What is Code-Switching?

This is not substandard speech. It is how most bilingual speakers talk, and it is especially common in Indian English, Singapore English and Filipino.

The hard part of this data is marking the switch points. The exact position where language A becomes language B. Different annotators will place it several words apart.

Varieties of code-switching

The main split is where the switch happens. In inter-sentential switching the change comes at a sentence or clause boundary: one sentence in one language, the next in the other, each grammatically complete. In intra-sentential switching the change happens inside a single sentence. The second is the hard case, because the model cannot use sentence structure to guess which language comes next.

The two languages are rarely equal partners. In most real speech one carries the grammar, the sentence structure, the verb endings and the function words, while the other supplies content: technical terms, numbers, or vocabulary that does not exist in the first language. Which one is which changes the product, and the same pair of languages produces different data depending on the direction of the mix.

Two neighbouring phenomena are often folded in and should not be. Borrowing is a word that has been adopted into the language and is used by monolingual speakers too; it is not a switch. Transliteration is writing one language in another language's script, which is a writing decision rather than something the speaker did. Keeping them apart matters, because each one changes the annotation rule.

What to specify when buying

Name the pair and the direction. Hinglish is not one product: the mix ratio, the topic and which language carries the grammar all vary, and a set collected from casual conversation about daily life is different from one collected from technical discussion. Say what register the target system will meet, and whether the mix should be heavy or occasional.

The switch-point convention is where the work is. The guideline has to fix where a boundary is placed, at the word or at the point where the grammar changes, because annotators naturally differ by several words on the same stretch of audio. It also has to fix how the switched words are written: in the script of the language they belong to, or romanised throughout.

Then the collection method, because code-switching cannot be read from a script. Handing a bilingual speaker prepared sentences produces careful, unnatural mixing. Natural switching comes from conversation, which is slower to collect and harder to control. The alternative is elicited switching, where a speaker is asked to switch on cue; it is more controlled and less natural. The brief should say which one the model is trained for.