The Dagstuhl Seminar "Large Language Models Meet Knowledge Graphs" is a five-day event designed to explore the intersection of Large Language Models (LLMs) and Knowledge Graphs (KGs). It aims to address the integration of parametric knowledge from LLMs with explicit knowledge from KGs, marking a shift towards hybrid knowledge representation. The event will gather approximately 40 participants, including leading international experts and promising young researchers, to discuss the opportunities and challenges in this evolving field.

The seminar is structured in three phases. Preparations include a survey to guide group formation based on participants' expertise and the distribution of key papers for preparation. The first day features keynotes and Q&A sessions, followed by the formation of breakout groups based on research questions, culminating in a plenary discussion to share initial ideas. The second phase spans days two to four, focusing on in-depth exploration through working groups. One group will investigate how LLMs can enhance KG tasks like entity resolution and link prediction. Another will explore how KGs can improve LLM training and prompt generation. The third group will analyze real-world applications, including discussions on future Neuro-Symbolic AI systems. Participants will rotate between groups to foster cross-pollination of ideas, with plenary sessions at the end of each day for feedback. The final day is dedicated to consolidating discussions and planning future work. Each group will present detailed reports, outlining key research questions and next steps. Sessions will formalize plans for joint papers, grant proposals, and future collaborative projects. A roundtable will synthesize the seminar's findings and outline concrete next steps, including joint publications and future research initiatives.

The seminar addresses key research questions, such as how Knowledge Graphs can benefit from Large Language Models by integrating LLMs with KGs while preserving correctness. It also explores how LLMs can leverage the structured knowledge from KGs and how the unification of KGs and LLMs can foster AI applications, such as improved QA systems and recommender systems. Topics include LLM-based Knowledge Graph Management, enhancing tasks like knowledge extraction and entity resolution, and the conception of KG-driven Large Language Models, using KGs to improve LLM training and prompt construction. The interplay in use cases is also explored, focusing on designing future applications and addressing challenges like noise and hallucination.

Expected outcomes include a roadmap to guide future research and development in LLMs and KGs, publications such as survey articles and guidelines, and collaborations through joint research projects and scientific events. The seminar targets a diverse group of experts from academia and industry, including specialists in Semantic Web, Data Management, Knowledge Representation, and NLP, with efforts to ensure gender diversity and representation of under-represented groups. Overall, this seminar aims to provide a foundational moment for the LLM+KG community, fostering collaboration and innovation in the integration of language models and knowledge graphs.

Creative Commons BY 4.0

Angela Bonifati, Jan-Christoph Kalo, Jeff Z. Pan, Simon Razniewski, and Luke Zettlemoyer