Knowledge graph matching is a critical task in data integration that consists of ontology matching and instance matching, i.e., finding correspondences between classes, properties, and instances from different knowledge graphs. Existing approaches for this task often rely on lexical similarities between instance attributes and separate ontology matching from instance matching, leading to suboptimal results. To overcome these limitations, we propose TIM - a tiered iterative knowledge graph matching architecture that leverages the inherent connectedness of knowledge graphs to iteratively find new correspondences. The tiered iterative structure enables the preferred use of high-precision matchers, with recall-oriented matchers running only after the more precise matchers cannot find any new correspondences. We evaluate our approach against several state-of-the-art approaches on multiple real-world datasets. We achieve the best accuracy for class and property matching by incorporating instance information into their matching process and an above-average accuracy for instance matching. Additionally, due to the efficient architecture of our approach, we achieve the fastest runtime for non-trivial baseline systems.
Artificial Intelligence (AI) methods based on Knowledge Graphs (KGs) have recently been proposed to generate explanations for scientific discovery tasks. Although these approaches offer a promising foundation for accelerating research and development, their practical adoption depends on a clearer understanding of which factors influence the usefulness and trustworthiness of the explanations they generate. We conducted an in-depth user study with eleven biomedical researchers on KG-based explanations for two key scientific discovery tasks in drug development: drug repurposing and drug–target interaction. We evaluated the relevance, completeness, and validity of path-based explanations extracted from KGs by state-of-the-art methods, and compared them along two dimensions: whether they produce single vs. multiple explanatory paths, and whether they incorporate ontological information. The evaluation combined participants’ ratings with a qualitative analysis of feedback to understand how experts interpret these explanations. Our results revealed that biomedical researchers prefer explanations that integrate ontological information and present diverse, biologically plausible mechanisms over those based solely on structural connectivity. These findings shed light on key design choices for KG-based AI systems that aim to support informed candidate selection in drug development and, more broadly, improve their impact in scientific applications.
Knowledge graphs (KGs) represent heterogeneous data as triples, including relational triples between entities and attribute triples linking entities to literal values like labels or numbers. Integrating multiple KGs is an essential task for maximizing data utility in downstream tasks. However, the integration process is challenging due to common issues such as noise, incompleteness, and varying coverages, which complicate accurate entity linking and can lead to misclassifications. As the size and the number of available datasets continues to grow, automated approaches to KG integration are indispensable for scalability. Existing methods often rely on KG embedding techniques and struggle when confronted with noisy or contradictory literal data. In this paper, we present Conel, a novel approach for unsupervised link discovery that leverages literal similarities and employs contrastive and entropy based scoring for robust entity matching. Our experiments show that Conel outperforms all baselines on 5 out of 7 datasets exhibiting robustness to the adverse conditions typical in real-world KG linking.
Question and feedback generation from ontologies has appeal for the high-quality content it can provide, and to alleviate the test setting and marking burden of teachers and quiz question setters. Existing approaches lack a formal and expressive technique, however, which limits their ability to represent and process complex and semantically rich questions, answers, and explanatory feedback patterns. They typically require post-processing to eliminate irrelevant outputs, rely on entity labels rather than represented semantics, and they often lack modularity, reducing scalability and reusability. We solve this by focusing on the content determination stage in the natural language generation pipeline, and any realiser of choice to be used afterwards. We propose a framework consisting of an OWL 2 DL-based model combined with an optimised determination algorithm to efficiently fetch relevant content from the ontology. The framework was implemented and evaluated on content selection performance with several ontologies.
