knowledge graph embedding

This paper addresses core challenges in knowledge graph (KG) relation modeling—namely, mapping complexity (e.g., one-to-many, many-to-many), semantic patterns (symmetry, invertibility, compositionality), implicit hierarchies, and dynamic sparsity. It systematically surveys mainstream approaches, including tensor factorization, geometric embeddings (e.g., RotatE, hyperbolic space), polar coordinate modeling, and neural networks. Methodologically, it introduces, for the first time, a unified four-dimensional analytical framework covering mapping properties, pattern awareness, hierarchical structure, and dynamic adaptability—establishing the most comprehensive taxonomy of relational characteristics to date. The work further identifies rule-guided learning, multimodal fusion, and dynamic relation representation as critical future directions. By providing a principled theoretical foundation for knowledge graph embedding (KGE) design, this study significantly advances robust relational representation under sparse and evolving KG settings.

This paper addresses the limitations of traditional rule- and statistics-based approaches in knowledge graph (KG) construction—namely, ontology engineering, knowledge extraction, and knowledge fusion. It proposes a novel “language-driven generative KG construction” paradigm that unifies schema-based and schema-agnostic methods, establishing synergistic mechanisms between large language models (LLMs) and symbolic KGs for structured organization and open-ended semantic expression. The method integrates prompt engineering, knowledge representation learning, automated reasoning, and multimodal modeling to enable dynamic, interpretable knowledge acquisition and fusion. The study comprehensively surveys technical pathways and bottlenecks, identifying three key research directions: LLM reasoning enhancement via KGs, agent memory modeling with KGs, and multimodal KG construction. Ultimately, this work advances the development of adaptive, neuro-symbolic intelligent knowledge systems.

This work proposes a novel knowledge graph representation paradigm that addresses the limitations of traditional approaches, which rely on predefined symbolic relations and consequently fail to capture the context-dependency, fine-grained semantics, and inherent uncertainty of real-world relationships—often leading to critical information loss. By reformulating relations as natural language descriptions rather than discrete symbols, the proposed framework leverages the generative and reasoning capabilities of large language models. It integrates prompt engineering with a minimal structural backbone, thereby harmonizing structured scaffolding with unstructured semantic expression. This hybrid design substantially enhances the semantic richness and context-awareness of relation modeling, offering a more adaptive and expressive pathway for knowledge graph construction in the era of large language models.

This work addresses the challenge that existing large language models often generate facts violating ontological constraints during knowledge graph extraction, thereby hindering symbolic reasoning and complex querying. The authors propose a neuro-symbolic framework that innovatively defers ontology-aware consistency correction to a post-processing stage following open-domain extraction. By integrating embedding-driven entity type and predicate normalization, ontology-guided large language model refinement, and SPARQL-based graph pattern analysis, the approach significantly reduces the computational cost of large model invocations while effectively enhancing the logical consistency of the resulting knowledge graph. Experimental results demonstrate that the method substantially decreases ontology violations without compromising question-answering performance and reliably supports graph pattern operations required for symbolic queries.

Current approaches to scientific discovery lack sufficient capacity for deep cross-domain knowledge association and dynamic reasoning. Method: This paper proposes a graph reasoning framework integrating category-theoretic isomorphic representation with preference-guided recursive modeling. It formalizes task-driven reasoning as structured mapping and enables principle-level alignment across heterogeneous domains (e.g., mythology and materials science) via a “Knowledge Garden Growth” strategy. The framework combines graph neural networks, symbolic abstraction, categorical semantic encoding, and reinforcement learning–inspired preference modeling to instantiate a 3-billion-parameter recursive language model. Contribution/Results: It achieves, for the first time, autonomous dynamic knowledge graph generation, abstract pattern extraction, and interpretable answer derivation. Empirical evaluation demonstrates significant improvements in reasoning depth and cross-domain adaptability across hypothesis generation, materials design, and creative reasoning tasks, thereby enabling multi-disciplinary, autonomous scientific discovery.

