HubLink: A Novel Question Answering Retrieval Approach over Knowledge Graphs

The rapid growth of scholarly literature poses challenges for efficient and effective information retrieval. Existing Question Answering over Knowledge Graphs (KGQA) systems, particularly those relying on Semantic Parsing, struggle with schema dependency and required training data. In this paper, we introduce HubLink, a schema-agnostic, training-free KGQA approach leveraging pre-trained Large Language Models to enhance scholarly search with semantic aspects. HubLink structures (research) knowledge graphs into conceptual hubs, enabling source-aware inference for literature. For evaluation, we use the Open Research Knowledge Graph as the underlying knowledge base and utilize a dataset from software architecture research to populate the graph. The empirical results show that our approach HubLink outperforms three state-of-the-art baselines, especially for complex queries, marking a major advancement in scholarly KGQA. In future work, we aim to explore more advanced techniques to improve the final answer generation.