Scholarlink

This chapter introduces an interactive learning system that uses OpenAI’s large language model (LLM) to automatically create and evaluate essay-style questions aligned with Bloom’s Taxonomy. The system processes educational PDFs using term frequency-inverse document frequency (TF-IDF)-based keyword extraction (from unigram to trigram) to connect question content to its source, reducing hallucinations. Semantic evaluation is performed with the ada-002 embedding model and cosine similarity to assess student answers. When similarity scores fall below a set threshold, the system provides follow-up questions to support adaptive learning. The platform supports multiple languages and was tested in a two-phase study with university students in Japan and Indonesia. Evaluation using the technology acceptance model (TAM) showed high perceived usefulness (0.991) and ease of use (0.850), with improved results after system updates. Findings also indicate that documents in their original language (English) produce more stable and accurate semantic similarity scores than translated versions, highlighting the limitations of cross-lingual embeddings in AI-assisted educational assessment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An LLM-Based Adaptive Learning System for Automated Essay Assessment Aligned with Bloom’s Taxonomy

  • 章节

  • 发表日期：

    2026年

查看全文
  • Laurentius Gusti Ontoseno Panata Yudha,
  • Hapnes Toba,
  • Oscar Karnalim,
  • Hendra Bunyamin,
  • Terutoshi Tada

摘要

This chapter introduces an interactive learning system that uses OpenAI’s large language model (LLM) to automatically create and evaluate essay-style questions aligned with Bloom’s Taxonomy. The system processes educational PDFs using term frequency-inverse document frequency (TF-IDF)-based keyword extraction (from unigram to trigram) to connect question content to its source, reducing hallucinations. Semantic evaluation is performed with the ada-002 embedding model and cosine similarity to assess student answers. When similarity scores fall below a set threshold, the system provides follow-up questions to support adaptive learning. The platform supports multiple languages and was tested in a two-phase study with university students in Japan and Indonesia. Evaluation using the technology acceptance model (TAM) showed high perceived usefulness (0.991) and ease of use (0.850), with improved results after system updates. Findings also indicate that documents in their original language (English) produce more stable and accurate semantic similarity scores than translated versions, highlighting the limitations of cross-lingual embeddings in AI-assisted educational assessment.