Abstract
Aims and Objectives/Purpose/Research Questions:
There is ongoing discussion as to the cost of language switching, with some studies indicating high cost and others showing low or no cost. The main research question in this paper is whether there are language switching costs in communities in which codeswitching is frequent.
Design/Methodology/Approach:
We conducted two on-line experiments, that is, a picture choice with sentence auditory stimuli and a word recognition task in sentence context. We constructed 16 sentences with differing degrees of ecological validity (16 sentences × 4 versions = 64). The sentences included verbs with different language preferences in natural conversations (first language (L1), second language (L2) or both).
Data and Analysis:
Thirty-seven simultaneous L1-Romani L2-Turkish bilinguals participated in Experiment 1 and 49 in Experiment 2. To analyse the results, linear mixed models (lmer) were constructed using the ‘lme4’ package in R.
Findings/Conclusions:
In Experiment 1, participants responded significantly faster for the all-Turkish sentences, followed by the mixed Romani-Turkish sentences, and the two types of ecologically non-valid sentences. However, there were no processing costs for the mixed sentences when they contained Turkish verbs that are more frequently used in Turkish in the spontaneous conversations. In Experiment 2, reaction times were similar for Turkish verbs (with Turkish verb morphology) in a mixed Romani-Turkish or a unilingual Turkish sentence.
Originality:
Taken together, these findings indicate that language switching costs in comprehension depend on the frequency of codeswitching in the bilingual community, as well as on exposure to specific lexical items.
Significance/Implications:
The Romani-Turkish data support a usage-based approach to bilingual processing and confirm the need to conduct experimental research that takes into account the communicational habits of the participants.
Processing and language switching costs
Depending on the interactional setting, bilinguals may switch from one language to another or consistently keep the two languages separate. This constant activity of language selection mobilizes executive functions that are claimed to also partake in a variety of cognitive tasks (Bialystok, Craik, Klein, & Viswanathan, 2004). Yet, it is still unclear to what extent switching between languages is an activity associated with high cognitive costs on par with costs that systematically arise in nonlinguistic task switching (see Monsell, 2003, for a review). Indeed, there is ongoing discussion as to the cost of language switching, with some studies indicating high cost (Alvarez, Holcomb, & Grainger, 2003; Costa & Santesteban, 2004; Grainger & Beauvillain, 1987; Jackson, Swainson, Cunnington, Jackson, 2001; Meuter & Allport, 1999; Proverbio, Leoni, Zani, 2004; Soares & Grosjean, 1984; Thomas & Allport, 2000) and others showing low or no cost (Gullifer, Kroll, Dussias, 2013; Ibáñez, Macizo, Bajo, 2010; Jackson, Swainson, Mullin, Cunnington, Jackson, 2004; Moreno, Federmeier, Kutas, 2002; Mosca & Clahsen, 2016), both in comprehension and production.
Attention has been paid to a number of factors that can modulate these costs, such as age and type of acquisition, proficiency, socioeconomic background and the type of tasks and stimuli. It has been shown, for example, that second language (L2) learners process switching differently from highly proficient and early bilinguals, the latter showing an unexpected processing ‘advantage of L2 (L3) over L1’ (Costa & Santesteban, 2004). Similarly, research based on neuroimaging shows that naming in the first language (L1) after naming in a L2 is associated with a greater effort in reactivating the dominant L1 that had been inhibited (Abutalebi & Green, 2016). Moreover, for comprehension tasks, switching costs across a L1 and a L2 are symmetrical (Thomas & Allport, 2000), but there can be asymmetrical switching costs across a L1 and a L2 in production (Costa & Santesteban, 2004; Meuter & Allport, 1999). Ultimately though, these costs may be overcome with more preparation time (Mosca & Clahsen, 2016). Also, although socioeconomic status has an effect on language and cognition, it appears to be independent from bilingualism (Calvo & Bialystok, 2014).
