Code-switching in computer-mediated communication by Gen Z Japanese Americans

1 Introduction

Code-switching refers to the phenomenon of switching between two or more languages when communicating, and it is argued that bilingual speakers code-switch to express themselves to the fullest (Appel and Muysken 2005). As bilingual speakers increase globally, code-switching has become a more common language phenomenon. In the United States, the percentage of people speaking a language other than English increased from 18 % in 2003 to 21.6 % in 2021 (US Census Bureau 2021). The rise in code-switching has sparked growing research interest, including consideration of the motivations for code-switching and the grammatical patterns of code-switching utterances. Code-switching between languages is not to be confused with temporary lexical borrowing from another language. As clarified by Myers-Scotton (2017), a monolingual person can borrow lexemes from another language, while only bilingual people engage in code-switching. Researchers have observed two main types of code-switching in terms of when the switch occurs: “intra-sentential” code-switching, where language switching happens within a sentence; and “inter-sentential” code-switching, where switching happens between sentences (Milroy and Muysken 1995). Studies have also categorized code-switching by the ratio of the two languages involved: either a host language with occasional embedded language or an even combination of both languages (Nishimura 1995).

With the popularization of social media and online communication tools (such as email and message apps), code-switching in computer-mediated communication (CMC) has become an emerging branch of research. Researchers have reported that code-switching in CMC presents some new linguistic, sociopragmatic, and communicative patterns that have not been previously observed and have argued that code-switching in CMC serves specific new sociolinguistic functions (Androutsopoulos 2007; Lin 2005; Montes-Alcalá 2024; Ting and Yeo 2019; among others).

However, little attention has been given to Japanese-English CMC code-switching, especially in bilingual texting, with most research focusing on oral code-switching data. Nishimura (1995) examined oral code-switching among second-generation Japanese Canadians (known as “Niseis”) from prewar Japanese neighborhoods and found that Niseis spoke primarily in English to fellow Niseis but switched between English and Japanese evenly when speaking to a group of native Japanese speakers and fellow Japanese Canadians. Shepherd (2021) confirmed that Nishimura’s patterns of bilingual speech (the “basically Japanese” pattern, the “basically English” pattern, and the “mixed” pattern) are still being used among contemporary Japanese Americans’ oral communications; however, the postwar second-generation Japanese Americans use primarily English when speaking to a group of native speakers and fellow Japanese Americans. While Nishimura’s and Shepherd’s studies focused on Japanese-English code-switching in English-speaking countries, Azuma (1997) obtained datasets from native Japanese speakers in Japan, including natural conversations among female college students returning from a year abroad at a US university and disk jockeys’ speech on FM radio stations playing American pop music. Azuma argued that significantly more code-switching in disk jockeys’ speech is likely attributable to the prevalence of English songs, but both datasets revealed that nouns, verbal nouns, and adjectivals (referred to as keiyoo-dooshi in Japanese) were the most frequently switched parts of speech. These parts of speech all possess the feature [+N] in Japanese, so they can meaningfully stand alone. In contrast to studies focusing on analyzing specific datasets, Nakayama et al. (2018) described the process of constructing a large-scale Japanese-English oral code-switching corpus. The corpus showed up to 59 code-switching instances within a 30-min audio recording, emphasizing the necessity for speech recognition models to handle both monolingual speech and code-switched speech.

The present study aims to contribute to the emerging study of code-switching in CMC by examining code-switching among Japanese Americans who code-switch between Japanese and English when sending text messages to each other. This study allows us to analyze when bilingual speakers choose to use one language over the other and notice patterns that might not be so obvious in spontaneous oral speech. Code-switching in CMC has been argued to be an identity marker among young persons (Gonzales and Tsang 2023; Lienard and Penloup 2011; Montes-Alcalá 2007). With a focus on CMC code-switching among Gen Z Japanese Americans, the present study can also contribute to the understanding of young persons’ CMC code-switching patterns and reveal insights into language dynamics and cultural identity among young bilingual speakers.

