2026-10-12T01:28:51Z https://api.figshare.com/v2/oai

oai:figshare.com:article/32216784 2026-09-20T20:31:38Z category_29110 item_type_3 month_year_09_2026

SciGraphChat: Retrieval-Augmented Generation over Scientific Knowledge Graphs Kaplan (23190841) Jan Keim (11661868) Knowledge and information management Retrieval-augmented generation (RAG) Knowledge Graphs View-based Interface FAIR Data Access Semantic Web in Practice Research Infrastructures Knowledge Graph Usability <p dir="ltr">Effective retrieval and information synthesis of scholarly knowledge remain challenging because of fragmented publication practices, LLM hallucinations, and the limitations of keyword-based search. We present SciGraphChat, an operational system deployed as an AI assistant within the NFDIxCS consortium, enabling discovery and exploration of scholarly knowledge represented in research knowledge graphs (KGs) through an integrated conversational and graph interface. The system builds upon a schema-agnostic, training-free retrieval approach for KG question answering, supporting source-aware and semantically informed access to structured scholarly data. By coupling KG-based Retrieval-Augmented Generation with a dialogue-oriented interface, SciGraphChat accompanies responses with retrieved RDF subgraphs and links to underlying resources. Unlike conventional chat-based assistants, the system discovers the retrieved graph evidence and enables users to interactively inspect and navigate the underlying knowledge structure through a dedicated graph panel. This paper presents the design and deployment of such a scholarly knowledge discovery system, along with practical design recommendations for integrating Semantic Web technologies and generative AI within research infrastructures. We evaluate the system with NFDIxCS stakeholders and external users via the System Usability Scale (SUS) and the User Experience Questionnaire–Short (UEQ-S), complemented by qualitative feedback. Results show a mean SUS score of 69.4, positive mean pragmatic quality (UEQ-S mean 0.93), and neutral hedonic quality (UEQ-S mean 0.70). Qualitative feedback indicates that users value the combination of conversational access and graph-based inspection, while also requesting tighter integration with scholarly workflows such as reference export, corpus scoping, and adaptive result presentation. Our findings suggest that combining conversational retrieval with transparent graph-based evidence inspection can improve trust, explainability, and usability in knowledge discovery.</p> 2026-09-20T20:31:38Z Dataset Dataset 10.6084/m9.figshare.32216784.v1 https://figshare.com/articles/dataset/SciGraphChat_Retrieval-Augmented_Generation_over_Scientific_Knowledge_Graphs/32216784 CC BY 4.0