Knowledge Graphs (KGs) are essential for organizing and retrieving knowledge from large text collections, yet constructing high-quality KGs from text remains a difficult task. An important challenge in constructing KGs is evaluation: how can we measure the quality of the graph under construction? Traditional evaluation relies on expensive, manually curated gold-standard graphs whose creation demands significant expert effort and is rarely available for every domain. To alleviate this cost, we need unsupervised metrics that compare generated graphs directly with the source texts. Unfortunately, most proposals in this direction require training or fine-tuning machine learning models, introducing additional costs that limit their applicability. To address this gap, we develop a fully unsupervised metric that compares graphs directly with the source texts. Our proposal, called Bidirectional Graph Similarity (BIGS), transforms each triple into natural language sentences and measures semantic alignment with original document sentences using embedding similarity. We evaluate BIGS on several document collections and standard benchmarks for supervised evaluation, finding strong correlations with supervised metrics, and that it remains robust across different text segmentation and verbalization strategies. Our results show that BIGS effectively captures the coverage of generated graphs without any manual annotations or additional training, offering a practical tool for KG evaluation.
Embedding models (KGEMs) constitute the main link prediction approach to complete knowledge graphs. Standard evaluation protocols emphasize rank-based metrics such as MRR or Hits@K, but usually overlook the influence of random seeds on result stability. Moreover, these metrics conceal potential instabilities in individual predictions and in the organization of embedding spaces. In this work, we conduct a systematic stability analysis of multiple KGEMs across several datasets. We find that high-performance models actually produce divergent predictions at the triple level and highly variable embedding spaces. By isolating stochastic factors (i.e., initialization, triple ordering, negative sampling, dropout, hardware), we show that each independently induces instability of comparable magnitude. Furthermore, for a given model, hyperparameter configurations with better MRR are not guaranteed to be more stable. Moreover, voting, albeit a known remediation mechanism, only provides a limited enhancement of stability. These findings highlight critical limitations of current benchmarking protocols, and raise concerns about the reliability of KGEMs for knowledge graph completion.
Ensuring the veracity of assertions is vital for building reliable and consistent knowledge graphs. A variety of automatic fact-checking approaches have been proposed over the past decade. Among these, path-based fact-checking approaches are particularly attractive due to their independence of supplementary external knowledge and their faster runtimes compared to methods reliant on external corpora or embeddings. However, the effectiveness of these approaches is fundamentally limited by the incompleteness of existing knowledge graphs, which often lack the paths necessary to support or refute assertions. To address this limitation, we propose ShallKnow, a framework that supplements the knowledge graph with shallow knowledge—automatically extracted RDF assertions from external unstructured sources—even if this additional knowledge may not always fit a well-defined ontology nor be fully verified. By appending such shallow knowledge, we enhance the graph’s coverage and increase the chances of finding relevant evidence for fact-checking. Comprehensive experiments on three widely used benchmark datasets demonstrate that integrating ShallKnow consistently and significantly enhances the performance of state-of-the-art path-based fact-checking approaches, yielding improvements of up to 0.24 in Area Under the Receiver Operating Characteristic Curve (AUROC). These results establish ShallKnow as a broadly applicable auxiliary component for improving the reliability and coverage of automatic fact-checking in knowledge graphs. Our code is open-source and can be found at https://github.com/dice-group/ShallKnow .
The increasing availability of heterogeneous cultural heritage (CH) data calls for semantic, scalable, and operational infrastructures capable of integrating symbolic knowledge with multimodal representations. Existing CH knowledge graphs (KGs) are typically built from individual institutions’ datasets and focus predominantly on semantic modelling. They seldom incorporate the results of AI methods, such as visual representations or neural retrieval methods, which limit their ability to support cross-collection interoperability and advanced multimodal retrieval and sustain continuous, automated data enrichment. In this paper, we introduce ArtKB: a modular end-to-end knowledge base (KB) that unifies semantic modelling, multimodal representation, and operational access to artefacts. Built from Wikidata and annotated using the CACAO ontology, the proposed system combines an RDF graph, a vector database for visual embeddings, and an object storage for digital assets, all orchestrated through a unified API gateway. This infrastructure enables hybrid and multimodal retrieval, metadata enrichment and automated text-to-graph generation, demonstrating how neuro-symbolic approaches can enhance CH data management. The paper details the architecture of the domain-specific knowledge base, demonstrating how the different components integrate to support different tasks. The resulting KB offers a scalable, interoperable, and domain-agnostic solution that can be adapted to broader CH contexts and other multimodal domains.