Moreover, the present study enables us to explore CMC code-switching patterns that may not be as apparent in code-switching scenarios within a single language family, where the languages have similar grammatical structure and writing systems (Barasa 2016; Callahan 2004; Estigarribia and Wilkins 2018; Halim and Maros 2014; Leppänen’s 2012; Montes-Alcalá 2005, 2024; Tsiplakou 2009; among others). Japanese, as a non-Indo-European language, has very different grammar from English. For example, Japanese follows a basic word order of subject-object-verb and has an extensive grammatical system for politeness, unlike English. Additionally, also unlike English which only has an alphabetic writing system, Japanese employs a mix of different writing systems: a logographic writing system (kanji), two syllabaries (hiragana and katakana), and a romanized alphabetic writing system (romaji) that has recently developed and is not widely used nor official (Matsuda 2017). Due to the distinct features of writing systems in English and Japanese, the present study can help us investigate the interactions between writing systems and language code-switching, which have barely been reported in the literature.

Before presenting the dataset and analysis, we would like to clarify some terminology-related points. First, we chose “code-switching” over “translanguaging” or “code-mixing” in this paper, although all three terms apply to multilingual data. The speakers in our dataset are aware of the two distinct languages as they message, and they have naturally switched between the languages since childhood. However, translanguaging views languages as integrated into one grammar with a unified set of linguistic features – a unitary grammar (Otheguy et al. 2019). Additionally, translanguaging tends to be widely used in relation to bilingual education contexts (Liu 2023). Thus, we adopted a code-switching framework for this paper. Code-mixing often refers to alternating between languages within a sentence, but code-switching is a broader term covering language switches within and between sentences (Mabule 2015). The second terminology-related point to note is that texting can be a semi-synchronous form of CMC, differing from blogs, emails, and online forums (Androutsopoulos 2013). Although texting is an informal CMC with conversational content like oral communication, it remains in written form. Typing and sending messages requires more deliberation and steps than simply verbalizing thoughts, making it valuable for studying code-switching patterns that may not be apparent in spontaneous speech.

2 Methodology and dataset

The dataset consists of text messages among three young adult female Japanese-English bilinguals (ages 21–22 at the time of data collection). One of the authors is one of the participants, but this author did not have the idea of a study such as this in mind when these messages were exchanged. The three speakers, referred to as speakers A, B, and C, are simultaneous bilinguals fluent in both Japanese and English who have frequently spoken both languages and have resided in the US since childhood (see Table 1). They are friends with overlapping social networks as they are all part of the local Japanese community. They have had regular exposure to both English and Japanese through their community and online media.

The 1,561 text messages were extracted from two chat groups. One set consists of direct messages (DMs) exchanged between speaker A and C between 21 August 2022 and 22 February 2023. The other set is from a group chat (GC) among speakers A, B, and C between 9 September 2022 and 20 February 2023. All three speakers gave their informed consent through online communication to allow the text messages to be exported from iMessages on a Mac using the terminal and SQLite commands. After exporting the text messages, the chat data was formatted into a CSV file.

The dataset cleaning involved three steps. First, blank messages, including those resulting from images or pictures, were removed, as visual messages are not directly related to the focus of Japanese-English code-switching. Subsequently, reaction messages such as “liked a message”, “loved a message”, and “emphasized a message” as well as emojis were removed. Lastly, we anonymized identifiable information by deleting social media links and birthdays and anonymizing names and addresses. Table 2 outlines the number of messages removed at each step. After cleaning up the data, there were 1,198 messages across the DMs and the GC. Each text bubble is considered a single message, so each message ranges from a single word to multiple sentences.

To observe the code-switching patterns in the data, we first analyzed the part of speech (POS) of each token using spaCy (Honnibal et al. 2020), a Python natural language processing library that performs tokenization and POS tagging. SpaCy has support for multiple languages and is an open-source tool, making it suitable for our Japanese-English bilingual data. We initially tried spaCy’s pre-trained pipelines for multilingual data but found the POS tagging accuracy lacking upon manual inspection. Instead, we conducted POS tagging on each message using both the English and the Japanese pre-trained models. Each token was identified as either English or Japanese first, and then the POS tag from the appropriate language was used as the POS tag for that token. The romanized Japanese tokens were manually identified from the data and were mapped to their equivalent in Japanese characters with a dictionary we created, then subsequently passed through the Japanese POS tagger. An overview of this process, including an example of the dictionary to map romanized Japanese words to Japanese characters, is illustrated in Figure 1.

Figure 1:

Part of speech tagging process of the Japanese-English code-switching dataset.

