical variants, it will be worth exploring the image effects by studying variants with gradient nature in future study.

5.3. Speakers’ association with the source language and its culture

As for speakers’ association with the source language and its culture, we deduced two types of predictions: the effects of relationship (Prediction 3a) and the effects of attitudes (Prediction 3b). First, the effects of relationship are not statistically supported in the current study. PC-people (i.e., relationship with Māori people) is not significant at all in the two experiments, despite the fact that some previous literature shows that the importation rates are affected by the relationship with the source language speakers (Poplack et al., 1988; Lev-Ari et al., 2014). The reason why the current study does not find this significant effect may be because most participants in the current study have a very weak relationship with Māori people and speakers. The answers to the questions about Māori people and speakers show that the participants have almost no Māori friends and are familiar with almost no Māori speakers. The participants in the current study might not be suitable to test this variable.

Although PC-culture (i.e., relationship with Māori culture and language) does not achieve a significant level, it does trend in the predicted direction (p = 0.08) in Experiment 1. This suggests that there is a slight tendency that speakers strongly related with Māori culture and language are more likely to produce imported structure. This finding is in line with Prediction 3a (see Section 2.3.3). The effect of relationship with the source language and its culture may also be supported by the significance of speakerPlace (p < 0.01) in Experiment 2, which suggests that North Islanders are more likely to produce imported structure in comparison with South Islanders. As expected in Section 4.1.2.3, this is unsurprising from the point of exemplar view. There are more Māori people living in the North Island, so North Islanders may have more opportunities to hear and store imported structure. As a result, they are more likely to produce imported structure because of its potential strength in exemplar space. This may be in line with Prediction 3a.

Prediction 3b (attitudes towards Māori) is statistically well-supported. It was found that the effect of PC-attitude is highly significant in the two experiments. As PC-attitude means “attitude towards Māori culture and language,” this finding suggests that speakers with more positive attitudes towards Māori culture and language are more likely to produce imported structure. To the best of our knowledge, this study is the first to show that importation rates are affected by attitudes towards the source language and its culture. The effect of this attitude can be captured by exemplar-based approaches as explained in Section 2.3.3; that is, speakers with more positive attitudes towards Māori may potentially activate a social concept ‘Māori.’ As a result, imported structure is also activated because it is cognitively linked with the social concept, and it is more likely to be produced. One of the anonymous reviewers pointed out another exemplar-based account for these attitude effects. The reviewer suggested that exemplars may be weighted in accordance with speakers’ attitudes towards social groups. In the case of the NZE loanword phonology, speakers with positive attitudes towards Māori may favorably weight exemplars encoding imported structure, and the exemplars are more likely to be selected in speech because of the potential prominence. The current study cannot tease apart the two theoretical interpretations, and it will be worth testing which theoretical interpretation corresponds to reality in future study.

5.4. Words’ association with the source language and its culture

We predicted that a loanword strongly associated with Māori is more likely to be produced with imported structure rather than adapted structure (see Prediction 4 in Section 2.3.4). This prediction is supported by the result of Experiment 2; that is, the effect of wordMaoriness is significant in the predicted direction (p < 0.05). Recall that this variable did not achieve significant levels in Experiment 1 (p = 0.11), although the direction was as predicted. The reason why this effect becomes statistically significant in Experiment 2 might be the number and type of loanwords employed in this experiment. The set of target loanwords in Experiment 2 is larger than that in Experiment 1, and it includes common nouns as well as place names whereas the set in Experiment 1 includes only place names. Consequently, Experiment 2 may allow us to explore a wider range of words’ association with Māori, and this factor achieves a significant level.

This effect can be captured in the following way using exemplar-based approaches: A loanword strongly associated with Māori is more likely to be heard with imported structure, and consequently exemplars with imported structure have higher strength amongst exemplars belonging to the lexical category. As exemplars with higher strength are more likely to be chosen in production, imported structure is more likely to be produced for loanwords strongly associated with Māori.

It is worth commenting on the effect of wordPlace. This variable shows a trend (p = 0.07) in Experiment 1. This variable suggests that there is a tendency that place names referring to places in the North Island are more likely to be produced with imported structure. Although this effect is tentative, it could be given the same account as Prediction 4. That is, place names in the North Island are more likely to be produced in Māori community, because there are more Māori people living in the North Island, and consequently North Island place names are more likely to be stored with imported structure.

5.5. Social meanings carried by adapted structure and imported structure in Māori loanwords

Before concluding this paper, we would like to mention social meanings carried by adapted structure [ɹ] and imported structure [ɾ] in Māori loanwords. Our results indicate that the variation in the NZE loanword phonology is a type of inter-speaker variation, because the selection of a variant depends on speakers’ attitudes towards Māori and may also depend on speakers’ relationship with Māori. The results also suggest that the variation in the NZE loanword phonology is a type of intra-speaker variation, because speakers choose the variants (i.e., imported structure or adapted structure) in accordance with topics in speech. That is, this variation is used to style speech in accordance with speech situations. Previous literature also demonstrates that the variation in loanword phonology can be an inter-speaker variation (see Poplack et al., 1988; Friesner, 2009) and an intra-speaker variation (see Lev-Ari & Peperkamp, 2014; Lev-Ari et al., 2014).

According to Bell (2014), inter-speaker variation itself does not attest the existence of social meanings, but intra-speaker variation does. On the other hand, Eckert (2016) notes that variation is always socially meaningful. In either case, our results suggest that the variants in loanword phonology carry particular social meanings. The attitude effects reported in the current study may suggest that adapted structure carries a social meaning dissociated from Māori while imported structure carries a social meaning associated with Māori. Exploration of the exact social message may require perceptual studies as in Campbell-Kibler (2007), and it is left for future study.

Loanword phonology may provide an interesting test case to explore social meaning in linguistic variation. Hall-Lew, Starr, and Coppock (2012) demonstrate that the second vowel of a loanword Iraq can be realized as either /æ/ or /a:/, and that this variation is used by politicians to express their social identity. It is worth exploring what kind of social message, in general, can be expressed by choosing a variant of loanword pronunciation.