Data Sharing Agreements (DSAs) remain largely unstructured, which hinders consistent interpretation, completeness checks, and reuse. This work introduces a general-purpose Data Sharing Agreement Ontology (DSAO) aligned with the Data Privacy Vocabulary (DPV) and an explainable validation pipeline that integrates SHACL-based structural checks with a guided human-in-the-loop review schema. The proposed approach translates guideline requirements into reusable ontology patterns and SHACL profiles, maintains traceability from requirements to patterns to shapes to competency questions, and records reviewer outcomes as data that can be distilled into warning-level hint shapes. The approach is evaluated on a synthetic corpus of 100 DSA graphs with controlled missing-element defects, measuring SHACL defect detection, competency-question answerability before and after repairs, and the extent to which hint shapes can pre-screen reviewer flags.
Traditional SPARQL provenance has focused on explainability, while data sharing today increasingly requires privacy preservation and cryptographic verification. We thus introduce zkRDF, a data-centric approach that enables a data holder (the prover) to guarantee soundness of SPARQL query results to a data consumer (the verifier) while exposing only the minimal information needed for the proof. Unlike existing methods that prove query execution, we establish soundness of query results by proving properties about the queried RDF dataset. Given a consumer’s query, the holder materializes a selectively disclosing view of the queried dataset, revealing required information and cryptographically hiding the remainder. Zero-Knowledge Proofs (ZKPs) guarantee both the integrity of the derived dataset and the adherence of hidden elements to desired constraints, such as numeric bounds. The consumer verifies the proofs and obtains the desired query results from the dataset. Importantly, proof verification ensures dataset validity and, therefore, guarantees sound query results. We present zkRDF’s methodology, detail the interpretation of SPARQL queries to produce sound results, and prove soundness of our approach. We discuss zkRDF’s support for SPARQL features and show that our proof-of-concept implementation outperforms an approach that proves query execution by three orders of magnitude.
Collaborative Knowledge Graphs (KGs) rely on evolving, “soft” constraints, making automated repair a challenge for rigid symbolic methods. In this paper, we propose a structure-aware neural framework that learns to repair violations directly from historical edit patterns. We perform a systematic study of graph encoding strategies, specifically comparing flattened predicate-as-node representations against multi-relational graphs equipped with a custom dual-representation module. We benchmark these encodings across different GNN backbones and feature initializations. Our experiments show that the Multi-Relational GIN yields the most robust performance, surpassing strong symbolic baselines by over 29
Visual Question Answering (VQA) is a challenging task that demands not only accurate alignment between images and language, but also multi-step reasoning, contextual understanding, and the ability to incorporate external knowledge, especially in multi-turn settings where follow-up questions depend on previous dialogue. In this work, we present a novel framework for generating knowledge-grounded, multi-turn VQAs datasets that has been integrated into the IBM Granite-Vision development pipeline. The main novelty of our method is the generation of multi-turn conversations using large language models (LLMs), but grounded in knowledge-driven prompting: we leverage structured and unstructured knowledge sources from Wikipedia articles, associated images, and the Wikidata knowledge graph (KG). By combining both unstructured and structured knowledge sources, our approach advances VQA beyond shallow perception tasks toward more profound, knowledge- and entity-aware reasoning. We demonstrate the effectiveness of this approach by using it to fine-tune and evaluate existing vision-language models (beyond the Granite-Vision models), and share valuable insights about the complexity of the task and the nature of available benchmarks.
Semantic parsing-based methods for knowledge base question answering have achieved leading performance, but rely on high-quality training data of question-SPARQL query pairs, which requires substantial manual effort. In contrast, question-answer pairs are much easier to obtain. In this paper, we propose a novel task, namely SPARQL Query Annotating (SQA), to automatically construct queries with given question-answer pairs. We propose a novel metric test suite score for the evaluation of this task, and collect a dataset ReQuMA to evaluate this task’s facilitation on manual annotation. We also present QuAD, a Question-Answer Driven method for this task. Our experiments show that the task effectively eases manual annotation, and helps semantic parsing-based KBQA methods to maintain competitive performance when using the constructed queries from QuAD as training data. Analysis shows our test suite score metric effectively reflects the quality of constructed queries. Meanwhile, the substantial performance of QuAD demonstrates its effectiveness as a dedicated SQA method. Our code is available at https://github.com/nju-websoft/SQA .