Tokenization was performed by spaCy to segment the sentence into words. An example of the result returned by spaCy’s token segmentation and POS tagging for one of the messages in our data is: [[‘最近’, ‘ADV’], [‘好き’, ‘ADJ’], [‘な’, ‘AUX’], [‘動画’, ‘NOUN’]]. Therefore, through the process outlined in Figure 1, we can count the number of English, Japanese, and romanized Japanese words and identify the POS of each word. The number of words of each language in the data can be seen in Table 3.

We further annotated the data by identifying whether the code-switched messages were intra-sentential or inter-sentential. This was done by manually going through the messages. Messages were marked as intra-sentential if two or more languages were used within the same sentence, as in (1a). Messages were marked as inter-sentential if two or more languages were used between sentences either in the same message, as in (1b), or in consecutive messages, as in (1c), where the symbol “//” indicates two separate messages. In the examples provided throughout this article, English translations of Japanese words, phrases, or sentences are included in parentheses.

For messages that were marked as inter-sentential code-switching, we further annotated the data by identifying the flow of languages used. For instance, example (1c) was labeled as “Eng to Jpn”, signifying that the speaker switched from using English to Japanese.

3 Descriptive results

In the dataset, including the direct message (DM) and group chat (GC) data, about 63.5 % of the data was in English, 35.9 % was in Japanese, and 0.612 % was in romanized Japanese, suggesting that the host language is English while the embedded language is Japanese. Since all three of these Gen Z bilinguals were born and raised in the US, it is expected that the primary language used would be English, but it is noticeable that there is a surprisingly high proportion of Japanese. Comparing the DM data and the GC data, the ratio of Japanese used in DMs (45.7 %) was much higher than the ratio of Japanese used in the GC (30 %). The difference in the ratio of Japanese messages across the DMs and GC could reflect the language backgrounds of the speakers and their language preferences at home in the respective chats: the messages in the DMs exchanged between speakers A and C have a much higher ratio of Japanese, which may come from speaker C primarily speaking Japanese at home and preferring to use Japanese when texting with friends from the same Japanese American community.

We also looked at the distribution of the parts of speech used within each language. Here, we looked at Japanese (written in the traditional Japanese writing systems) and romanized Japanese as two different categories. Table 4 presents the top 10 parts of speech for each language within each chat type, DMs and GC.

Table 4 indicates that nouns are consistently in the top five parts of speech, regardless of the language or chat type, echoing previous studies on Japanese-English code-switching (Azuma 1997). Verbs also tend to be in the top five parts of speech except in the romanized DM data. A notable difference between English and Japanese is that in both the DMs and the GC, the percentage of auxiliary verbs in Japanese (14.26 % in DMs and 13.74 % in GC) is about twice that of auxiliary verbs in English (7.18 % in DMs and 6.52 % in GC). Furthermore, pronoun is a commonly seen part of speech in English for both the DM and GC data; however, it does not appear in the top 10 parts of speech for Japanese, which may result from the fact that pronouns, in general, are much rarer in Japanese than in English.

We also conducted an analysis following Derrick’s (2015) sentence-by-sentence analysis to study the linguistic nature of the messages. We manually categorized the 1,198 messages into five categories: monolingual English messages, English-base bilingual messages, monolingual Japanese sentences, Japanese-base bilingual messages, and hybrid messages. Examples of each type of message are presented in (2). The frequency of each type of message in both the DM and the GC data is shown in Figures 2 and 3, with approximately 13 % of the messages being bilingual or hybrid messages. This result suggests a very common use of code-switching in CMC texting among these Gen Z Japanese Americans, since the percentage of bilingual or hybrid messages is much higher than reported in literature (Derrick 2015; Montes-Alcalá 2024; Shepherd 2021).

Figure 2:

Analysis of the linguistic nature of the direct messages.

Figure 3:

Analysis of the linguistic nature of the group chat.

Interestingly, the hybrid sentences have a much higher distribution in the GC than in the DMs. In contrast, the DMs have a much higher proportion of monolingual Japanese sentences and Japanese-base bilingual sentences. We conjecture that the more frequently a speaker orally code-switches at home, the more they will code-switch over text messaging with friends. Likewise, the more frequently a speaker uses an embedded language at home, the more likely they will use that embedded language over text messaging with friends who are in the same bilingual community. To further testify to these hypotheses, we further analyzed the linguistic nature of the messages by speaker.

