Beyond Translation: The Cultural Nuances of Code-Switching

When I switch from English to Pidgin mid-sentence, I'm reaching for a feeling that comes more naturally in Pidgin.

The identity dimension

On social media, I notice influencers using this choice to shape their audiences. A switch can make an explanation easier to follow or invite listeners into a shared way of speaking. It can also give a joke its familiarity or make a rebuke more pointed. At work, a change of language may signal a change in formality.

Allan Bell's account of audience design relates speech style, including bilingual language choice, to the audience a speaker addresses. It gives me a way to think about the identity dimension of switching: speakers position themselves in relation to other people. They can invoke familiarity or establish distance, and those positions can change within an exchange.

The social meaning of a language choice varies with the relationship and setting. A familiar expression may establish solidarity in one relationship and be unwelcome in another. Interpreting a switch therefore involves asking what relationship the speaker is invoking and whether the listener shares that understanding.

Why NLP struggles

An influencer narrating an altercation with a cashier described it this way:

"I approached the cashier and she said, 'Go back and join the queue.' In my mind I'm like, 'u dey whyne'. Go to which back?"

The narration separates the cashier's reported instruction from the narrator's internal response. The shift into Pidgin helps mark that change of perspective. An English rendering such as "Are you kidding me?" conveys disbelief; in this account, the reaction also expresses confrontation because the narrator feels disrespected. The influencer's later telling gives the online audience access to that reaction. We are interpreting a representation of an encounter, with the narrator's perspective built into it.

John Gumperz's work on contextualization treats language choice as one of the cues through which listeners infer how speech should be understood. For NLP, this creates a distinction between translating an utterance's wording and identifying its pragmatic force: what it accomplishes in an exchange. A fluent translation can leave unresolved whether a remark is teasing, challenging someone or recounting an unspoken objection.

The account tells us how the narrator experienced the instruction, without establishing the cashier's intentions. A system should preserve that distinction when explaining the exchange.

Building more inclusive models

In my work at Sussex on evaluation frameworks for code-switched environments, I want to examine how models use conversational evidence to interpret social meaning. LinCE provides task-based evaluation for code-switching, including language identification and sentiment analysis. Assessing pragmatic interpretation requires an explicit account of whose attitude is represented and what evidence supports a proposed reading.

I would evaluate related utterances across different relationships and settings, checking whether a model changes its interpretation when the context warrants it. Where feasible, holding the wording constant would help isolate the effect of context. The task would require the model to explain which parts of the exchange support its judgment. Scoring could assess the interpretation and whether the cited conversational evidence supports it.

Fluent speakers should help construct these cases and assess the answers. Their disagreements should remain visible in the evaluation, since an expression may be understood differently across communities or relationships. When the supplied context is insufficient, a model's recognition of uncertainty should count toward a successful answer. Reference answers could record the acceptable interpretations and the contextual evidence for each.