Large language models (LLMs) have demonstrated remarkable capabilities in natural language processing (NLP) tasks, yet they face significant challenges when applied to educational contexts. This paper introduces WisdomBot, a novel educational LLM that addresses these limitations by incorporating educational theories and knowledge into its training process.
WisdomBot employs Bloom’s Taxonomy to guide the acquisition of knowledge concepts and the design of instruction-tuning data. By leveraging LLMs and self-instruction techniques, the model acquires a comprehensive understanding of educational content and diverse cognitive processes, ranging from basic to advanced abilities. Additionally, the model utilizes retrieval augmentation strategies, including local knowledge base and search engine retrieval, to enhance its factual knowledge and generate more accurate and professional responses.
Experiments conducted on Chinese LLMs, such as Chinese-LLaMA and Qwen-7B-Chat, demonstrate the effectiveness of WisdomBot. The fine-tuned models outperform the original models in various educational tasks, including professional question answering, test problem generation, and intelligent tutoring. Moreover, WisdomBot exhibits superior advanced cognitive abilities, such as creativity, personalized ability, and logical reasoning.
Overall, WisdomBot represents a significant advancement in the field of educational AI. By combining the power of LLMs with educational theories and knowledge, it opens up new possibilities for personalized and effective learning experiences.
The work titled “WisdomBot: Tuning Large Language Models with Artificial Intelligence Knowledge”, was published on Frontiers of Digital Education (published on August 29, 2024).
DOI: 10.1007/s44366-024-0005-z