Baxnet Ideas · Technical explainer
The Language Switch Is Part of the Message
TL;DR: Translating every message into one language can preserve what was said while erasing the moment somebody chose to say it differently.
A translated message can be accurate and still be incomplete.
In a multilingual chat, language choice is part of the conversation. A person may move into the language they use with family, switch back to the one they use at work, borrow a phrase that only sounds right among old friends, or quote somebody in the words they actually used.
Translate all of those messages into English and the practical meaning may survive. The change of voice can disappear.
Imagine a family chat that usually runs in English. One sibling sends an ordinary request in the language everyone grew up speaking at home. A translation might render it perfectly. The family may still hear something else in the switch: warmth, mock seriousness, impatience, or simply the sound of home.
There is no universal decoder for that moment. That is precisely why the original matters.
People do not switch only because a word is missing
Code-switching and code-mixing are sometimes treated as problems to clean up before software can understand a conversation. The assumption is that everybody was trying to write in one language and occasionally failed.
Real multilingual conversation is more interesting than that.
A 2023 study of Cantonese-English use on WhatsApp examined messages and survey responses from 24 Hong Kong university students. Participants did describe using another language to fill conceptual gaps. They also described switching to make conversation feel more casual and natural, strengthen solidarity, and express bilingual or Hongkonger identity.
It was a small, locally specific study, not a dictionary of what every language switch means. Its useful point is that language alternation can be social as well as lexical. Sometimes the switch is not repairing a sentence. It is helping to place that sentence inside a relationship.
The scale of the technical challenge is becoming clearer too. A Findings of ACL 2026 paper introduced a corpus of more than 355,000 user-generated WhatsApp and Telegram messages, focused mainly on English, Mandarin, and other languages. The messages are kept in conversational context rather than reduced to a bag of isolated mixed-language sentences.
That does not make the social meaning of each switch automatically knowable. It does make it harder to defend treating multilingual chat as untidy monolingual data.
Translation needs the conversation around it
Even literal translation benefits from context.
A separate 2026 study in Transactions of the Association for Computational Linguistics tested translation systems that could use preceding turns in a conversation. Across customer-support chats and user-assistant interactions, its context-aware approach improved translation quality and consistency over the systems used for comparison.
Those were task-oriented conversations, not private family or friendship chats. The study does not prove that a translation system can recover intimacy, irony, or identity from a language change. It establishes a more basic boundary: translating a conversational turn without its surroundings can lose information needed to translate the words themselves.
There is a further problem when an entire chat is normalised into one language. Even a good translation may hide that the original language changed.
The reader sees a smooth English transcript. They do not see that two people moved into a shared language when a parent joined, that one phrase repeated the wording of an old joke, or that a normally formal colleague answered in the language used outside work. The translated archive may be easier to search while becoming less faithful to the interaction.
The translation should remain a layer
A translated view of a chat should sit alongside the original, not silently replace it. The source message, its position in the exchange, and the fact that the language changed should remain inspectable.
Conversation analysis also needs restraint. A switch is evidence that something changed in the language of the exchange. It is not a reliable sentiment label. The same move can carry affection in one relationship, efficiency in another, and no special significance at all in a third. Topic, audience, habit, community, and the surrounding turns all matter.
For a product such as Mimoto, the design implication is fairly direct. If translated text is ever used to help someone search or understand a multilingual history, the original conversation should remain the source of record. An observation based on the translation should still lead back to the messages as they were sent.
Multilingual people are not producing a defective version of a single-language chat. They are using another part of conversation. Translation can help us understand the words. It should not make us forget that the language changed.
Further reading: The NUS ABC Codemixed Corpus, context-aware translation for conversations, and code-switching and identity on WhatsApp.