More importantly for the purpose of our study, it has been argued that the cost of translation is different from that of inter-sentential switching (Gullifer et al., 2013; Ibáñez et al., 2010) and that the cost of ecologically valid codeswitching is similar to that of unilingual passages in the L1 as opposed to unexpected, irregular codeswitching and L2 (Chan, Chau, Hoosain, 1983). Moreno et al. (2002) confirm that an unexpectedness effect may arise from switchings if they are not the ones usually encountered in a given bilingual community. Moreno and colleagues further observe that written stimuli are predominantly used in experiments, despite the fact that codeswitching generally occurs orally. We suggest that stimuli should also reflect the codeswitching constraints that prevail in a given community. For example, the use of an English determiner preceding a Spanish noun is possible but not frequent in naturalistic data (Herring, Deuchar, Parafita Couto, & Quintanilla, 2010). In consequence, stimuli such as “He heard a knock at the puerta”, used in Moreno et al. (2002, p. 191), may be unexpected not because of the language switch from English to Spanish, but because of the type of switch.
The importance of the communicational habits in language processing was put forward by Green and Abutalebi (2013) through the “adaptive control hypothesis”, which stresses the capacity of the language control network to adapt to the needs of the interactional setting. More generally, usage-based research argues for a strong link between cognitive representations and frequency of use (Tomasello, 2003, for language acquisition; Goldberg, 2006, for constructions; Backus, 2015, for codeswitching). Recent studies also demonstrate how language processing is facilitated by statistical frequencies within the experimental study as well as beyond (Jaeger & Snider, 2013; MacDonald, 2013; Wells, Christiansen, Race, Acheson, MacDonald, 2009). Such accounts consider that language processing is shaped by expectations and prediction errors based on recent and long-term exposure. Speakers opt for the least costly means in language production based on greater statistical regularities while simultaneously reinforcing them, and comprehenders rely on these regularities when processing linguistic input. In that sense, the cognitive processes involved in language switching and translation tested in the laboratory may differ from those involved in codeswitching as it occurs in natural communicative settings. More specifically, Green (1998, 2011) suggests that bilinguals from communities with frequent codeswitching rely more heavily on the joint activation of the two languages, as opposed to bilinguals from communities that do not frequently codeswitch and therefore rely more heavily on their language control network to avoid conflict between the two languages.
In order to fully comprehend the adaptive capacities of the bilingual brain, researchers call for more research taking into consideration the conversational practices of bilinguals (see Abutalebi & Green, 2016; Hofweber, Marinis & Treffers-Daller, 2016). To date, however, there is still little research on cognitive processing among bilinguals from communities that codeswitch systematically. This paper is an attempt to help fill this gap by providing the first experimental study on sentence processing in a community showing intense, well-established codeswitching between Romani and Turkish.
Romani-Turkish as spoken in Greek Thrace
Romani is an Indic language of the Indo-European family. Roma had arrived in the Byzantine Empire by the 11th century, in what are now Turkey and Greece (Matras, 2002). At present, approximately 200,000 Roma live in Greece (Bakker, 2001). In this paper we focus on a Romani community settled in the Drosero neighbourhood at the outskirts of the city of Xanthi in Greek Thrace; see the map in Figure 1. The community has approximately 4,000 members of low socioeconomic status.
Map of the area in Thrace, Greece. The research was conducted in the town of Xanthi.
Roma in Xanthi are typically trilingual in Romani, Turkish and Greek. The dominant Romani dialect spoken in Xanthi is characterized by the frequent insertion of Turkish words from all parts of speech with the exception of pronouns, amounting to approximately 15% of all words in natural conversation (Adamou & Granqvist, 2015, p. 531). The resulting Romani-Turkish variety is called Xoraxane Romane ‘Turkish Romani’ by its speakers. It is the home language and the language of in-group communication in this tightly knit community.
The use of Romani-Turkish among the Muslim Roma of Greek Thrace has attracted some attention in the literature because of the typologically rare use of Turkish verb morphology with Turkish verbs inserted in Romani dominant speech (Adamou, 2010; Adamou, 2016; Adamou & Granqvist, 2015). An example illustrating Romani with Turkish codeswitching insertions is given in (1).
