Researchers at CU Boulder are using artificial intelligence to develop digital models representing children that learn one language at home early in life and then begin learning another language in pre-school.
The interdisciplinary work sits at the crossroads of language development, education theory and the future role of artificial intelligence in both. Specifically, it’s an early step towards gaining a better understanding of how children learn language and developing ways to support and teach them. The work could also be used to improve other classroom and learning environments.
Katharina Kann, an assistant professor in the Department of Computer Science, is co-leading the work with a focus on the natural language processing aspects. She said the team’s ultimate goal is to support young children as they learn a new language without losing knowledge and skills they built around their first.
“To do that, we are adapting machine learning models to help detect language in teaching materials that is difficult for non-native speakers to understand,” she said. “Those models can be used to rewrite that information and make it more accessible quickly.”
Kann is working with Associate Professor Eliana Colunga from the Department of Psychology and Neuroscience. Colunga’s team is currently testing children who are learning English in combination with another language – or who are learning English alone – to see how their language development differs. That data will be used to validate and refine the models from Kann’s team so that they better match how children learn.
The research is funded through a seed grant from the Engineering Education and AI-Augmented Learning (EER-AIL) Interdisciplinary Research Theme within the College of Engineering and Applied Science. Interdisciplinary Research Themes help researchers coordinate faculty hires, share facilities and use seed funding to leverage work that could provide transformational societal impact.
The scope of the EER-AIL theme includes research in engineering and computing education and assessment, as well as artificial intelligence, machine learning and the convergence between all those areas.
One key goal of the theme is to develop the theories, technologies and technical expertise for advancing student-centered learning and creating next-generation learning environments in K-16, graduate and professional engineering and computing education situations. Kann said the simplification process being developed could be used for any topic, including complex course and learning materials in engineering disciplines – making it a great fit for that part of the initiative’s goals.
With the work still in its early stages, the team is currently studying if neural network language models that have been pre-trained on a source language like Spanish and are then trained on English make the same grammatical errors in practice as their human counterparts.
“One challenge we encountered bringing these two disciplines together was finding ways to compare the model and child performance so we can use the children’s performance to better refine the models,” said Kann. “There are many things state-of-the-art AI cannot do well but creating computational models that make errors similar to children language learners is still a major challenge we hope to address.”
As work progresses, the team plans to use the models to design personalized educational materials that will incorporate a variety of factors, such as the language children are learning at home. The team is currently seeking families with children that are 3-to-5 years old and are interested in participating in the research.
“It is all done virtually, only takes 30 minutes and the children really seem to enjoy it," Colunga said. “Language is a critical way in which we all interact with the world, but it doesn’t come easy to everyone. If we can build good AI models of how children learn language, we will be able to find better ways to help those children who struggle, and the applications go out from there.”