Recent advances in ontology matching have increasingly relied on language models to capture the lexical information of entities. However, this focus on surface terminology often overlooks the formal semantics encoded in the ontological structure and limits the interpretability of alignment decisions. We introduce EXACT-OM, a context-aware model that integrates ontology-derived semantics directly into the alignment scoring process. It integrates lexical similarity, contextual similarity based on informative relation-specific subgraphs, and language model-based signals within an adaptive scoring framework. The system is intrinsically interpretable: it decomposes every decision into component contributions, provides per-triple importance via perturbation analysis, and renders graph-based visual explanations with concise natural-language summaries. On the OAEI Bio-ML benchmark, EXACT-OM achieves competitive performance while delivering fine-grained, auditable explanations. In a user study (n=12), the majority of users preferred EXACT-OM explanations to more traditional tools, and they were found to substantially improve validation efficiency for high performing users, enabling faster decision-making without compromising accuracy. Importantly, both performance and preferences differed across users, suggesting that explanation effectiveness depends more on individual strategies than on domain knowledge or technical skill. Overall, this shows that the principled integration of semantic structure with language models can enhance explainability and expert validation workflows without sacrificing performance.
Automated Knowledge Graph Completion (KGC) is a crucial task in the semantic web community, focused on discovering missing structured information within Knowledge Graphs (KGs), such as identifying links between entities or classifying relation types. The continually improving performance of pre-trained Large Language Models (LLMs), with their inherent ability to learn world knowledge, has shown significant promise for KGC. However, as KGs evolve with new factual knowledge, continual fine-tuning of such models becomes necessary, making them vulnerable to catastrophic forgetting and incurring significant computational and storage overhead. In this paper, we propose a dynamic Parameter-Efficient Fine-Tuning method that introduces a novel and effective combination of masking and growing strategies for continual KGC, addressing a core limitation of existing methods: static architectures that suffer from parameter saturation over time. Our method enables LLMs to continually adapt to evolving KGs while preserving previously acquired knowledge. It supports knowledge transfer, mitigates catastrophic forgetting, and incrementally expands model capacity as needed. Evaluation is conducted in two continual learning settings, task-incremental learning for link prediction and class-incremental learning for relation extraction. Experimental results show that the proposed method outperforms both rehearsal-based and rehearsal-free baselines, offering an effective and scalable solution for continual KG modeling.
A bounded description of a resource is a useful default subgraph that represents relevant information. When accessing RDF knowledge graphs, bounded descriptions are returned for SPARQL DESCRIBE queries or for HTTP GET requests on Linked Data. Concise Bounded Description (CBD) is the de facto default, but only provides conditional reachability of connected resources, limiting its usefulness for knowledge graph traversal. This paper outlines alternative bounded description specifications consisting of chained and parallel combinations of construction steps. Furthermore, characteristics of bounded description specifications are introduced, including ontology awareness, applicability, and reachability. The evaluation challenges the traversal of a knowledge graph with n-ary relations using retrieved bounded descriptions. Results confirm limitations of the CBD specification, highlight the impact of reachability, and bring attention to ontology-aware and set-specific bounded descriptions.