Greek Thrace Romani < Romani (in regular font), Turkish (in bold) (1) always 1 me 1 me 1 “Am I
An analysis of a free-speech corpus with Romani as the dominant language indicates that speakers use 12% of Turkish verbs and that these verbs are always inserted along with the Turkish person and tense-aspect-modality morphology (Adamou & Granqvist, 2015). Adamou and Granqvist (2015) suggested that the Romani-Turkish data resemble the early stages of so-called “mixed languages”, which were shown to result from codeswitching (Auer, 1998; McConvell & Meakins, 2005; O’Shannessy, 2012).
Although insertion of non-integrated contact verbs is rare cross-linguistically (Wohlgemuth, 2009), it has been reported for several Romani dialects in Europe in contact with typologically diverse languages, that is, North Russian Romani in contact with Russian (Rusakov, 2001), Finnish Romani in contact with Finnish (Adamou & Granqvist, 2015), Crimean Romani in contact with Turkish (Elšík & Matras, 2006, p. 135), Lithuanian Romani in contact with Russian (Elšík & Matras, 2006, p. 135) and several dialects spoken in the Balkans in contact with Turkish (Friedman, 2013). This suggests that sociolinguistic factors are at play, such as extensive and intensive bilingualism without any strong prescriptive attitudes (Adamou, 2010).
Besides the Romani-Turkish variety, the Roma of Xanthi also speak a non-native variety of Turkish that they use in traditional trade, most likely since Ottoman times – from the 15th to the early 20th century (Adamou, 2010). In the Greek Thrace context, Turkish is the minority language of the Greek Thrace Muslim minority. As a consequence, all members of the Muslim community, independent of their L1, have the right to Greek-Turkish bilingual education in primary school. Secondary education is strictly provided in Greek, which is also the language of administration and services. In practice, the effects of formal education on Roma should be viewed with some caution, as access to schooling is strongly affected by the social exclusion that characterizes the relations between the Greek State and the Romani communities. Indeed, Roma over the age of 30 from Xanthi have at best attended Greek primary school, with high rates of illiteracy. Most Roma under 30 have low school attendance rates and often drop out before completing middle school.
Roma in Xanthi generally acquire both Romani-Turkish and Turkish before the age of three at home and in the community with differing degrees of exposure. We will refer here to the traditional community language, Romani-Turkish, as the L1 and Turkish as the L2. Despite a current strong shift to Turkish, one must bear in mind that Romani children partake in a wide range of interactions: they often grow up in households alongside their parent(s) and siblings and their paternal grandparents, as well as their parents’ siblings, spouses and children. It is also important to note that most families receive daily visits from extended family, friends and neighbours and that children are quite free in their movements within the community where Romani is still in use. In addition, children of all ages often follow their primary caregivers to work and are thus exposed to Turkish and Greek as spoken outside their community. Table 1 summarizes the trilingual setting for most adult Roma in Xanthi.
The trilingual setting for most adult Roma in Xanthi.
L1: first language; L2: second language; L3: third language.
Experiment 1: Picture choice with auditory sentence stimuli
Experiment 1 is an on-line, bimodal picture–sentence matching task with auditory stimuli. This experiment is inspired by the visual world paradigm with eye-tracking used in Dahan and Tanenhouse (2004) in that it also uses visual and aurally presented stimuli. The research question that this experiment addresses is whether mixed Romani-Turkish sentences have higher processing costs than unilingual L2-Turkish sentences. We therefore wanted to test the reaction times (RTs) of the Romani participants with respect to Romani-Turkish mixing within sentences (local costs) in a demanding task with frequent language switches between sentences (global costs). In line with usage-based models, Romani-Turkish mixing should be treated similarly to unilingual speech, as it is very frequent in the community (Adamou & Granqvist, 2015). However, in line with a number of studies on language switching costs, the mixed Romani-Turkish sentences should be associated to higher costs than unilingual speech (i.e. speech in a single language).
