Malayalam clinical AI

Kerala does not consult in one language at a time.

A consultation here moves between Malayalam, English and Malayalam written in Latin characters — often inside a single sentence. Clinoria's research programme exists because that mixing breaks most clinical NLP assumptions, and because we would rather publish the breakage than ship around it.

Experimental research · not a supported production language

Read this first

Clinoria does not claim Malayalam support. English is the supported language. Malayalam is under active research. In the current build, Malayalam written in its own script behaves reasonably on negation and symptom extraction, while Malayalam typed in Latin characters (“Manglish”) can fail completely or, in one documented case, invert clinical polarity.

What works

Malayalam script, clause-scoped negation.

This is the case the current implementation handles correctly, and it is the reason the Malayalam research is worth continuing rather than starting over.

Synthetic transcript · Malayalam script
Doctor പനി ഉണ്ടോ?
Patient ഇല്ല, പനി ഇല്ല.
Patient ചുമ ഉണ്ട്.
Extracted facts Correct
ConceptStatusNote
feverQUESTIONEDfrom the question
feverABSENTdenial recognised
coughPRESENTseparate clause

Verified output The denial and the present symptom coexist: “The patient denies fever” appears alongside cough as a current finding. Negation is scoped to its clause rather than to the whole utterance.

What fails

The same conversation, typed the way people actually type.

Most Malayalam written on a phone at speed is Latin script. The same clinical content, transliterated, currently produces either nothing or the opposite of what the patient said.

Failure 1 · recall

Manglish yields zero facts

“Pani undo? Illa, pani illa. Chuma undu.” — a complete, unambiguous consultation — extracts no clinical facts at all. The rendered note is fourteen sections of “Not documented”.

Failure 2 · polarity

A denial becomes an assertion

“Chest pain illa.” is recorded as chest pain: PRESENT and rendered as “The patient reports pain in the chest.” This is the most dangerous defect in the system: it is fluent, plausible and exactly wrong.

Why this happens — the actual mechanism

Malayalam negation and scoping rules live in the normalisation layer that processes speech-recognition output. Text submitted through the typed path is deliberately not passed through that layer, on the reasonable assumption that pasted text has already been normalised by whoever typed it.

That assumption holds for English and fails for Malayalam. Setting the language to Malayalam does not change it, because the normalisation stage is skipped before language is considered.

Honest framing This is a pipeline-architecture defect surfaced by a language feature, not a model-quality problem. It is fixable, and it is unfixed today.

The research problem

Code-switching is not a translation problem.

In a real Kerala consultation, a symptom may be introduced in English, quantified in Malayalam, denied with a Malayalam verb on an English noun, and hedged in a third register. Each of those switches can carry a clinical distinction.

  1. What we have

    Rules and tests

    • Malayalam-first negation handling on the recognition path
    • 248 Malayalam-focused unit tests
    • A documented, reproducible failure taxonomy
  2. What is missing

    Evidence

    • No consented code-switched corpus
    • No clinician-authored gold documentation for Malayalam consultations
    • No published evaluation in this language
  3. What would change it

    A Kerala corpus

    • Consented recordings from multiple districts and specialties
    • Two independent clinician annotations per recording
    • Held-out evaluation that we commit to publishing, including failures

Language policy

Four rules we hold ourselves to.

  1. Never claim Malayalam support until held-out evaluation in Malayalam exists and is published.
  2. Never publish a real consultation transcript, in any script.
  3. Never let a Malayalam example stand in for a general capability claim.
  4. Never quietly remove a failure from this page when it is fixed; mark it resolved with a date.

If you are a Malayalam-speaking clinician

You are the missing input. Not encouragement — transcripts, with your own documentation of what the correct note should have said, under a protocol you approve. That is the only route from “the engineers think this looks right” to evidence.

Evaluation protocol

Collaborate

Help us find where Malayalam clinical language breaks.

A demonstration can run Malayalam script, Manglish and English cases side by side, including the failures documented above.