The rapid growth of the Internet of Things (IoT) has created vast networks of interconnected devices generating continuous data streams, offering opportunities for real-time analytics across domains like smart cities, healthcare, and transportation. On the one hand, traditional knowledge representation methods provide interpretability, precise semantic modeling, and structured reasoning but often struggle with scalability, making predictions, and handling uncertainty. On the other hand, machine learning models offer powerful predictive capabilities and efficient handling of large, uncertain datasets, but typically lack interpretability, adaptability, and require extensive labeled data. Neuro-symbolic AI tries to use the best of both symbolic reasoning and machine learning—combining interpretability and structured inference from symbolic approaches with the predictive power and generalization capabilities of machine learning. Existing neuro-symbolic approaches, such as DeepProbLog and NeurASP, although successful in general logical programming contexts, they do not natively support streaming data or querying over RDF. This paper introduces Kolibrie, a hybrid stream reasoning engine developed in Rust that integrates both at the language and system level (1) SPARQL for querying, (2) Datalog-style rules for symbolic inference, (3) RSP-QL semantics for RDF stream processing and (4) machine learning predictions. Resource type: Software framework License: Mozilla Public License Version 2.0 DOI: https://doi.org/10.5281/zenodo.17808656 URL: https://github.com/StreamIntelligenceLab/Kolibrie
Although time series analytics plays an important role across diverse application domains, efficiently managing the resulting insights remains a significant challenge. While specialized ontologies structure information of analytical models, they provide limited support for standardizing and arranging inferred knowledge. The absence of a unified data model to categorize findings such as anomalies, trends, or patterns complicates reuse and inhibits synergy effects between subsequent utilization stages. This paper presents the Semantic Time Series Ontology - SemTS, an ontology designed to classify time series characteristics as explicit knowledge entities, facilitating their consistent and semantic representation. The associated integration of related concepts through specially created vocabularies improves the dissemination of insights across various abstraction levels. SemTS further enables the description and incorporation of scenario-specific information, including domain expertise, to enrich analytical contexts. To demonstrate its practical utility, we showcase various aspects of SemTS through competency questions that illustrate how the ontology can be employed to efficiently query and validate semantic time series information. By systematically combining inferred knowledge and predefined facts, SemTS provides a comprehensive framework for improving the reusability and integration of insights affiliated with time series data.
The widespread digitization of cultural heritage (CH) collections has enabled institutions to make vast artefacts accessible through knowledge graphs (KGs). Yet existing retrieval systems predominantly rely on keyword-based search with semantic enrichment via linking to controlled vocabularies, while only a few systems support visual information search, and none support queries involving relational metadata constraints such as temporal ranges, or combinations of metadata and visual information. Although advanced approaches, including multimodal retrieval and knowledge reasoning, have emerged in research, few have been deployed in operational CH environments, creating a substantial gap between academic innovations and real-world practice. In this paper, we present a knowledge-enhanced multimodal retrieval system that operates over KG, integrating a domain-adaptive fine-tuned CLIP model with an LLM-based Text2SPARQL and unifying their results through a fusion strategy. We deployed the system as a backend API integrated into a CH platform and conducted a comprehensive evaluation through quantitative metrics and qualitative user studies with museum curators. Results demonstrate that our domain-adaptive CLIP achieves high retrieval accuracy, while Text2SPARQL effectively further enhances precision through knowledge reasoning. Our work provides practical experience into the benefits and limitations of these technologies in production environments, informing broader adoption across CH institutions.
Document-level relation extraction is an essential task in natural language processing. Achieving high performance on this task is challenging due to the complexity of reasoning over long texts. While large language models have shown promising results in natural language processing, their effectiveness in document-level relation extraction remains limited. Techniques such as chain-of-thought and few-shot prompting offer insufficient improvements for real-world applications. Reinforcement learning with human feedback (RLHF) has proven beneficial in enhancing large language model capabilities; however, its reliance on manual annotation makes it costly and impractical at scale. To address these challenges, we introduce RLKGF, an approach that incorporates knowledge graph feedback into RLHF and replacing human feedback with knowledge graph feedback. Rather than integrating knowledge graphs into the learning process, RLKGF utilizes them solely as a source of feedback for the document-level relation extraction, reducing reliance on manual supervision while lowering annotation costs and improving feasibility. The proposed method complements existing RLHF techniques and can be easily incorporated with them. In our evaluation on two benchmark datasets, our method outperforms both traditional document-level relation extraction and large language model-based approaches, achieving F1-score improvements of 10