Methodology
Participants
Thirty-seven Roma participated in the task. They were contacted through a local non-profit organization based on their proficiency in both Romani-Turkish and Turkish. All gave oral informed consent and, in agreement with the organization’s representatives, they received no compensation for their participation in the study. 1 Participants were all residents of the community of Drosero in Xanthi and were of similar socioeconomic status, that is, low income levels. All but one had low education levels, that is, 10 attended at most primary school, 26 attended at most secondary school and one had a university degree. Twenty-three participants were female and 14 male. The age of the participants ranges from 13 to 51 (M = 22.59, SD = 11.13). All the participants declared understanding and speaking Romani and Turkish. They have received no formal education in Romani, a language with no written tradition, nor Turkish, which they acquired in oral interactions within the community and with outsiders. All declared that they acquired Romani-Turkish and Turkish before the age of three, with varying degrees of exposure to each language that it was not possible to measure. Twenty-seven participants declared Romani to be their primary language of communication (age M = 25.6, SD = 11.77), while 10 declared it was Turkish (age M = 17.87, SD = 1.34).
Materials
Sixteen two-sentence stimuli were created, consisting of a first sentence with a Turkish noun at its beginning and a second sentence with a Turkish verb at its beginning. The first and second sentences were related in meaning and were culturally adapted; see Appendix 1 for the list of sentences translated into English. Four versions of each sentence were recorded with a fluent female, native Romani speaker from the locality, providing a total of 64 auditory stimuli. The sentences were carefully constructed in collaboration with this speaker and were repeated orally until reaching a fluent and naturalistic version. The number of words for the four versions was kept as similar as possible, that is, between five and seven words per sentence. The recordings were done with a Tascam DR-100 solid-state recorder via a supercardioid head-worn microphone. Using Audacity, long hesitations, typically occurring between the first and the second sentence, were reduced to achieve similarity in the total duration of the audio stimuli. The resulting sentences for the four versions were similar in length with a mean duration of 6800 ms. The length of the sound input in milliseconds was also integrated in the statistical models.
In Version a, the sentences were recorded using the Romani-Turkish mixing patterns that are frequent in the community. In this version, exemplified in (2a), the Turkish noun is integrated into Romani morphology, while the Turkish verb maintains its original Turkish morphology.
Version b contains strictly Turkish material and was produced in the local Turkish variety. This resulted in verb initial sentences, parallel to the Romani sentences, even though Turkish is a verb final language; see (2b).
Version c, shown in (2c), is close to so-called classic codeswitching and was constructed by enunciating one part of the sentence in Romani and the other in Turkish with morphologically non-integrated nouns and verbs and with words from word classes that are otherwise used in Romani, such as pronouns. This option is possible in other Romani varieties, but is not attested in the free conversational corpus of the community (Adamou & Granqvist, 2015).
Finally, Version d included sentences in Romani with Turkish verbs adapted in the Romani morphology, like borrowings; see example (2d). This possibility is attested in other Romani varieties of the Balkans but is not used in the local Romani variety.
Romani (in regular font), Turkish (in bold), Greek (underscored) (2) ”The neighbours were having parties very often. They drank and danced until late.” Mixed Romani-Turkish a. e DEF.PL neighbour-PL very often make-3PL-IMPF party-PL drink-AOR-PST-PL eat-3PL-IMPF very late every day Turkish b. neighbour-PL every day party make-PROG-3PL-PST drink-AOR-PST-PL eat-AOR-PST-3PL every day very late time- until DAT Romani with Turkish codeswitching c. neighbour-PL every day make-3PL-IMPF party 3PL drink-AOR-PST-PL eat-3PL-IMPF dance-3PL-IMPF until very late Romani with Turkish borrowings d. e DEF.PL neighbour-PL every day make-3PL-IMPF party-PL drink-LVM-3PL-IMPF eat-3PL-IMPF dance-3PL-IMPF very late
In sentence (2a), we note that the Turkish noun “neighbour” is adapted to Romani morphology through the plural suffix –je and the use of the definite article. As expected, the Turkish verb “drink” has Turkish morphology, but in the rest of the sentence, the two other verbs are in Romani, and bear Romani verb morphology. The noun “party”, [γlendja], can be identified as a codeswitch to Greek (natural conversations contain approximately 4% of Greek words, mainly nouns). We also note the use of the Turkish adverb “late”.