Figures 4 and 5 show how many messages of each category (Monolingual English messages, English-base bilingual messages, Monolingual Japanese sentences, Japanese-base bilingual messages, and hybrid messages) each of the speakers sent.^[1] Figure 4 shows that speaker B (who frequently code-switches with a sibling at home) has more bilingual and hybrid messages, and speaker C (whose whole family speaks Japanese at home) sends the most monolingual Japanese messages to friends. In Figure 5, we observe a similar pattern: speaker C sends more monolingual Japanese sentences and tends to code-switch less than speaker A (who speaks both English and Japanese at home). Both Figures 4 and 5 confirm the conjectures: a high frequency of oral code-switching at home or a high frequency of using both languages at home is reflected in a high frequency of code-switching in text messages with friends, while a high frequency of speaking the embedded language (i.e., Japanese in this case) at home is reflected in a high frequency of texting in that embedded language with bilingual friends.

Figure 4:

Analysis of the linguistic nature of the group chat by speaker.

Figure 5:

Analysis of the linguistic nature of the direct messages by speaker.

4 Discussion

Our dataset shows that Japanese usage accounts for as much as 45.7 % of the messages in one of the text chats and 35.9 % of the overall dataset. The high proportion of Japanese usage in our bilingual text dataset is contrary to Shepherd’s (2021) observation that postwar second-generation Japanese Americans primarily use English with occasional Japanese. Our dataset thus suggests that Japanese-English bilinguals (in particular, those in Generation Z) switch between Japanese and English frequently in their text messages with their fellow Japanese-English bilinguals, despite English being the dominant language in their daily life.

4.2 “Texting with bilingual friends reflects our uses of two languages at home”

A closer look into the speakers’ language background and their code-switching frequency (Figures 2–5) reveals an interesting but barely reported code-switching pattern in the literature: a higher frequency of oral code-switching or using both languages at home is reflected in a higher frequency of code-switching in text messages with friends, while a higher frequency of speaking the embedded language at home is reflected in a higher frequency of texting in that embedded language with bilingual friends. Example (5) presents a message bubble among the three speakers and shows their preferences for bilingual messages or monolingual text.

(5)

Speaker A:	i think im going home tuesday and 火曜日までプロジェクトで忙しいかも (‘I’m busy with a project until Tuesday’)
Speaker B:	Sad. perhaps do something more up north towards end of the week? Only if you can tho!! 無理なしに (‘No pressure’)
Speaker A:	im not sure i’ll be able to gomennn (‘sorryyy’)
Speaker C:	日本帰ってきたら学校始まる前に会おうよ!はるかちゃんも一緒に! (‘Let’s hang out when we’re back from Japan before school starts! With Haruka too!’)

When zooming in on the non-monolingual messages, like that in (5), we observed slightly higher but very similar occurrences of inter-sentential code-switching (120 messages) than intra-sentential code-switching (103 messages). This is different from the pattern reported in Poplack (1980) and Koban (2013), which were based on oral speech data: there, intra-sentential code-switching was more common than inter-sentential code-switching among proficient Spanish-English bilingual speakers and among proficient Turkish-English bilingual speakers in New York. However, we do not have enough evidence to attribute the differences in inter-sentential and intra-sentential code-switching patterns to our dataset being CMC texting data and the data of Poplack (1980) and Koban (2013) being oral. A recent study by Gonzales and Tsang (2023), based on CMC text data, also found that the vast majority of code-switched utterances (99.5 %) in their Cantonese-English WhatsApp dataset were intra-sentential, similar to Poplack’s and Koban’s observations. Previous studies proposed that the distribution of different types of code-switching is related to language fluency or the medium of communication. Poplack (1980) and Lipski (1985) argue that more fluent bilinguals do more intra-sentential code-switching while less fluent bilinguals do more inter-sentential code-switching. Gonzales and Tsang (2023) argue that CMC texting makes bilinguals tend to text short messages or one-word messages, which consequently makes it very difficult to find inter-sentential code-switching, that is, two distinct sentences of different languages in one speech bubble. However, our data does not support either an “incompetence” account nor a “CMC constraint” account.