Knowledge graph reasoning faces two key challenges: the fragmentation between neural and symbolic approaches, and poor adaptability across diverse reasoning scenarios. To address these, this paper proposes Tunsr, a unified neural-symbolic reasoning framework that— for the first time—integrates transductive, inductive, interpolative, and extrapolative reasoning within a single model. Its core innovations include: (i) a unified reasoning graph structure with forward logical message passing to synergistically combine neural computation and symbolic rules; (ii) the FARI algorithm, enabling dynamic rule merging and automatic induction of first-order logic rules; and (iii) attention-enhanced structural awareness. Extensive experiments across 19 benchmark datasets demonstrate that Tunsr significantly improves cross-scenario generalization, validating both the effectiveness and scalability of deep neural-symbolic integration.

Traditional knowledge graphs struggle to capture high-order interactions among multiple entities in scientific reasoning, limiting intelligent inference in complex research scenarios. This work proposes the first hypergraph-based framework for scientific knowledge representation, constructing a global hypergraph comprising 161,172 nodes and 320,201 hyperedges by extracting multi-entity co-occurrence relations from over 1,100 publications on biomaterial composite scaffolds. By leveraging hypergraphs, the approach avoids the combinatorial explosion inherent in pairwise expansions and preserves the contextual integrity of scientific statements. Furthermore, it enables verifiable, teacher-free cross-concept reasoning through hypergraph traversal constrained by node intersections. Experimental results demonstrate the generation of novel, mechanistically interpretable material design hypotheses, such as cerium oxide–chitosan–PCL scaffolds.

To address prompt sensitivity, factual hallucination, and ontological inconsistency of large language models (LLMs) in domain-specific knowledge graph (KG) construction, this paper proposes GraphMERT—the first efficient, scalable neuro-symbolic architecture. GraphMERT is a lightweight encoder-only graph model that performs end-to-end semantic relation extraction and structured KG generation directly from unstructured text (e.g., PubMed diabetes literature) via internal representation learning and knowledge distillation. Crucially, it eliminates reliance on LLMs, thereby ensuring factual verifiability (with provenance) and strict ontological consistency. Experimental evaluation on the diabetes domain demonstrates GraphMERT’s superior reliability: it achieves an FActScore of 69.8% and a ValidityScore of 68.8%, significantly outperforming LLM-based baselines. This work establishes a novel paradigm for constructing domain KGs that is highly reliable, interpretable, and scalable.

To address the challenge non-expert users face in directly querying knowledge graphs, this paper proposes a natural language-driven interactive query construction method. The approach employs a two-stage constrained language model that integrates ontology-based semantic constraints to generate syntactically and semantically valid query prototypes—thereby avoiding invalid classes, relations, and grammatical errors. A visual editor enables users to iteratively refine queries via natural language descriptions and graphical adjustments. Finally, an interpretable SPARQL translation pipeline converts the refined prototype into standard SPARQL. Evaluated across multiple ontologies and language models, the system consistently produces correct SPARQL queries without manual intervention, outperforming existing baselines in both retrieval accuracy and efficiency. Validation through synthetic data experiments and an initial user study confirms the method’s effectiveness, usability, and practical applicability.

This work proposes a modular differentiable neuro-symbolic reasoning framework to address the challenge of efficiently and accurately integrating neural models with knowledge graphs in knowledge-intensive multi-hop question answering. The approach innovatively combines soft unification-based symbolic reasoning, a neural path evaluator, and a value-guided Monte Carlo tree exploration mechanism to jointly optimize reasoning accuracy and computational efficiency. Evaluated on standard KGQA benchmarks, the model achieves high answer accuracy while substantially reducing the number of expensive graph queries and model invocations, demonstrating its effectiveness and practicality.

This study addresses the lack of systematic understanding regarding the applicability, timing, and methodology of integrating graph structures with large language models (LLMs) across diverse scenarios. It proposes a structured analytical framework that categorizes existing graph-LLM approaches according to task objectives, graph types, and integration strategies—encompassing techniques such as prompt engineering, data augmentation, joint training, and agent-based architectures—and covering multiple graph modalities including knowledge graphs and causal graphs. Through cross-domain evaluation, the work identifies optimal practices and boundary conditions for various fusion schemes in tasks like reasoning and retrieval, offering researchers a principled guideline for selecting appropriate methods based on task requirements, data characteristics, and reasoning complexity.