Schmidt, Wilma JohannaWilma JohannaSchmidtRincon-Yanez, DiegoDiegoRincon-YanezKharlamov, EvgenyEvgenyKharlamovPaschke, AdrianAdrianPaschke2024-10-022024-10-022024-09https://publica.fraunhofer.de/handle/publica/476879The increasing amount of available research data leads to the need to scale scientific knowledge discovery, e.g., the conduction of systematic literature reviews (SLRs), to keep up with fast developments in research and further support decision-making in the industry.AI-based methods are gaining importance in these tasks and have been integrated into many SLR tools.Yet, several challenges are still open on applying especially neural methods on scientific knowledge discovery tasks.To address this, we evaluate various neural and neuro-symbolic scenarios on a specific generative writing task.While confirming existing concerns on pure Large Language Model (LLM) approaches for these tasks, we obtain a heterogeneous picture of Retrieval-Augmented Generation (RAG) approaches.The most promising candidate is a Knowledge Graph (KG) based context-enhanced LLM approach for Knowledge Discovery.enNeuro-Symbolic AIKnowledge GraphLarge Language ModelRetrieval-Augmented Generation (RAG)Systematic Literature Reviewconference paper