Patient dialects and accents degraded translation-app accuracy in aged-care wards

Description #

A major subtheme of the "Communication" theme was the concept of language: translation apps did not work when the patient's language was a dialect of a main language, and for Google Translate the accuracy of the translation was affected by the patient's accent, producing wrong words and confusing output. Linguistic variation among older CALD patients thus degraded app performance.

"Translation apps did not work if the language of the patient was a dialect of a main language "Some patients use different dialects" (Nurse 1, site 4), and for Google Translate, the accuracy of the translation was affected by the accent of the patient "Does not pick accents correctly so uses wrong words making the translation confusing" (Volunteer healthcare worker 1, site 3)." (Hwang, 2022, p. 582)

Methods Context #

What? #ⓘ

The observable: translation accuracy/coverage as a function of the patient's spoken language variety (dialect, accent), characterized qualitatively from staff/observer responses.

"A major subtheme within this theme was the concept of language." (Hwang, 2022, p. 582)

How? #ⓘ

Inductive qualitative content analysis of open-ended survey and observation responses under the "Communication" theme; apps were the fixed-phrase A-0018ArtifactA-0018Initial AI draftCALD Assist translation appA mobile healthcare translation app used by ward staff to communicate low-risk, everyday care messages to older CALD patients with limited English proficiency when a professional interpreter is not available or feasible.… / A-0017ArtifactA-0017Initial AI draftTalk To Me translation appA mobile healthcare translation app used by ward staff to convey low-risk everyday care messages to older CALD patients with limited English proficiency in the absence of a professional interpreter. "Talk To Me was devel… and the free-text A-0016ArtifactA-0016Initial AI draftGoogle TranslateA free, general-purpose machine translation app increasingly used as an ad-hoc communication tool in healthcare settings to bridge language barriers with LEP patients, including via voice-to-voice translation. "One such….

"Two researchers each read the responses to the questions independently of other questions to gain familiarity with the content. Each response was interpreted for meaning and codes developed based on the interpretation." (Hwang, 2022, p. 580)

Who? #ⓘ

Nursing/allied health staff and older CALD patients with limited English proficiency across four aged-care hospital wards; observed languages spanned Italian, Greek, Arabic, Vietnamese, Cantonese, Serbian, Macedonian and Spanish.

"The language of the observed interactions was as follows: Italian (6), Greek (4), Arabic (3), Vietnamese (2), Cantonese (2), Serbian (2), Macedonian (1) and Spanish (1)." (Hwang, 2022, p. 581)

Other Notes #

Accuracy failure from dialect/accent is distinct from, and additive to, the speed/usability failure captured in the parallel Google Translate EVD.