Our data suggests a slightly higher occurrence of inter-sentential code-switching in a CMC dataset for proficient bilinguals, and moreover, considering the data by speaker, speakers A and B have no strong preference between inter-sentential code-switching and intra-sentential code-switching, while speaker C strongly prefers inter-sentential code-switching (Figure 6). Speaker C’s proficiency in either language is no lower than that of the other two speakers, as all three speakers are simultaneous bilinguals and have all reported full command of both languages. The only difference among speakers is that speaker C is the only one who primarily uses Japanese at home with both parents and siblings. Therefore, higher occurrences of a certain type of code-switching may not come from language competency or the constraints of CMC as a medium, but rather the language preference among family members extends to CMC texting among bilingual friends.

Figure 6:

Type of code-switching by speaker.

The proposed explanation is also supported by the analysis of language flow. For messages that were marked as inter-sentential code-switching, we further annotated the data by identifying the flow of languages used (Tables 5 and 6). Speaker C, who primarily uses Japanese at home, uses both Japanese and English in daily life, and produces the most monolingual Japanese messages, no doubt has full command of Japanese. If speaker C preferred inter-sentential code-switching because of a lack of proficiency in English, then we would expect a much higher English-to-Japanese-switch in speaker C’s data. If speaker C was influenced by the CMC medium constraint, we would expect speaker C to have very few English to Japanese and Japanese to English inter-sentential code-switched sentences, since switching keyboards to type two sentences of different languages in one speech bubble is not very convenient when texting. However, neither the “incompetence” account nor the “CMC constraint” can explain the higher occurrence of inter-sentential code-switching for speaker C.

Table 5:

Inter-sentential language switch frequency in direct messages.

Direct messages	Speaker A	Speaker C	Total
Eng to Jpn	14	10	24
Jpn to Eng	15	13	28
Eng to Rom	2	1	3
Jpn to Eng to Jpn	0	1	1
Rom to Eng	1	1	2
Eng to Jpn to Eng	2	0	2
Rom to Jpn	1	0	1
Jpn to Eng to Jpn to Eng	1	0	1
Total	36	26	62

Table 6:

Inter-sentential language switch frequency in group chat.

Group chat	Speaker A	Speaker B	Speaker C	Total
Eng to Jpn	6	18	3	27
Eng to Jpn to Eng	2	3	0	5
Eng to Rom	2	0	0	2
Eng to Rom to Eng	1	0	0	1
Jpn to Eng	3	11	2	16
Jpn to Eng to Jpn	0	0	2	2
Rom to Eng	2	0	1	3
Total	16	32	8	56

5 Summary and future work

The dataset created and analyzed in this study is one of the first of its kind on Japanese-English code-switching in CMC, to the best of our knowledge. The findings from this dataset challenge commonly cited motivations and reasons for code-switching. We found that code-switching among the Japanese-English bilinguals is a result of their full mastery of both Japanese and English, not due to a lack of language proficiency. This study also highlighted rarely discussed patterns and features of written code-switching: the frequency at which the speakers orally code-switch and speak the embedded language (Japanese) at home is reflected in their code-switching patterns over text messages. Additionally, this study highlighted the development of an unique CMC style of Japanese-English code-switching, characterized by the use of romanized Japanese words, elongation of Japanese words using romanization, mixing writing scripts, and English words written in Japanese characters.

The present study has certain limitations that we aim to address in future research. In the present study, we removed emojis from the data analysis, as the processing tool we used could not handle emojis properly. However, emojis are widespread in CMC and have been argued to represent the morphosyntax of a lexically typed stem and to undergo linguistic changes (Storment 2024). In a future study, we would like to incorporate recent work on annotating emojis (Shoeb and de Melo 2021) to broaden our investigations into understanding the role of emojis in code-switching in CMC texting and the factors influencing speakers’ choice between English, Japanese, romanized Japanese, or English words in Japanese script, along with emojis. Another limitation of the present study is that the dataset is unbalanced. In the present study, the dataset was collected from female Gen Z bilinguals only and there is no data from male bilinguals nor from older bilinguals. However, studies have shown that female bilinguals have different code-switching patterns from male bilinguals (Gonzales and Tsang 2023). In future studies, we hope to expand the dataset to create a balanced and large-scale corpus of bilingual code-switching texting data so that we can analyze the sociopragmatic and communicative patterns of these data, and thoroughly investigate the potential factors that could influence the code-switching patterns in CMC texting. We also hope to expand the corpus by including both oral and text bilingual data from the same speakers, allowing us to observe differences in code-switching patterns in writing versus oral production.