In sentence (2b), entirely in Turkish, we note that the pronunciation of the word “every” [ör] not only differs from the usual Turkish, [her], but also from its usual pronunciation as a borrowing in Romani, [er].
Sentence (2c) contains codeswitching alternations between Turkish and Romani. It includes the Turkish noun “neighbours” with the Turkish plural, which is never used as such in the Romani-Turkish mixed speech, as illustrated in (2a). We also note the use of the free pronoun for the Turkish third-person plural, onlar, which is also distinct from the Romani-Turkish mixed form, onnar, Romani on and Turkish onlar.
In sentence (2d), the Turkish verbs and nouns bear Romani morphology, for example, the noun “party” keeps its Turkish phonetics but takes the Romani plural marker -a.
Moreover, in order to test the difference in effect between a highly predictable switch and a less expected one, we took into consideration the lexical frequency of the verbs as observed in natural speech (Adamou, 2016). Five sentences included a Turkish verb that is more frequently used in Turkish than in Romani (“Turkish”), six sentences included a verb that is attested in the corpus once and for which the Romani variant is more frequently encountered (“Romani”) and five sentences included a Turkish verb that is used with equal frequency in both Turkish and in Romani (“Variable”). We predict that the verbs that are preferably used in Turkish will be associated with short RTs, that the verbs preferably used in Romani will be associated with longer RTs and that the verbs that are used in either Romani or Turkish will be associated with long RTs.
We further conducted an offline norming study for the four versions of the sentences. Four Roma from the community (ages 18–25) who did not participate in the experiment were asked to rate the sentences with respect to the likelihood of hearing them in the community. They were asked to provide their ratings on a Likert five-point scale, from 1 (not likely) to 5 (very likely). The ratings were noted by the researcher. Each participant listened to 16 sentences in total and versions were counterbalanced across participants. Version a sentences received a mean rating M = 4.6, Version b M = 4.1, Version c M = 2.3 and Version d M = 0.4. In combination with the naturalistic evidence, this result confirms that Versions a and b are the most natural in the community, whereas Versions c and d are the most unnatural.
Sixteen pairs of pictures served as material for the target sentences and two pairs of pictures were used for the trials. See Appendix 2 for an example of visual stimuli. We did not conduct a norming study for agreement between the pictures and the sentences, since we did not predict any pattern with respect to accuracy in the picture choice.
Procedure
The participants were tested in the office of a local non-profit organization, in a calm environment. They were seated in front of a computer screen and wore headphones. The experiment was conducted on a computer using Open Sesame (Mathôt, Schreij, Theeuwes, 2012). The stimuli were fully randomized. The instructions were provided in Greek by the researcher, that is, in the language of school, and in Romani-Turkish and Turkish by a local assistant, thus creating a multilingual environment (Soares & Grosjean, 1984). The Romani participants were invited to watch several pairs of pictures and listen to the synchronized auditory stimuli, which appeared simultaneously. They were told that they would listen to various languages, Romani as spoken in their community (i.e. Xoraxane Romane “Romani-Turkish”), Turkish and Greek, so that they could anticipate the frequent inter-language switchings (see Ibáñez et al., 2010, who note the capacity of translators to anticipate language switches). Participants were instructed to select the picture that appeared to be more closely related to the meaning of the audio stimuli by pressing a button on the computer: a left arrow button if they wanted to select the picture on the left of the screen, and a right arrow button for the picture showing on the right. They were told that they should press the button as soon as possible and not to wait for the completion of the sentences. Inter-trial intervals were not controlled for and were left up to the participant. The task started with two warm-up trials. Each participant responded to 16 trials consisting of four sentences for each version (a, b, c and d). For 36 participants, this experiment took place after Experiment 2 and in this case the participants were familiarized with the sentences; 13 participants first completed Experiment 1 before completing Experiment 2. 2 Changes in the order of the tasks were made to avoid any confounding effects. There were short breaks (2–5 minutes) between the two tasks. A pilot study with three participants included five fillers in the experiments. However, we noted that participants were distracted before the end of each task or did not wish to complete it. For this reason, we decided not to include the fillers in the final version. Once the experiment was completed, participants declared that they did not find the task difficult and appeared amused by the pictures. They were also happy to listen to their native language via a computer.
