Multilingual Chatbots

The problem

A chatbot trained in English and translated to French does not understand how French speakers phrase their questions. It understands French translations of English questions, which is a different, more constrained problem.

Real user input does not follow translation patterns. A French speaker asking “je ne retrouve plus ma facture” is not translating “I can’t find my invoice”; they are expressing the question in a naturally French way. A bot trained on translated English utterances will not reliably match this to the billing intent, because the phrasing is too different from what the translated training examples looked like.

The same applies to response templates. A translated response in French reads like a translated response in French. Native speakers notice this: it is the same register problem that applies to all machine translation, except in a conversational context where unnatural phrasing is more conspicuous because the user expects a natural conversation.

How we approach it

We localize at the data level, which means starting earlier in the process than most chatbot localization projects begin.

Intent examples in each language are written by native speakers who understand how users in that market ask questions naturally. We do not start from English utterances and translate them; we start from the intent definition and generate locale-native examples. The bot trains on language-natural data, not translation artifacts.

Entity definitions are extended per locale. Date formats, address structures, name ordering, and locale-specific variations all need to be represented in the entity definitions. A bot that can recognize “25 avril 2026” and “25/04/2026” as the same date entity, and “Herr Schmidt” and “Schmidt” as the same name entity in German context, is more robust than one trained on the English patterns transposed.

We also review the intent taxonomy before localization begins. Some intents map cleanly across languages; others reflect English-language communication patterns that do not transfer directly. When we find intents that will need restructuring for specific locales, we flag them before data generation. Restructuring after the training data is built costs significantly more.

Response templates are written per locale, not translated. The goal is a response that sounds like it was authored by a native speaker for a native-speaking user, not like it was composed in English and passed through translation.