Adaptive Exploration: Elevating Educational Impact of Unsupervised Knowledge Graph Question Answering

Xi Yang, Zhangze Chen, Hanghui Guo, Tetiana Shestakevych

Abstract: Question answering (QA) systems have been widely used in educational domain, significantly contributing to immediate information access and enhancing learning experiences. This paper introduces an efficient educational Knowledge Graph Question Answering (KGQA) framework that operates without relying on annotated training data. Our method enables the swift deployment of new knowledge graphs by simulating human-like information acquisition through a symbolic exploration module. Leveraging diverse program generation and large language models (LLMs), we formulate natural language questions to optimize the query for each program. To address semantic accuracy challenges in LLMs due to the absence of contextual training data, we propose adaptive strategies, including dynamic contextual re-ranking. These techniques significantly enhance question generation precision, showcasing robust performance in unsupervised settings. The framework demonstrates exceptional adaptability across varied queries and outperforms state-of-the-art models in zero-shot queries across various knowledge graph scales and datasets, underscoring its efficacy and scalability.

External IDs:doi:10.1007/978-981-96-0055-7_20