Twi-English maternal-health dataset: What Mixed Questions Reveal About AI Understanding

A 20,000-pair Twi-English maternal-health dataset matters because it gives researchers and builders real mixed-language questions to study, test, and improve against. It treats the way people actually speak, moving between Twi and English within one thought, as useful language data rather than an error to remove.

The sentence a benchmark can miss

Consider Adwoa, an illustrative composite, standing in her kitchen in Kumasi with a phone pressed between her shoulder and ear. Her aunt has sent a voice note about a pregnant relative who feels “dizzy paa” after taking a medicine, then asks in English whether the clinic should be called now.

Adwoa understands every word. A system trained to expect clean language boundaries may not.

The risk is not a mildly awkward reply. If the system drops the Twi intensifier, treats the English medical term as unrelated, or loses who took the medicine, the family could leave with an answer that sounds confident while missing the urgency of the question. The clinic call might be delayed because the message was misunderstood at the first step.

A larger benchmark score can tell us something about performance on a test. It cannot, by itself, show whether the test contains the mixed sentences people use when they are worried, tired, speaking quickly, or passing on a family voice note. A dataset built around Twi-English maternal-health questions makes that gap visible.

Code-switching carries meaning

Code-switching is often treated as a complication: separate the languages, translate one into the other, then continue. That approach can flatten the message before the system has understood it.

In ordinary Ghanaian conversation, the switch can signal emphasis, familiarity, precision, or the easiest available word. A speaker may use Twi for the feeling of a symptom and English for a term heard at a clinic. They may ask a question in English, then add a short Twi phrase that changes how urgent it feels.

The dataset’s value begins there. With 20,000 Q&A pairs, developers have more than a handful of polished examples. They have material for checking where a model loses the thread, where it over-translates, and where a response needs to preserve the speaker’s original meaning.

That matters especially in maternal health, where a vague answer can create false reassurance. The dataset does not make any AI medically reliable, and it does not replace qualified care. It gives teams a clearer way to test whether a system understands the question before it attempts to answer it.

For a closer look at why mixed sentences deserve their own evaluation, read what happens when a voice assistant loses the thread between Twi and English.

Better data changes what builders can measure

Benchmarks still have a role. They can expose progress and regressions. The important question is what they ask a model to do.

A benchmark made from mostly single-language prompts may reward a system that handles Twi in isolation and English in isolation. That same system can still stumble when a speaker says both in one breath. If code-switching appears only as an edge case, teams can miss the failure until someone tries to use the product in a real conversation.

A 20,000-pair dataset creates a different discipline. Builders can test whether the system keeps the subject, symptom, timeframe, and question intact across a language switch. They can look for patterns in failures rather than treating each one as a strange exception. They can compare answers for clarity and safety, instead of celebrating a score that hides the conditions under which the score was earned.

This is also a reminder to be precise about claims. “Supports Twi” can mean many things. It may mean translation, a limited set of prompts, or a conversation that stays coherent when the speaker naturally mixes languages. Those are different experiences, and people deserve to know which one they are getting. This plain comparison of mixed-sentence handling explains the distinction.

The useful standard is understanding before fluency

Back in Adwoa’s kitchen, the best outcome is simple. The system should recognise that the family is describing a concern, preserve the meaning of the mixed-language message, and avoid pretending it can settle a medical question. It should help make the next step clearer, including when the next step is to contact a qualified health professional.

That is a higher standard than sounding polished in a demo. It asks whether the system can follow the language people bring to it, especially when the stakes make people speak as they normally do.

For teams building Twi and Ghanaian English AI, the practical work starts with the data: test mixed-language conversations as mixed-language conversations; inspect failures for lost context; and make room for uncertainty where a health question requires professional care. A dataset at this scale gives that work a firmer place to begin.