Reasoning over Knowledge Graphs : How can LLMs help?

Abstract:

Knowledge graphs (KGs) provide a structured way of representing entities and relations, but entity alignment-identifying pairs of entities across KGs that refer to the same real-world object-remains a major cross-domain challenge. Many existing approaches rely on graph structure or symbolic rules, yet often struggle with generalization. In this talk, I will present recent work that reframes entity alignment as a natural language inference (NLI) task, using neural models that approximate logical reasoning in natural language. I will also discuss how large language models (LLMs) can be combined with this framework to improve recall on difficult cases, while introducing new trade-offs. By bringing together KG context, pretrained entailment models, and LLM reasoning, we can move toward more generalizable and explainable solutions for KG integration. The talk will conclude with open challenges and future opportunities at the intersection of KGs and LLMs.

Suggested readings:

Sun, Z., Zhang, Q., Hu, W., Wang, C., Chen, M., Akrami, F., & Li, C. (2020). A benchmarking study of embedding-based entity alignment for knowledge graphs. Proceedings of the VLDB Endowment, 13(11), 2326-2340. https://doi.org/10.14778/3407790.3407828

Jradeh, M., Raoufi, E., David, J., Larmande, P., Scharffe, F., & Todorov, K. (2025). Graph Embeddings Meet Link Keys Discovery for Entity Matching. Proceedings of The Web Conference. https://doi.org/10.1145/3696410.3714581

Bio:

Ensiyeh Raoufi is a PhD candidate in Computer Science at the University of Montpellier (LIRMM, recently joined IRD, France). She has also recently joined the University of Montpellier, Polytech, as a lecturer (ATER, enseignante-chercheuse). She studied computer science for both her bachelor’s and master’s degrees, focusing on algebraic graph theory during her master’s. Her current research explores how neurosymbolic, neural, and large language models can be used to address the challenges of entity alignment and knowledge graph integration, with the goal of building approaches that are more generalizable and explainable.
More information is available on the LinkedIn profile: https://www.linkedin.com/in/ensiyeh-raoufi/