Building a Better Way to Assess Kids’ Language Skills

In 2021, during the COVID-19 pandemic, U of T education scholar Eunice Eunhee Jang was researching how AI could improve language testing when she met a newcomer family from China who feared their Grade 4 daughter’s communication skills had suffered during school closures.

Using BalanceAI, the literacy assessment platform she was developing, Jang evaluated the student and found typical development patterns, along with clear abilities in cognitive understanding and reading – and a need to improve her oral language.

“The mother was in tears – not because of bad news, but because for the first time, her child’s abilities had been seen fully and fairly,” recalls Jang, who leads the Innovative Dynamics in Educational Language Assessment Lab at OISE. “That moment crystallized why assessments have to be comprehensive enough to reveal each child’s potential.”

The encounter helped shape BalanceAI. When teachers use it to assess literacy and language development, they receive granular results that allow them to tailor instruction. Unlike common AI assistants such as ChatGPT, Google Gemini and Microsoft Copilot – which rely largely on adult-generated content – BalanceAI is trained on primary students’ writing and speech samples. This enables it to recognize the distinct patterns in children’s language.

“Teachers need an educational AI that is informed by how children communicate,” Jang says. “This is how they can understand the potential and needs of their students. Without such evidence, how can they personalize the way they teach?”

BalanceAI addresses other problems with conventional literacy assessment. Reading comprehension, vocabulary and spelling are typically tested separately – a laborious process that often yields little more than numerical scores. Oral language skills are rarely evaluated – a gap Jang considers inequitable, since students from diverse linguistic backgrounds may struggle with reading and writing but excel verbally.

BalanceAI measures multiple literacy and language skills at once. In one assignment, learners watch a short video and write a paragraph expressing their opinion. Within 30 seconds, the system generates a bar graph analyzing their vocabulary, grammar and spelling, and how they organize their ideas. It also gathers students’ reflections on how they perceive their learning and translates these into indicators of motivation, self-regulation and enjoyment – data that can help teachers refine their approach.

“Rather than teaching everybody with one-size-fits-all instructional materials, teachers can customize their content based on the evidence,” says Jang.

The platform builds on a decade of research, drawing from authentic student language – what children actually say, write and read aloud – to form the foundation of the large language model she developed with machine learning engineers and computer programmers, including U of T students.

Peer-reviewed studies show that BalanceAI can reliably assess writing and oral language, and provide feedback that supports students’ development in reading, writing and speaking. The platform is now used by students at OISE’s Dr. Eric Jackman Institute of Child Study laboratory school and by English language learners at a primary school in China. Jang is working to expand its use in Ontario public schools.

“I really believe this technology can level the playing field in public education,” she says.