Sirvi Autor "Nkem-Eze, Chioma Jessica" järgi

Nüüd näidatakse 1 - 1 1

Design and Evaluation of an AI-Assisted COMPS Tutor for Students with Learning Difficulties in Mathematics
(Tartu Ülikool, 2025) Nkem-Eze, Chioma Jessica; Barbu, Eduard, juhendaja; Lipmaa, Kateryna, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This thesis presents Nutikas, an AI-assisted tutor that automates Conceptual Model-Based Problem Solving (COMPS) for early-grade additive word problems, designed with learners with special educational needs (SEN) in mind. Nutikas uses a four-step prompt pipeline: (i) super-category classification (Change / Combine / Compare). (ii) 12-way subtype selection, (iii) schema slot filling (e.g., Start/Change/End), and (iv) story-grammar questions to align large language model (LLM) outputs with instructional scaffolds. Three current LLMs (GPT-4.1, Claude Sonnet 4, and Gemini 2.5 Flash) are evaluated on a 120-item corpus covering all COMPS additive subtypes and score four dimensions: category, subtype, mapping (equation fidelity), and answer. Answers are near the ceiling (≥99.2%), while residual errors concentrate in schema mapping, especially the polarity of Change-Separate problems, where the COMPS convention requires a non-negative change magnitude. Mapping accuracy ranges from Gemini at 98.3% to Claude at 91.7% to GPT-4.1 at 85.0%, suggesting that the remaining variance reflects representation conventions rather than arithmetic capability. A small usability pilot with two SEN students (SUS-Kids mean 68.8) and one teacher indicates acceptable usability and highlights the need for clearer analytics on the teacher dashboard. While Tier-2 findings are formative and the scope is additive only, Nutikas already delivers accurate solutions with actionable paths to close the remaining mapping gap.