Building Multilingual NLP
English, Nepali, and Hinglish in one conversation — and why code-switching breaks naive systems.
Users in Nepal do not speak “one language per message.” They code-switch. A single line can mix English verbs, Nepali nouns, and Hindi-adjacent fillers. If your NLP pipeline assumes monolingual purity, it will misunderstand daily speech.
Hinglish is not a third official language. It is a usage pattern. Treating it as noise is how products alienate the exact audience that would love them most.
BolKharcha’s approach combines language hints, mode context, and fallbacks. When confidence is low, ask a clarifying question instead of committing a bad transaction. Silence is worse than a quick confirm.
Speech-to-text adds another layer. Numbers and brand names break often. Post-processing that normalizes currency words and local payment names recovered more accuracy than swapping providers every week.
I keep a living glossary: eSewa spellings, Khalti variants, common expense slang. Dictionaries are unglamorous and incredibly high leverage.
Testing multilingual UX means recruiting real speakers, not only bilingual engineers. What feels “obvious Nepali” to you may be unnatural to someone else.
If you build for South Asia, multilingual is not a v2 feature. It is the interface. Design the conversation for mixing from day one.