Beyond Monolingual Assumptions: A Survey of Code-Switched NLP in the Era of Large Language Models across Modalities

1 Introduction

• Code-switching challenges NLP systems built on monolingual assumptions, and multilingual models still struggle with real-world transfer and low-resource settings.The survey treats switching alongside script alternation, orthographic variation, and spontaneous conversational behavior.Source: §1 · pp. 1–2
• LLM-era code-switching research builds on successive eras of statistical, representation-learning, contextual, and foundation models.The survey notes that multilingual pretraining alone proved insufficient for robust code-switching modeling.Source: §1 · p. 1
• The paper positions code-switching as a core modeling challenge and calls for structured data, human-in-the-loop annotation, and language- and script-aware architectures.The proposed roadmap also emphasizes inclusive datasets, equitable models, and fair metrics across dialogue, speech, and multimodal contexts.Source: §1 · p. 2
• The survey analyzes 327 studies across 15 NLP tasks, 30+ datasets, and 80+ languages, and organizes research by architecture, training, and evaluation.It also identifies gaps in low-resource coverage, script diversity, cross-lingual transfer, and unified evaluation frameworks.Source: §1 · p. 2

2 Pre-LLM-Era works

• Early code-switched NLP used rule-based and statistical methods for language identification, POS tagging, NER, and sentiment analysis.Approaches included CRFs and CNNs for LID, CNNs with n-grams for POS, character-level RNNs and SVMs for NER, and SVM sentiment classification.Source: §2 · pp. 2–3
• BiLSTM-CRF models with embeddings later improved LID and NER, while switch-point sampling enhanced LID performance.The passage reports reduced perplexity for the BiLSTM-CRF models.Source: §2 · p. 3
• Task- and language-specific designs, shallow features, scarce labeled data, and poor cross-linguistic transfer limited early approaches and kept solutions fragmented.The fragmented work lacked shared representations and unified frameworks across diverse code-switching contexts.Source: §2 · p. 3

3.1 Traditional Tasks

• Traditional tasks gained methods including fine-grained LID, contextual POS tagging, synthetic-data NER, and prompt-based approaches, but multilingual models still underperform on code-mixed inputs.The survey links underperformance primarily to limited representation of code-switching in pretraining, with greater challenges in low-resource and typologically distant languages.Source: §3.1 · p. 3

3.2 Emerging Contemporary Tasks

• Emerging NLI, QA, intent, and dialogue tasks show advances, but fragmented semantic evidence, cultural ambiguity, and weak discourse grounding remain challenges.Generative models also struggle to maintain discourse coherence and stable switching patterns; text generation faces scarcity for typologically distant and non-Latin-script languages.Source: §3.2 · pp. 3–4

3.3 Underexplored Frontiers Tasks

• Conversational, speech, reasoning, safety, and visual-processing applications remain underexplored, despite emerging gains in dialogue systems and document processing.Code generation from mixed prompts achieves only moderate functional correctness, and reasoning tasks struggle with logical complexity and semantic drift.Source: §3.3 · p. 4

4.1 Datasets

• CSW datasets span broader language and modality coverage, including SwitchLingua with 420k samples and over 80 hours of audio across 12 languages and 63 ethnic groups.MEGAVERSE covers 22 datasets in 83 languages for multimodal evaluation, while several resources target underrepresented language pairs.Source: §4.1 · p. 5
• LLM-driven dataset expansion raises authenticity concerns: human evaluations report only 60–65% acceptability, and high-quality resources for underrepresented languages still rely on expert curation and community involvement.The survey describes trade-offs between expanded multilingual coverage and natural code-switching authenticity.Source: §4.1 · pp. 4–5

4.2 Frameworks and Toolkits

• Annotation and generation tools include speech annotation, LLM-assisted text annotation, translation refinement, video annotation, and linguistically grounded synthetic text generation.Examples include CoSSAT, COMMENTATOR, CHAI, ToxVidLM, GCM, and CodemixedNLP.Source: §4.2 · p. 5
• LLM-augmented resources can expand low-resource CSW data and improve performance, while human-in-the-loop annotation can support more authentic and equitable benchmarks.The survey cautions that synthetic data may lack naturalness and cultural nuance, introducing biases.Source: §4.2 · pp. 5–6

5.1 Mainstream Pre-training Approaches

• Specialized pre-training on real code-switched corpora often outperforms multilingual baselines, while task-adaptive strategies encode switching structure for downstream tasks.Examples include boundary-aware masked language modeling, model merging, alignment-based methods, and synthetic augmentation.Source: §5.1 · p. 6

5.2 Mainstream Fine-tuning Approaches

• Fine-tuning can improve in-domain performance but depends on curated data and may generalize poorly to unseen language pairs or introduce negative transfer.The surveyed work includes task-specific and multitask methods across LID, translation, parsing, NLI, QA, and sentiment analysis.Source: §5.2 · p. 6

