Knowledge Processing in the Era of Large Models: New Opportunities and New Challenges

The article begins by highlighting that both knowledge graphs (KG) and large language models (LLM) are essential tools for representing and processing knowledge, with LLMs enhancing language understanding and KGs enriching structural knowledge.

It then contrasts language and knowledge, explaining how ChatGPT encodes vast world knowledge in neural parameters, while KGs use graph structures to capture hierarchical, relational, temporal, and logical information, and argues that their deep integration can provide more reliable and controllable AI systems.

The discussion moves to the role of KGs within the LLM technology stack, covering knowledge enhancement, structural enhancement, prompt engineering as knowledge engineering, and the use of KGs to improve instruction fine‑tuning, knowledge editing, and alignment, illustrating each point with examples such as RetroPrompt, KnowPrompt, and KG‑of‑Thoughts.

Conversely, the article explores how LLMs can benefit KG pipelines: automated KG construction, knowledge extraction models (e.g., DeepKE‑LLM/KnowLM), structured‑knowledge LLMs that translate natural language to query languages (SPARQL, Cypher), and LLM‑augmented KG reasoning that combines symbolic and neural approaches.

Finally, the authors summarize three perspectives—"Language vs. Knowledge," the emerging "New KG" technology stack, and the transition from large language models to large knowledge models—emphasizing that future AI will require scalable models capable of handling diverse structured knowledge representations.