Analysis
Linear mixed models (lmer) were constructed using the “lme4” package (Bates et al., 2015) in R (R Core Team, 2013) to analyse the results. The dependent variable is the RT, and the independent variables are “Language preference” for the verb in the free-speech corpus (Variable, Romani and Turkish), and “Versions” of the sentences in the experiment (a, b, c and d). The “Subjects”, “Sentences” and “Duration” of audio files in terms of the acoustic input time in milliseconds are coded as random factors. We eliminated the outliers of the RTs that were faster than 500 ms and slower than 20000 ms, which was about 0.7% of the total data.
Results
The participants responded the fastest for the unilingual Turkish sentences (Version b), followed by the Romani-Turkish mixed sentences (Version a) and the codeswitches (Version c). The mean RTs are as follows: Version a = 5756.4 ms; Version b = 5178.4 ms; Version c = 5914.1 ms; and Version d = 6409.9 ms; see Figure 2.
Mean reaction times (RTs) in milliseconds (ms) of different versions of the sentences (a = mixed Romani-Turkish; b = all Turkish; c = Romani with Turkish codeswitching; d = Romani with Turkish borrowings) in Experiment 1 (picture choice).
The analysis of variance (ANOVA) shows that there are significant “Version” differences (the logLikelihood is −6900.9, χ23 = 19.0, p < .001). Sentences entirely in Turkish, Version b, are significantly shorter than all others, that is, Version a (t = 2.7), Version c (t = 2.3) and Version d (t = 4.3). The codeswitching Version c is also significantly shorter than Version d, which contains atypical Turkish verbal forms with Romani morphology (t = 2.0). Mixed Romani-Turkish Version a is marginally shorter than Version d (t = 1.5). However, there is no significant difference between the mixed Romani-Turkish version (Version a) and Version c involving codeswitching (t < 1).
We then conducted a more fine-grained analysis by considering the RTs with respect to the language preference of the verbs in natural speech. Although this experiment does not target the verbs in particular, the RTs indicate that participants reached the beginning of the second sentence, including the Turkish verb, before pressing the button. Verbs were classified in three types depending on preferred language when Roma speak in the Romani-Turkish variety. They are tagged “Variable” when speakers in the community use verbs in Romani or in Turkish, “Romani”, when they mainly use the verbs in Romani, although some peripheral occurrences in Turkish can be observed, and “Turkish” when the verbs are mainly used in Turkish. The mean RTs of the experiment are presented in Table 2.
Experiment 1: mean reaction times of different versions of the sentences (a = mixed Romani-Turkish; b = all Turkish; c = Romani with Turkish codeswitching; d = Romani with Turkish borrowings) with different language preferences for the verbs as observed in a free-speech corpus (Variable, Romani, Turkish).
The analysis of variance shows significant “Version” differences (the logLikelihood is −6900.9, χ23 = 19.0, p < .001) and “Language preference of verb in the corpus” differences (the logLikelihood is −6898.3, χ22 = 11.5, p < .003).
For “Language preference of verbs in the corpus” we note that “Variable” verbs have the longest RTs (ts > 2.5). For “Variable” and “Romani” verbs, all-Turkish (Version b) and codeswitching (Version c) have shorter RTs than mixed Romani-Turkish (Version a) and borrowing (Version d) (ts > 1.9). For the verbs most frequent in “Turkish” in natural speech, both the mixed Romani-Turkish and the all-Turkish versions (Versions a and b) are shorter than the two ecologically non-valid versions (Versions c and d) (ts > 1.9); see Figure 3.
Mean reaction times (RTs) in milliseconds (ms) of different versions of the sentences (a = mixed Romani-Turkish; b = all Turkish; c = Romani with Turkish codeswitching; d = Romani with Turkish borrowings) with different language preferences for the verbs as observed in a free-speech corpus (variable, Romani, Turkish) in Experiment 1 (picture choice).
