Code-Switching

The most complex form of code-switching, where the alternation occurs within a single sentence or clause. This requires the speaker to navigate two distinct grammatical systems simultaneously.

• Example: 'I went to the tienda to buy some leche for the café.' (English/Spanish)
• Violates the assumption of monolingual sentence structure in traditional NLP pipelines
• Demands models that can process mixed-language token sequences without segmenting by language first

Alternation that occurs at sentence boundaries, where a speaker completes a full sentence in one language before switching to another for the next sentence.

• Example: 'I finished the report yesterday. Mañana vamos a revisarlo juntos.'
• Easier for NLP systems to handle than intra-sentential switching
• Still requires cross-lingual coreference resolution to track entities across language boundaries

A technical approach where language identifiers are prepended to subword tokens to signal code-switching boundaries to the model.

• Example: <en> I bought <es> pan dulce <en> at the bakery
• Enables a single multilingual model to process mixed-language input without separate language detection
• Used in architectures like XLM-RoBERTa and mBERT to handle code-switched corpora

A foundational linguistic theory positing that in code-switched utterances, one language acts as the matrix language providing the syntactic frame, while the other is the embedded language contributing content morphemes.

• The matrix language determines word order and grammatical morphemes
• Critical for syntactic parsing of code-switched text
• Informs the design of grammar-aware multilingual models

A data augmentation technique that artificially creates code-switched training examples by aligning parallel corpora and substituting spans from one language into another.

• Uses bitext mining and word alignment to identify substitutable phrases
• Addresses the severe scarcity of naturally occurring code-switched training data
• Must respect equivalence constraint to avoid generating ungrammatical switches

Code-switched queries present a retrieval challenge where a user mixes languages in a single search string, requiring the search engine to understand cross-lingual intent without explicit translation.

• Example query: 'best restaurantes near me con outdoor seating'
• Demands multilingual dense retrieval models that encode mixed-language queries into a shared semantic space
• Hybrid systems may combine language identification with cross-lingual query expansion

The classification task of automatically detecting the natural language in which a given text is written. Code-switching breaks traditional language identification models that assume a single language per document or sentence. Sub-word level identification is required to segment a mixed-utterance like 'I'm going to la tienda for some leche' into its constituent languages before routing to downstream translation or embedding models.

The pre-processing step of converting text into a canonical script form. Code-switching often involves script-switching, such as typing Hindi words in Roman script (Hinglish: 'main kal market ja raha tha'). Transliteration modules must convert Romanized text back to its native script (Devanagari) before tokenization, or the model must be trained on massive Romanized corpora to handle the variation natively.

A pre-training objective where a model learns to predict randomly masked tokens in a concatenated stream of text from multiple languages. For code-switching, synthetic code-switched training data is generated by randomly swapping words with their translations. This forces the model to learn contextual representations that bridge languages, enabling it to predict a masked Spanish word using English context clues and vice versa.