Integrating Large Language Models with Knowledge Graphs: Current Status and Future Directions

Before large models, knowledge graphs already had many applications such as knowledge Q&A. When combined with search engines, e.g., knowledge cards, they provide direct factual answers and objective knowledge display, and they remain more proficient in this area. This mode will not be greatly affected by the arrival of large models because the knowledge acquisition scenario itself is already broad.

So, what new changes does the arrival of large models bring? Large models actually enhance human‑machine interaction, lower the interaction threshold, and enable AI to understand users' natural language inputs more fully, finely, and accurately. For more casual user questions, large models are more flexible and their answers more natural.

Therefore, large models and knowledge graphs are not a complete replacement relationship; they can be combined to solve problems that were previously unsolvable. Past KG‑based conversational systems involved many aspects such as chit‑chat, knowledge Q&A, and task‑oriented Q&A (e.g., booking flights or hotels). These tasks made the system very complex and heavy to deliver. After combining large models with knowledge graphs, chit‑chat and knowledge Q&A can be better fused, making dialogue systems more natural and end‑to‑end.

When to use a knowledge graph, large models will naturally find the optimal way; task‑oriented Q&A can also be organically combined with other APIs, improving the architecture of customer service and dialogue systems.

Moreover, knowledge graphs still have an irreplaceable role in fields with highly structured data characteristics such as risk control and financial market analysis, so they will continue to exist in the long term.

For many years, knowledge graphs have faced a serious problem: high construction complexity and cost, lacking a unified building capability to solve all knowledge construction issues, because many knowledge expressions are highly diverse and non‑standard at the text or multimodal level.

The industry is actively exploring how to leverage the powerful semantic understanding of large models to enhance the standardized construction of knowledge, and some progress has been made, but many challenges remain.

Knowledge graphs need to reduce cost; if the cost is low enough, structured decisions will be more rigorous than those made by large models. Therefore, we must both accommodate large models and insist on the long‑term existence of knowledge graphs.

Thus, these two modes should be integrated and viewed in an inclusive manner.

To deeply discuss the current status and future development of knowledge graphs and large models, the DataFunSummit2024: Knowledge Graph Online Summit will be held online on March 23, 2024, 9:00‑17:00, and we welcome practitioners to participate and exchange ideas!