KG-NeSy 2026: The Third Workshop on Knowledge Graphs and Neurosymbolic AI

Biomedical knowledge exists in two complementary but distinct forms: unstructured scientific literature and structured knowledge graphs (KGs). Aligning them is essential for knowledge grounding, evidence retrieval, and KG completion, yet existing methods do not explicitly align free-text evidence with KG triples. We present a unified framework for systematically studying design choices for aligning biomedical text and KGs. With a text encoder and a KG embedding model both frozen, we learn only a lightweight projection between their spaces via a contrastive objective. This enables a fair comparison across six design dimensions: text encoder, KG embedding model, projection head, triple composition, training direction, and hard-negatives sampling. We construct CTD-Align, a corpus of over 22K one-to-one triple-document pairs linking chemical-gene interactions from the Comparative Toxicogenomics Database to supporting PubMed passages. We evaluate alignment on it in two retrieval settings: document-to-triple and triple-to-document. We find that the triple composition and the training direction (i.e., shared retrieval space) have the greatest impact, whereas the text encoder and hard-negatives sampling matter little. Overall, simple choices win: projecting text into the KG space with a linear head over concatenated subject, predicate, and object embeddings performs best. These findings establish lightweight contrastive alignment as an effective, practical foundation for bridging biomedical text and KGs.

Paper

Neuro-symbolic (NeSy) systems are commonly described using taxonomies and design patterns that capture different perspectives of the interaction between neural and symbolic components. However, these representations remain largely human-oriented and do not support interoperability or reasoning across taxonomies. This paper introduces a semantic framework for representing neuro-symbolic design patterns, taxonomies, and architectures in a machine-interpretable form. The proposed approach formalizes existing NeSy taxonomies using Boxology patterns, introduces the NeSy ontology for representing taxonomies, design patterns, and their alignments, and materializes these representations in a Neuro-Symbolic Knowledge Graph (NeSy-KG). Building on this representation, we define an axiomatic reasoning framework that infers implicit pattern hierarchies, equivalences, and cross-taxonomy alignments. An empirical evaluation on a corpus of published neuro-symbolic systems demonstrates that the proposed framework identifies interoperability relationships across alternative taxonomic representations and enriches the knowledge graph with previously implicit knowledge. The results establish a foundation for representing, comparing, and reasoning about NeSy systems using semantic technologies.

Paper

Neurosymbolic research typically assumes a pre-existing symbolic specification, leaving the upstream challenge of acquiring and formalizing requirements and constraints largely unaddressed. We present an architecture that fills this gap by using an OWL configuration ontology to mediate between neural constraint sources and downstream consumers. In this framework, LLM assistants elicit soft stakeholder preferences, while hardware specifications define hard physical and engineering limits. The ontology unifies these heterogeneous inputs, leverages description logic to identify unsatisfiability, and generates symbolic explanations that enable LLMs to interactively renegotiate terms with users. Any remaining conflicts are resolved downstream via priority-based relaxation. We illustrate our approach on a microgrid use case from the FLEXI project and argue its generalizability to multi-stakeholder domains where constraint acquisition is distributed across human and automated sources of unequal authority.

Paper

The actions afforded by an object depend on its situation: changing a location, time, instrument, or co-occurring object can change which of several plausible actions is preferred. We present TRACE-Afford, a neural-to-symbolic pipeline that converts screened language-model descriptions of everyday human actions into a reusable, typed affordance knowledge graph. The graph separates entities, intrinsic attributes, action contexts, bare actions, and object-bound affordance instances. Repeated observations are retained as frequency evidence and normalized locally to support inspectable shortest-path retrieval from multiple inputs. We also introduce a 255-pair Context Intervention Benchmark. Each pair holds the base object and two candidate actions fixed and requires their expected ordering to reverse between two situations that differ in instrument or resource, location, temporal or task-stage context, or co-occurring object. Built from 10.0 million screened sentence occurrences elicited with 2,000 seed terms, the resulting graph contains 3.35 million nodes and 12.08 million distinct labeled edges. TRACE-Afford achieves 82.4% pair accuracy; replacing local relative-frequency costs with absolute-frequency decay reduces accuracy to 71.8%, and removing frequency information reduces it to 53.3%. These results show that the proposed node schema and edge structure, together with relative frequency, enable a broad symbolic graph to recover contextual preferences while retaining the paths that support each ranking.

