- UNIST develops personalized AI 'ExPerT' that analyzes both question content and typing behavior

- Expertise estimation error cut by 18.4%; applications expected in medicine, law, finance and education

Kong Tae-sik (left), a professor at Ulsan National Institute of Science and Technology, and researcher Park Ye-ji, who conducted the study. [UNIST]
Kong Tae-sik (left), a professor at Ulsan National Institute of Science and Technology, and researcher Park Ye-ji, who conducted the study. [UNIST]

"If someone types technical jargon quickly and without hesitation, they are likely an expert."

Researchers have developed a generative AI technology that goes beyond analyzing what users ask — it also reads how they type to assess their level of expertise. The advance is seen as a step closer to realizing truly personalized AI that adjusts the depth and style of its answers based on each user's knowledge level.

A research team led by Kong Tae-sik, a professor in the Department of Computer Science and Engineering at Ulsan National Institute of Science and Technology (UNIST), announced Sunday that it has developed ExPerT, a personalized AI system that assesses a user's expertise in real time by simultaneously analyzing the content of their questions and their typing behavior when entering technical terms.

Existing personalized AI systems typically rely on user profiles set up in advance or on past conversation histories. But such approaches have limitations: a computer scientist may know nothing about accounting, and even within the same field, a person's knowledge level can vary depending on the specific topic.

To address this, the research team adopted an approach that judges expertise on a question-by-question basis in real time. Particularly, the team used not just the content of a question but the act of typing itself as a clue to the user's level of knowledge.

The researchers noticed that even when typing the same technical term, experts and non-experts differ in three measurable ways: how long a key is held down and released, the interval before moving to the next key, and how often they use the backspace key.

ExPerT feeds both "semantic information" — the expressions and technical terms in a question — and this typing data into a generative AI model such as ChatGPT to estimate the user's expertise across five levels. It then passes that expertise estimate along with the original question back to the AI, which tailors the difficulty and explanatory style of its response to match the user's level.

When the team analyzed 1,270 questions entered by 40 participants across the fields of chemistry, computer science and business administration, the average gap between the AI's expertise assessment and participants' self-evaluations was 0.488 levels when only the semantic content of questions was used.

A comparison of conventional static AI personalization and ExPerT's personalization approach. [UNIST]
A comparison of conventional static AI personalization and ExPerT's personalization approach. [UNIST]

When typing behavior was added to the analysis, that gap narrowed to 0.398 levels — meaning the addition of typing data alone reduced the expertise estimation error by 18.4 percent.

The gap with existing personalization methods was even wider. Conventional approaches that extract a user's tendencies and background from past conversations showed an average discrepancy of 1.162 levels from actual expertise. By contrast, the team's method — which analyzes the meaning of each individual question in real time — brought the error down to 0.488 levels.

The research team expects the technology to find applications in specialized chatbots for fields with large knowledge gaps, such as medicine, law and finance, as well as in edtech platforms that tailor explanations to each student's level of understanding and in business-to-business customer support systems.

"We sought to overcome the limitations of existing personalized AI, which relied solely on fixed profiles or previous conversation histories and could not reflect real-time context," Kong said. "We will be able to provide a personalized AI interface optimized to each user's real-time level of understanding across a wide range of fields."

The findings were accepted for oral presentation at ACL 2026, an international conference on natural language processing, placing the paper in the top 4 percent of all submissions.


nbgkoo@heraldcorp.com