Source-linked AI summary
A Survey of Code-switching: Linguistic and Social Perspectives for Language Technologies
A. Seza Doğruöz, Sunayana Sitaram, Barbara E. Bullock, Almeida Jacqueline Toribio
TL;DR
Research on code-switching (C-S) needs broader attention to linguistic and social dimensions, as well as better representation across language pairs and C-S types. This paper surveys linguistic and social studies in European and Indian contexts alongside language-technology issues. It concludes that current models and benchmarks do not represent the full spectrum of C-S, while context-specific analysis matters because speakers may not switch in every multilingual setting.
Problem
Existing C-S models and benchmarks do not represent the full range of language pairs, C-S types, and applications, leaving an incomplete picture of model capabilities.
Source: §7.1 · p. 7, §7.2 · p. 7Method
The paper surveys linguistic and social studies of C-S in European and Indian contexts and relates them to computational language-technology issues.
Source: §1 · p. 1Results
Current benchmarks do not cover the full spectrum of C-S, and state-of-the-art models are only slightly better than chance on NLI.
Source: §7.2 · p. 7Takeaways & Limitations
Systems for arbitrary language pairs may be of little use to targeted users without systematic analysis of C-S’s linguistic and social aspects in context.
Source: §8 · p. 8C-S cannot be assumed across all multilingual contexts; historical, social, and cultural factors may lead speakers not to switch.
Source: §8 · p. 8
Abstract
from arXiv · showhide
The analysis of data in which multiple languages are represented has gained popularity among computational linguists in recent years. So far, much of this research focuses mainly on the improvement of computational methods and largely ignores linguistic and social aspects of C-S discussed across a wide range of languages within the long-established literature in linguistics. To fill this gap, we offer a survey of code-switching (C-S) covering the literature in linguistics with a reflection on the key issues in language technologies. From the linguistic perspective, we provide an overview of structural and functional patterns of C-S focusing on the literature from European and Indian contexts as highly multilingual areas. From the language technologies perspective, we discuss how massive language models fail to represent diverse C-S types due to lack of appropriate training data, lack of robust evaluation benchmarks for C-S (across multilingual situations and types of C-S) and lack of end-to-end systems that cover sociolinguistic aspects of C-S as well. Our survey will be a step towards an outcome of mutual benefit for computational scientists and linguists with a shared interest in multilingualism and C-S.
1 Introduction
Code-switching (C-S) involves speakers alternating between languages, sometimes within a word or phrase. This survey aims to connect computational research with linguistic and social accounts of C-S in multilingual contexts.
1 Introduction
- A bilingual child’s German-Turkish example attaches a Turkish possessive marker to a German noun.The example describes the child’s teacher and gives the marker as -si.Source: §1 · p. 1
- Speakers in multilingual communities switch between languages in both speech and writing.The passage presents this as a common practice.Source: §1 · p. 1
- The survey aims to inform computational linguists about linguistic and social aspects of C-S in multilingual contexts, including Europe and India.It seeks to bridge the fields so their work can be mutually reinforcing.Source: §1 · p. 1
2 Competing models of C-S
The section compares grammatical accounts of code-switching (C-S), including the Matrix Language Frame and constraint-based alternatives. It also notes limits in applying these models and distinguishing C-S from borrowing.
2 Competing models of C-S
- A single Spanish word in an English utterance may be borrowing rather than C-S, unlike more syntactically integrated language alternation.The section illustrates the single-word case with “baile.”Source: §2 · p. 1
- The Matrix Language Frame (MLF), adapted by Myers-Scotton (1997) from Joshi (1982), treats one language as the clause’s grammatical frame for material from another.Its scope is the clause, so it does not cover alternational C-S where separate clauses use separate languages.Source: §2 · pp. 1–2
- The MLF is contested: matrix-language criteria are ambiguous, and Bullock et al. (2018) found different criteria reliably converge only for simple insertions.The passage also notes that consistent identification is not always possible and that the model’s scope is limited.Source: §2 · p. 2
- Alternative accounts include the Equivalence Constraint, Government Constraint, Functional Head Constraint, and constraint-free theories.Incorporating the Functional Head Constraint in NLP experiments has improved computational and speech-model accuracy, while testing constraint-free theories requires rare parsed mixed-language corpora.Source: §2 · p. 2
3 Why do speakers code-switch?
Speakers’ reasons for switching can affect the structure and placement of code-switches. The section describes pragmatic, interactional, cognitive, and social factors that shape language choice.