Paper

We propose a methodology for generating knowledge graphs from semi-structured data that integrates large language model (LLM) queries directly into RML mapping rules: structured fields are handled by standard RML rules, while LLMs extract ontological terms from unstructured text fields. This targeted use of LLMs–fed only short, focused text snippets–reduces the risk of errors compared to end-to-end LLM solutions, while preserving correctness guarantees for the structured part of the data. We evaluate our approach on a Medicines Information use-case using pharmaceutical data in JSON format, and assess the correctness of the resulting knowledge graphs and query answers against pure LLM-based baseline. This is an extended abstract of the paper published at DL 2025 titled “Accessing Semi-structured Data with RML and LLMs” https://ceur-ws.org/Vol-4091/paper54.pdf.

Paper

ℰℒ⁺_⊥ is a lightweight description logic designed to capture the expressive requirements of large practical ontologies while supporting tractable reasoning, as implemented by reasoners such as ELK. Recent efforts have developed scalable methods for learning continuous representations of such ontologies for inductive inference, but their coverage of the normalised axiom forms varies, they provide no formal guarantees relative to an inference calculus, and they cannot place individuals with no observed assertions. We present ELK-Em, a novel geometric ontology embedding model that embeds (i) concepts as axis-aligned boxes; (ii) roles with a new parametrisation that allows coverage of every normalised ℰℒ⁺_⊥ axiom form, yielding, at zero loss, a model of the ontology and satisfaction of all of ELK’s inference rules, with role chains enforced by a stronger sufficient condition; and (iii) individuals as points regularised using domain-specific encoders. On GALEN, GO, and Anatomy, ELK-Em ranks the atomic subsumptions ELK derives above non-entailed candidates more accurately than geometric baselines. On CAFA5 human protein-function data it also predicts GO concept assertions for held-out proteins from ESM-2 embeddings; nearest-neighbour transfer scores higher, a gap we locate in box-membership scoring rather than placement. Code and a technical report with full proofs are available at our repository.

Paper

Knowledge graph embeddings map symbolic graph structures into continuous vector spaces that can be processed by machine learning models. However, it remains unclear to what extent the resulting subsymbolic representations preserve meaningful aspects of the symbolic structure encoded in the original knowledge graph. This work investigates whether symbolic information grounded in information theory can guide embedding generation and improve the alignment between symbolic and subsymbolic representations. We consider two Information Content measures that quantify entity informativeness from complementary structural perspectives, based respectively on occurrence frequency and on the rarity of local graph features. Information Content is used as a symbolic relevance signal to control which graph contexts are more prominently considered during embedding generation. We introduce two weighting strategies: one following the conventional information-theoretic view by emphasizing highly informative contexts, and another shifting the focus toward moderately informative ones. The weights produced by these strategies are integrated into RDF2Vec by biasing random-walk transition probabilities, thereby controlling which symbolic graph contexts exert greater influence on the learned representations. We evaluate five embedding configurations on FB15K-237, NELL-995, and WN18RR by comparing neighborhoods induced by graph-based features with those emerging in the embedding space. Their agreement is assessed using Overlap and Rank-Biased Overlap at multiple cutoff values. The results show that Information Content-based weighting can improve the alignment between symbolic and subsymbolic neighborhoods, although its effectiveness depends on the dataset and weighting strategy.

Paper

Knowledge Graph Embedding (KGE) models for Knowledge Graphs (KGs) can be used to predict missing links between nodes in a knowledge graph. Link prediction quality depends on the integrity of these vector representations, as even small modifications can lead to significant changes in predictions. The total size and the number of vector representations grow with the number of entities in a KG. As such, methods that can compress and approximate KGEs are beneficial. Additionally, methods that prune the search space by reducing the number of entities considered for a query are of great interest. In this paper, we study the possible extent of approximation of entity embeddings and its influence on link prediction across two methods: (1) a vector quantization method that compresses individual entity embeddings into compact integer representations and (2) a triple-based partitioning method that groups entities defined by shared neighbors to create representative aggregate embeddings. We evaluate both methods on a number of standard KGs and KGE models. We measure the change embeddings undergo using two similarity measures and analyze the relationship with link prediction performance. Our empirical results yield concrete thresholds, as measured by the Euclidean distance and signal-to-noise ratio, prescribing how much embeddings can be modified before downstream task accuracy degrades. Additionally, we show that these thresholds can be overcome when using grouping methods such as triple-based partitioning.

Paper

The explainability and interpretability of learned latent spaces remain important and active avenues of investigation. In this work, we examine the effect of various materialization techniques that we hypothesize rigorously and consistently impact learned representations of knowledge graphs. In particular, we examine the effect of materializing the transitive closure for various relations over numeric literals-as-objects. We also consider the impact of subsumptive windowing techniques. We test these theories through a series of controlled experiments over synthetic knowledge graphs (SKGs) representing persons and their ages. Our results demonstrate consistent ordering in the latent space, and the capacity to learn mappings across different monotonic number distributions.