5.3 Post-training Approaches

• Post-training enables rapid adaptation with little labeled data, but results vary across language pairs and tasks, and prompting or retrieval struggles with informal mixing and low-resource pairs.Zero-shot outcomes can also be sensitive to pretraining language composition.Source: §5.3 · pp. 6–7

6 Evaluation & Benchmarking

• Code-switching benchmarks have expanded from narrow task evaluations to frameworks covering switching patterns, cross-language performance, and contextual understanding.The survey covers benchmarks across text, speech, and multimodal tasks.Source: §6 · p. 7
• Evaluating code-switching systems requires standard performance measures, code-switching-specific metrics, and human-centric evaluation.Source: §6 · p. 7
• The survey directs readers to its appendices for comprehensive benchmark details and descriptions of evaluation methods and metrics.Source: §6 · p. 7

7.1 Speech Processing

• Speech research spans translation, ASR adaptation, and audio-visual recognition, while limited data and switch-point errors remain challenges.Monolingual fine-tuning can outperform multilingual baselines, but may degrade performance in the non-target language.Source: §7.1 · p. 7

7.2 Vision-Language Processing

• Vision-language work covers Hinglish visual question answering, harmful-meme detection, and code-switched image-text retrieval, with domain-, language-, and region-aware adaptation still needed.Source: §7.2 · pp. 7–8

7.3 Cross-Modal Integration

• 8–10% ASR error reductions are reported for Wav2Vec2 fusion in high-resource switching scenarios, and multimodal fusion combines audio, visual, and textual signals.The survey identifies scalable, phonetic-aware multimodal pretraining as a future need.Source: §7.3 · p. 8

8 Open Problems and Future Directions

• More than 72% of speech datasets and 92% of social-media datasets involve English, while regional biases and inefficient tokenization further constrain coverage and generalization.Source: §8 · p. 8

Model Architecture and performance gaps

• LLMs can confuse languages and show sharp accuracy drops in zero-shot code-switching transfer, despite their ability to generate synthetic code-mixed data.Source: Model Architecture and performance gaps · p. 8
• LLM-based evaluators can overestimate performance against human judgments, while common metrics may miss linguistic diversity or cross-lingual equivalence in mixed contexts.Source: Model Architecture and performance gaps · p. 8
• Potential applications include multilingual public services, education, healthcare, and digital preservation, with examples in customer service and language learning.Source: Model Architecture and performance gaps · p. 8
• Research and datasets favor high-resource pairs, while less common pairs and trilingual switching remain understudied and under-resourced.The survey reports mid-90s F1 for English-Spanish and English-Hindi systems, versus substantially lower accuracy for less common pairs.Source: Model Architecture and performance gaps · pp. 8–9

Future Directions

• Large-scale conversational datasets with naturalistic code-switching remain scarce; existing multimodal, multilingual, and audio resources have diversity or spontaneity limits.Source: Future Directions · p. 9
• Future systems should jointly model text, speech, and vision, while evaluation frameworks should assess switch-point accuracy, semantic consistency, fluency, sociolinguistic appropriateness, and regional variation.Source: Future Directions · p. 9

Ethics and Safety in Multilingual Contexts

• Multilingual safety alignment gaps disproportionately affect low-resource and marginalized communities, and failures persist on unseen language mixtures.The survey calls for bias-aware training, fairness-sensitive evaluation, and participatory data curation.Source: Ethics and Safety in Multilingual… · p. 9

• CSW research has shifted from task-specific statistical methods to unified multilingual and instruction-based frameworks.Source: §9 · p. 9
• LLM gains remain largely confined to high-resource language pairs, while spontaneous mixing, reasoning, and sociolinguistic variation remain difficult in low-resource settings.Source: §9 · p. 9
• Limited dataset coverage and weak CSW-aware evaluation frameworks amplify these challenges; the survey calls for targeted data curation, linguistically informed architectures, and real-world evaluation protocols.Source: §9 · p. 9

• The survey may miss indigenous or minority code-mixed languages because its coverage highlights widely studied language pairs.Source: Limitations · p. 10
• Some surveyed approaches and benchmarks may soon be outdated as LLM research evolves rapidly.Source: Limitations · p. 10
• Speech and multimodal research lags behind text-based NLU, so the taxonomy covers more text-based NLU.Source: Limitations · p. 10
• The survey focuses mainly on academic progress, giving less attention to ethical, computational, and accessibility concerns in deployment.Source: Limitations · p. 10
• Quantitative claims are representative values from cited primary studies, not aggregated or meta-analyzed statistics.Source: Limitations · p. 10

Ethics Statement

• The review analyzes published research and public datasets without collecting or analyzing human or user data.Source: Ethics Statement · p. 10

A Methodology

• Searches across ACL Anthology, arXiv, and Semantic Scholar, followed by screening and categorization, yielded 327 papers across 15 tasks, 30+ datasets, and 80+ languages.Source: A Methodology · p. 30

B Taxonomy

• The taxonomy organizes CSW research by task maturity, methodological evolution, language-pair coverage, and performance gaps as task complexity rises.Source: B Taxonomy · p. 30