3 Why do speakers code-switch?
- Pragmatic and interactional functions can condition C-S patterns, including signalling identity, quoting, gaining attention, or softening and intensifying a message.Some functions are associated with constructions such as clause-edge tags, verbs of saying, vocatives, and sequential translation equivalents.Source: §3 · p. 2
- C-S choices can reflect interactional factors, politeness, prestige, face-saving, and political positioning.Clyne (1991) identifies factors including topic, interlocutors, role relationship, and communication channel; Heller (1992) describes French-English C-S in Quebec as potentially signalling political choice.Source: §3 · p. 2
- C-S patterns can differ across generations and speakers, and computational studies suggest switching is speaker-dependent.The passage also relates C-S practices to gender and identity and describes work predicting Turkish-Dutch choices in computer-mediated interaction.Source: §3 · pp. 2–3
4 Code-switching, Borrowing, Transfer, Loan Translation
The section distinguishes code-switching from borrowing, transfer, and loan translation, noting that structurally similar mixing can have different analyses. It also states that the extent to which transfer and loan translation co-occur with C-S is unknown.
4 Code-switching, Borrowing, Transfer, Loan Translation
- A lone-word insertion may be a nonce borrowing rather than C-S, while structurally similar mixing can also represent a community’s single code.Auer (1999) distinguishes prototypical pragmatic C-S, Language Mixing without a pragmatic purpose, and Mixed Languages.Source: §4 · p. 3
- Loan translation and transfer involve words from one language combined with semantics or grammatical constructions from another.The passage illustrates loan translation with a Turkish verb taking on the Dutch meaning of “do.”Source: §4 · p. 3
- Transfer can involve borrowing a grammatical construction without borrowing words, as in French constructions translated from Dutch or English equivalents.The examples include Brussels French chercher après and Ontario French chercher pour for “look for”; how often transfer and loan translation occur alongside C-S is unknown.Source: §4 · p. 3
5 C-S across Languages: European Context
European code-switching (C-S) research spans diverse language-use settings, from children’s multilingual development and education to immigrant communities, media, and historical records. It also covers switching between standard languages, dialects, and minority varieties.
5 C-S across Languages: European Context
- European multilingualism includes simultaneous and sequential language acquisition, as well as switching between local or minoritized varieties and dominant languages.Multilingual communication also occurs at the institutional level.Source: §5 · pp. 3–4
- Studies examine C-S in bilingual and multilingual children’s speech, parent-child interaction, and educational settings, including its use in resolving disputes.The research spans multiple language pairings and countries.Source: §5 · p. 3
- European research documents C-S between standard languages, dialects, and minority languages, including in computer-mediated communication.Examples include Greek and Cypriot Greek, Welsh and English, and Swiss-German dialects.Source: §5 · pp. 3–4
- Immigrant-community studies cover many language pairs across Europe and examine both spoken interaction and computer-mediated communication.The surveyed work also includes historical documents and C-S in written media and advertising.Source: §5 · p. 4
6 C-S across Languages: Indian Context
The Indian context combines widespread multilingualism with a large number of languages and varied historical, social, and cultural influences on C-S. Research covers everyday interaction, media, education, and health, while systematic comparisons and data collection remain challenging in some settings.