Paper

AI applications across many domains increasingly draw on an emerging paradigm that combines two complementary approaches: symbolic and sub-symbolic AI. From the symbolic side, Knowledge Graphs (KGs) encode complex domain knowledge in formal graph structures. To make this knowledge usable by sub-symbolic methods that uncover latent patterns, KGs must first be embedded into low-dimensional vector spaces via KG Embeddings (KGEs), which serve as a bridge between symbolic structures and vector representations and are a pivotal component of modern AI systems. Nevertheless, a critical gap persists: how KG characteristics (e.g., expressivity) shape KGE performance remains poorly understood. So far, KGE research has relied on a small set of benchmarks with limited structural diversity, capping performance especially for specialized-domain KGs (e.g., medical KGs) with distinctive properties. The DORSET project (https://dorset-project.org/) aims to address this gap from two aspects: (1) building a systematic, experimentally grounded understanding of how KG characteristics shape KGE performance, and (2) using these insights to advance both KGE and knowledge engineering (KE) research. The project pursues this through iterative, cross-disciplinary collaborations between KGE and KE experts. Expected outcomes include diverse benchmarks, KG-characteristic-aware embedding methods (KGE), and new KG engineering methodologies and tools (KE), all with broader impact on AI applications of societal value.

Paper

While Large Language Models (LLMs) demonstrate impressive retrieval of isolated facts, their ability to maintain logical consistency across complex relational paths and temporal shifts remains under-evaluated. This paper introduces a novel benchmarking framework that utilizes Wikidata to perform automated fact-checking of LLM outputs. We focus on four core reasoning dimensions: spatial traversal, temporal sovereignty, relational succession, and quantitative aggregation. By grounding our "Ground Truth" in the Wikidata knowledge graph, we expose systemic factual incorrectness and reasoning failures. We evaluate multiple local LLMs across varying scales (4B, 7B, 9B, 27B) demonstrating that while models are factually aware, they frequently fail at tasks that require combining facts across time, space, or multiple reasoning steps. While larger LLMs seem to be more robust, the model size alone does not guarantee consistent improvements. We also find that evaluation method matters: fuzzy matching reveals many partially correct answers missed by exact string matching, whereas off-the-shelf LLM-as-judge frameworks appear less reliable for computing precision, recall, and F1 in settings that require structured factual comparison and numerical correctness.

Paper

While emerging regulations mandate strict transparency for AI systems, modern Machine Learning Operations (MLOps) pipelines remain primarily sub-symbolic, managing opaque neural artifacts, raw data, and numerical metrics. This presents a challenge for automated regulatory oversight. In this paper, we present an implementation of a Semantic MLOps Engine that establishes a neuro-symbolic lifecycle for AI development. By dynamically tracking sub-symbolic ML processes and instantiating them into a formal Knowledge Graph based on the AIDOCAP ontology, we bridge the gap between model execution and symbolic governance. Furthermore, we demonstrate a Completeness Dashboard that evaluates the instantiated graph through ontology-mediated SPARQL competency queries to compute documentation completeness. Using a real-world energy forecasting scenario, we illustrate how our approach automates the acquisition and validation of documentation evidence, and makes the residual human-provided items explicit, without disrupting the developer workflow.

Paper

Before a Knowledge Graph (KG) is compiled into a hard safety constraint for a reinforcement learning (RL) agent, its classification logic needs to be validated against more than its own internal consistency. We take up this question in a safety-critical setting: mass casualty incident (MCI) triage. Our system couples a Deep QNetwork (DQN) agent with a reasoner-validated Web Ontology Language (OWL) / Semantic Web Rule Language (SWRL) ontology encoding the Simple Triage and Rapid Treatment (START) protocol and its chemical, biological, radiological, nuclear (CBRN) decontamination extensions. We first cross-validate two inference paths within the ontology (DL classification and SWRL-based inference, both grounded in the same START encoding), then check its base classification logic against Syn-STARTS [1], an independently constructed synthetic benchmark generated by a large language model that played no role in authoring our ontology. Agreement is perfect (630/630, 100.0%) on every scenario whose four required features could be mechanically and unambiguously derived from its structured fields. Only once the ontology has passed both checks do we compile it into a per-state action mask that structurally prevents the RL agent from selecting clinically contraindicated actions, rather than penalizing unsafe behavior after the fact. Across an exhaustive sweep of all 120 enumerable patient profiles and three random seeds, the constrained agent (KG-DQN) achieves a 0.0% ± 0.0 violation rate and 100.0% ± 0.0 triage accuracy, against 92.5% ± 1.8 accuracy and 72.2% ± 13.4 violations for an unconstrained DQN baseline: the safety constraint costs nothing in accuracy. We also examine, through per-step action traces, how the agent behaves on the rarest patient category - the natural stress test for a masking-based guarantee.

Paper