Discussion
In this experiment, participants responded faster for the unilingual Turkish sentences (Version b) than for the mixed Romani-Turkish sentences (Version a). The mixed Romani-Turkish sentences (Version a) were similar to codeswitching (Version c). This result can be understood in relation to other studies on bilingual processing, which indicate that understanding speech in a single language is the least costly. However, when language preference of the verbs in the natural speech of the community is taken into consideration, it appears that for the verbs that were more frequently Turkish in the corpus, mixed Romani-Turkish sentences were as fast as Turkish unilingual sentences. It is interesting to note that the slowest RTs arise when both Romani and Turkish verbs are used in the community. This might indicate that variation in language choice entails processing costs in comprehension.
In conclusion, we consider that the models of processing that best accommodate the Romani-Turkish results are those that integrate the expectations of the comprehenders based on communicational habits established prior to the experiment (Jaeger & Snider, 2013). These expectations are not limited to the likelihood of listening to Turkish words in a Romani sentence, but appear to be sensitive to more fine-grained probabilities related to specific lexical items, as widely demonstrated in usage-based research.
A limitation of this experiment is that RTs provide information on the processing of entire sentences, since participants could press the button at any point in a sentence, but not directly on the processing costs of the Turkish verbs. Also, possible confusion may have arisen from difficulties with picture association rather than with processing. Experiment 2 was designed to overcome these limitations.
Experiment 2: Word recognition in sentence context
Experiment 2 is an on-line task with auditory sentence stimuli using the word monitoring paradigm (Kilborn & Moss, 1996). The experiment is designed to determine whether the morphologically non-integrated Turkish verbs have higher processing costs when they occur in a mixed Romani-Turkish environment than in a unilingual Turkish environment. In line with studies showing language switching costs, the prediction would be that the morphologically non-integrated verbs will have high processing costs in Romani-Turkish sentences. However, Adamou and Granqvist (2015) show that the Turkish verbs are frequent switches in the Romani community of Xanthi. In line with usage-based models, Turkish insertional switches should be processed with low costs, similar to unilingual speech.
Methodology
Participants
Forty-nine trilingual Romani-Turkish-Greek speakers participated in this experiment. As in Experiment 1, participants were contacted through a local non-profit organization. All gave oral informed consent and received no compensation for their participation in the study following agreement with the organization’s representatives. Participants were of similar socioeconomic status, that is, low income levels and low education levels, that is, 18 attended at most primary school, 30 middle school or secondary school and one had a university degree. Thirty-four participants were female and 15 male. Ages ranged from 13 to 50 (M = 24.10, SD = 11.5). As in Experiment 1, all the participants declared that they acquired Romani-Turkish and Turkish simultaneously before the age of three, although they had varying degrees of exposure to each language at the time. Thirty-nine declared Romani to be their primary language of communication (age M = 26.5, SD = 11.72) and 10 declared that Turkish is their primary language (age M = 17.87, SD = 1.34).
Materials
We used the same auditory stimuli as in Experiment 1. The same four versions that were used in Experiment 1 were also used for this experiment, that is, (a) mixed Romani-Turkish, (b) all Turkish, (c) Romani-Turkish codeswitching and (d) Romani with Turkish borrowings. Similar to Experiment 1, language preference in a free-speech corpus for the Turkish verbs that were used in the experiment was also taken into consideration: Variable, Romani, Turkish (see Adamou, 2016).
The target word in this experiment was the word that immediately followed the Turkish verb, the latter being the prime. As can be seen in Table 3, the prime was practically the same in Versions a, b and c, and differed slightly in Version d. The target was the same in Versions a, c and d, but different in Version b.
Stimulus examples from Experiment 2 (Romani in regular font, Turkish in bold, Greek underscored). Transcriptions are given in IPA the International Phonetic Alphabet (IPA) for versions a, c and d and in Turkish script for version b.
| Version | Auditory stimuli for the sentences: |
Target | Prime |
|---|---|---|---|
| a. Mixed Romani-Turkish | E |
Xanas |
|
| b. All Turkish |
| Yerdiler |
|
| c. Romani with Turkish codeswitching | Xanas |
|
|
| d. Romani with Turkish borrowings | E |
Xanas |
