Google to Teach AI Aboriginal English
The project could take longer than expected, however, due to a lack of reference materials and cultural sensitivities.
Aboriginal and Torres Strait Islander people are being excluded from technology that does not understand the way they speak, the terms they use, or the cultural significance behind them.
However, a multi-year project between Google and researchers at the University of Western Australia will try to improve access to a range of technologies by training an artificial intelligence model in Aboriginal English.
The internet giant revealed the project on Feb. 19, in the latest investment from its $1 billion Australian Digital Future Initiative announced in 2021.
The project, which is due to add Aboriginal English to Google services in mid-2026, could take longer than expected, however, due to a lack of reference materials and cultural sensitivities.
Despite its lack of recognition by AI or voice recognition technology, University of WA adjunct lecturer Glenys Collard said most Indigenous children began school speaking Aboriginal English they learnt at home.
The powerful language featured different grammatical structures to Australian English as well as different meanings and cultural references which often went unrecognised, the Nyungar scholar said.
"People aren't getting services through the whole system," she told AAP.
"I want to give my mob the choice, the option to use technology."
However, capturing the nuances of Aboriginal English will be complicated, University of WA associate professor Celeste Rodriguez Louro said, as linguists typically reached for a dictionary or a set of grammatical rules as a first step.
"There's no standard language ideology that applies to Aboriginal English in the same way we think about mainstream varieties of English," she said.
"We need to be careful what we are imposing from our own ideologies that we carry as speakers of majority languages."
Storytelling sessions that the university's Language Lab had used to capture Aboriginal English previously would not be appropriate for this project, she said, as it could expose stories of trauma.
"Things shifted in the work we're trying to do for Google ... because we did not think our structured yarning sessions would be culturally safe," she said.
Instead, she said the university would recruit and train Indigenous research assistants in different parts of Australia to capture Aboriginal English speakers' responses to questions and stimuli.
Once recorded, the language could be added to a range of services, Google Research scientist Ben Hutchison said, including AI-powered voice assistants in cars and phones, as well as text message dictation and YouTube video captions.
"Ideally, everyone should be able to talk to AI systems in ways that are natural to them but AI systems haven't always encountered different ways of speaking before," Hutchison said.
"Previous research actually suggests that if you're from an under-represented or minoritised group (and) an AI system fails to understand you, you're more likely to feel self-conscious, both about yourself and about your community."
Language collected through the research project will be vetted by an Indigenous advisory committee before being made available to Google, he said, and the technology would be evaluated and tested before launch.