B.1 Code-Switching Task Landscape

• The task landscape ranges from foundational analysis such as language identification, POS tagging, and NER to underexplored safety, multimodal, and reasoning tasks.Source: §B.1 · pp. 30, 32

B.2 Datasets and Resources

• The resource taxonomy includes low-resource and multilingual datasets, synthetic data, and frameworks for annotation and data generation.Source: §B.2 · pp. 32–33

B.3 Model Training and Adaptation

• Model adaptation spans code-mixed pretraining, task-adaptive objectives, cross-lingual alignment, task- or multi-task fine-tuning, and instruction tuning.Source: §B.3 · p. 33

Post-training and Inference-time Adaptation

• Post-training and inference-time strategies include zero-, one-, and few-shot learning and example-based prompting under minimal supervision.Source: Post-training and Inference-time… · p. 33

B.4 Evaluation and Benchmarking

• Evaluation covers comprehensive and domain-specific benchmarks, traditional and CSW-specific metrics, human judgments, and intrinsic quality assessment.Source: §B.4 · p. 33

C.1 Traditional Tasks

• Traditional CSW tasks have moved from statistical and feature-based methods toward transformer models, with persistent challenges in low-resource settings, script variation, and syntactic alignment.Work spans language identification, sentiment, syntax, and machine translation.Source: §C.1 · pp. 34–36

C.2 Emerging Contemporary Tasks

• Emerging work covers code-mixed generation, summarization, cross-lingual transfer, and transliteration, but evaluation, semantic alignment, and low-resource transliteration remain difficult.Source: §C.2 · pp. 36–37

D Pre-training Approaches

• Pretraining approaches use naturally occurring or synthetic code-switching, alignment, and token-replacement strategies to support downstream transfer and task performance.Source: D Pre-training Approaches · p. 37

E Fine-tuning Approaches

• Fine-tuning includes instruction tuning and parameter-efficient methods, which support CSW adaptation but can rely on generated data or underperform on divergent and transliterated pairs.Source: E Fine-tuning Approaches · pp. 37–38

Reinforcement Learning for CSW Adaptation

• Reinforcement-learning approaches use AI feedback and preference optimization for CSW generation and translation, though computational demands and reliance on broader RLHF remain concerns.Source: Reinforcement Learning for CSW… · p. 38

F.1 Benchmarks

• CSW benchmarks have expanded in language and task coverage, but are often better suited to classification and retrieval than complex reasoning, multimodal interaction, or long-form generation.Source: §F.1 · p. 38

F.2 Evaluation Metrics

• CSW-specific metrics complement conventional scores, yet existing measures still miss discourse consistency, semantic adequacy, and natural code-switching in generation.Source: §F.2 · pp. 38–39

G Future Directions

• Transfer remains weak for complex or geographically varied CSW, while models also overlook speaker identity, discourse function, cultural norms, and pragmatic intent.Source: G Future Directions · p. 39, Sociolinguistic and Pragmatic… · p. 39

H Supplementary Material

• Figures 3–5 illustrate failures on code-mixed queries: ChatGPT returns Indian states instead of vacation countries, GLM-4.6 misreads a Bengali sandhi query, and Perplexity ignores a constraint about destroyed Goan temples.The examples also include incorrect grammar in ChatGPT’s response.Source: H Supplementary Material · pp. 39–40, Figure 3 · p. 40, Figure 4 · p. 40
• 40% of analyzed collections are “Large” datasets (50–100k instances), while COMI-LINGUA (125,615) and MEGAVERSE (100,000) are the largest prominent datasets.Figure 6 pairs a comparison of prominent dataset instance counts with an overall size-category distribution.Source: H Supplementary Material · p. 40, Figure 6 · p. 41
• Hinglish and Spanish–English combinations strongly dominate the language-pair dataset and benchmark counts, indicating a research focus on these pairs relative to others.Figure 7 covers the top 15 language pairs.Source: H Supplementary Material · p. 40, Figure 7 · p. 42
• Code-switched NLP tasks are organized by linguistic level and modality, with research concentrated on lower-level tasks such as language identification and sentiment analysis.The task distribution also reports sentiment analysis at 26.2% and machine translation at 18.5%; “Other” and “Understanding” categories account for 55.9% and 31.1%, respectively.Source: H Supplementary Material · pp. 41–42, Figure 8 · p. 42
• Tables 1–5 cover selected code-switching datasets, models, benchmarks, representative datasets, and speech datasets.The model and dataset selections use citation counts and community adoption, while the resources span tasks, languages, data sources, and real or synthetic speech data.Source: Table 1 · p. 44, Table 2 · p. 45, Table 3 · p. 45
• Tables 6–8 summarize representative LLM, speech, multimodal, and evaluation-metric failure modes in code-switching settings.The tables highlight reasoning, safety alignment, mixed-language understanding, cross-modal grounding, acoustic disambiguation, transliteration variation, and ungrammatical mixing.Source: Table 6 · p. 47, Table 7 · p. 48, Table 8 · p. 48