6 C-S across Languages: Indian Context
- India’s census records 121 major languages and 1,599 other languages, while excluding languages with fewer than 10,000 speakers.The passage reports that 26% of the population is bilingual and 7% trilingual.Source: §6 · p. 4
- Studies describe C-S in everyday Indian settings as shaped by social and functional contexts, including market interactions and long-standing language contact.In one documented setting, fluent Konkani-Kannada bilinguals did not switch between the languages, a pattern attributed to Konkani’s higher prestige.Source: §6 · p. 5
- Media studies report changing language patterns in film scripts and examine C-S in television, radio, and advertising.For Bollywood films, Hindi’s earlier dominance with occasional English switches gave way to greater English dominance in later productions.Source: §6 · pp. 5–6
- Research on education documents C-S in Indian schools and immigrant settings, while systematic comparisons of Bengali-English C-S across India, Bangladesh, and immigrant communities are lacking.The passage also describes generational shifts toward English in the UK Bengali community.Source: §6 · p. 6
- Studies of bilingual aphasia and mild dementia report differences in C-S amount or frequency, but both studies had limited subjects.The passage presents these findings as insights from health settings.Source: §6 · p. 6
7 Computational Approaches to C-S
The survey reviews computational work on C-S across tasks and language pairs, but finds that current technologies still do not process C-S as seamlessly as monolingual data. It identifies limitations in training data, evaluation coverage, and user-facing systems.
7 Computational Approaches to C-S
- Computational C-S research spans tasks including language identification, POS tagging, sentiment analysis, and speech recognition, but systems do not yet handle C-S as seamlessly as monolingual data.The survey identifies data, evaluation, and user-facing applications as three main limitations.Source: §7 · p. 6
7.1 Data
- Multilingual language models may miss C-S during training or lack coverage of language pairs, types, and functions encountered in daily life but absent from web data.The passage notes that mBERT is trained primarily on monolingual Wikipedia and XLM-R on Common Crawl.Source: §7.1 · pp. 6–7
7.2 Evaluation
- GLUECoS and LINCE introduced benchmarks spanning multiple C-S tasks and language pairs, but they do not represent the full range of applications and C-S varieties.The passage describes GLUECoS as covering Spanish-English and Hindi-English, and LINCE as covering four language pairs.Source: §7.2 · p. 7
- Current C-S models perform well on LID, POS tagging, and NER but are only slightly better than chance on NLI, and existing benchmarks cover only part of C-S.Many benchmark tasks use annotated tweets, limiting the picture of model performance across C-S types and applications.Source: §7.2 · p. 7
7.3 User-facing applications
- The survey identifies a lack of end-to-end user-facing applications for multilingual communities, and notes that current approaches usually omit social and linguistic factors influencing when and why people switch.Grammatical C-S models do not necessarily generate utterances that bilingual speakers would produce; bilingual chatbot users preferred bots that mimicked their own C-S patterns.Source: §7.3 · p. 7
8 Conclusion
The survey finds both shared themes and context-specific differences in European and Indian C-S research, alongside gaps in comparisons and computational resources. It argues for attention to negative evidence and social context when developing C-S technologies.
8 Conclusion
- European and Indian C-S research shares themes but differs with social, cultural, and historical context, and systematic comparisons across some language pairs are lacking.The conclusion notes gaps for newer immigrant and minority languages in Europe and less-studied pairs in India.Source: §8 · p. 8
- The conclusion cautions against assuming C-S occurs in every multilingual context, citing a Konkani-Kannada setting where speakers preferred not to switch.The preference is attributed to historical, social, and cultural factors in that setting.Source: §8 · p. 8
- Developing C-S systems for a language pair without systematic analysis of its linguistic and social context may not be useful to targeted users or language technologists.The conclusion connects this concern to the need for negative as well as positive evidence.Source: §8 · p. 8
- The survey aims to connect computational linguistics and linguistics, whose publishing traditions and citation practices can make relevant research less visible across fields.The conclusion says improved visibility could save time and resources for computational research.Source: §8 · p. 8