Stepwise AI tutoring helped pupils with dyslexia tackle maths word problems

For 83 Taiwanese pupils with dyslexia in grades four to six, a six-week study tested assistance with arithmetic word problems. The 39 assigned to chatbot practice improved more than the 44 receiving traditional instruction. Both groups had twelve 30-minute sessions. The 18-problem assessment was shown in print and read aloud. Li-Chih Wang and Hsu-Chan Kuo published the trial on 31 May in the Journal of Learning Disabilities.[1]

The learning problem sits at the meeting of language and mathematics. A learner can understand subtraction and still misread who has the larger quantity in a written story. Making the wording audible can remove one obstacle; understanding the relationship between the quantities remains another task.

Practice that kept the steps visible

The bespoke GPT-based platform offered graduated hints, text-to-speech and coloured cues. Pupils answered successive prompts before advancing. Staff supervised use and restarted the prompt when they spotted an obvious error. The comparison teaching matched lesson time, but did not match the frequency of individual feedback.[1]

These instructional ingredients have a history beyond generative AI. The US What Works Clearinghouse’s 2021 elementary mathematics guide recommends systematic teaching, clear mathematical language, suitable visual representations and deliberate instruction in word problems. For a multistep problem, it describes helping pupils understand what is being asked and connect the quantities through a diagram, table or equation. It also recommends gradually reducing visual guides as pupils become more comfortable.[2]

Consider two illustrative sentences. “Ana has eight pencils; Ben has two more than Ana” gives Ben ten. “Ana has eight pencils, two more than Ben” gives Ben six. The same word, more, appears in both. Working out whose quantity is being described determines which calculation answers the question.

That example also shows why highlighting words needs a clear teaching purpose. The What Works Clearinghouse warns against training pupils to map a keyword automatically to an operation. The wording has to be interpreted as part of the whole situation. A cue that helps someone locate information can be useful; an automatic shortcut from one word to one calculation can lead them astray.[2]

The result belongs to a supervised programme

The Taiwan trial measured short-term performance, with no later follow-up. Novelty and extra adult attention may have contributed. The paper describes a research platform, without giving a public access route; opening an ordinary chatbot would not reproduce the intervention.[1]

For schools, the study raises a precise design question: can software help a learner get from the words of a problem to the relationship they describe? The pupil’s equation and explanation provide a way to judge that transition. A useful system would make the language easier to work with while keeping the mathematical decision in view.

Report an error or suggest a correction