The rapid development of educational technology has significantly expanded the range of academic degree programs, associated courses, and corresponding learning resources. This expansion has led to substantial overlaps in course content across programs, making it increasingly difficult for learners to construct coherent and efficient learning paths. Educators must also adapt their materials more effectively to meet the diverse needs of their students. A personalized course and curriculum path recommendation system offers a promising solution by guiding both students and instructors in selecting relevant content and structuring appropriate learning trajectories. Among the techniques available, knowledge graphs (KGs) are particularly effective. By representing information as entities and their interrelationships, KGs provide a structured, semantically meaningful structure for organizing educational data. This conceptual paper introduces a Knowledge-graph driven recommender framework that supports the generation of personalised learning paths aligned with individual learners prior knowledge and goals, while enabling instructors to adapt curriculum design and instructional delivery to identified learner needs. By semantically linking learner’s background knowledge to relevant curricular components by leveraging placement assessments, the framework supports individualized assessments and recommends adaptive learning trajectories. The work outlines its theoretical foundations and explores its potential through relevant literature and graph-based design principles, demonstrating how structured semantic representations support goal-aligned, personalized learning. Through this approach, systems based on the framework have the potential to reduce content redundancy, enhance coherence, and improve the quality of teaching and learning, while fostering a more dynamic and responsive educational environment.
AI technologies play an important role in diagnosing the performance of a service system and ensuring its continuous operation (AIOps) by detecting anomalies, localizing root causes, and adapting system configurations. Large language models (LLMs) possess broad knowledge of computing systems and offer new opportunities for effective root cause analysis and system adaptation. However, their inference capabilities for complex problems remain challenging. In this abstract, we discuss how a knowledge graph-based representation of performance anomalies, root causes, and adaptation strategies can be used to enhance LLMs in AIOps through an agentic system model.
This paper presents ABEL (Artificial Buddy for Effective Learning), a modular, knowledge-graph-driven chatbot designed to enhance and support education in Data Science and Artificial Intelligence. At the core of ABEL, is a hybrid retrieval architecture that integrates a dynamic Knowledge Graph and a Retrieval-Augmented Generation (RAG) pipeline. The knowledge graph, constructed from curated learning resources, enables a concept-driven retrieval of semantically relevant and specific educational content through multi-hop graph queries and embedding-based similarity search. This approach enhances the contextual grounding and supports the generation of personalized, specific, and explainable responses by Large Language Models (LLMs). ABEL is also complemented by a Frequently Asked Questions (FAQ)-based RAG approach, thus offering flexible access to learning content while ensuring traceability and correctness. We present the system’s architecture, evaluate its performance using both retrieval and user-based metrics, and discuss the benefits of combining symbolic graph structures. Our results demonstrate that this approach can significantly improve the relevance and adaptability of chatbot-driven learning platforms.
There is a recent trend of using LLMs for generating SPARQL queries from natural questions, in order to exploit the generated query for retrieving the answer over a Knowledge Graph (KG). This can be quite challenging for popular domains, such as sports, that include billions of fans worldwide that are not familiar with KGs and query languages, and they typically express their questions in natural language. In this paper, we focus on basketball games which are complex events including several factors and participants, such as teams, players, statistics and others. However, all this information is usually given in separate HTML tables or JSON files, thereby, it is not trivial to be combined. The objective is to show a generic pipeline including the following steps: i) data collection, ii) the definition of competency questions, iii) the creation of an event-based ontology and the KG, and iv) the exploitation of the KG through the LLMs, for enabling natural QA through a text-to-SPARQL method based on ontology and literals path patterns. As a use case, we construct a KG including 5 million triples from the EuroLeague games from 2000–01 until 2024–25, a web application, and an evaluation benchmark with 100 natural questions in 3 languages (English, Greek, Chinese). Finally, we present an experimental evaluation for the QA task; indicatively by using the proposed method, we reached 0.79 accuracy through GPT-4 and 0.71 accuracy through DeepSeek, for the English benchmark, instead of 0.3 for the baseline method.
Recently, Large Language Models (LLM) have been used to generate structured graph data from unstructured text. However, parts of the extracted nodes lack information and retraining is necessary whenever the graph changes. We propose an unsupervised approach based on embeddings similarity, to link entities extracted by a LLM from unstructured text data to nodes of a knowledge graph, generated according to an ontology. Tested on three datasets from a maritime cyber security use case, results indicate an accuracy improvement from 86.2
Massive volumes of raw astrophysical data are collected every day, and subsequently processed into public catalogues used for scientific research. While deep learning has accelerated astrophysics in recent years, the field has yet to benefit from the Semantic Web’s advances in data representation and management. As a result, most of those catalogues are typically stored as flat CSV/ASCII files in a tabular format, limiting their interoperability and reuse. In this work, we introduce a semantically enriched astrophysical catalogue, in the form of an ontology-backed Knowledge Graph (KG). Our KG is based on a custom, lightweight ontology (DOCBO), and was populated by transforming the Heraklion Extragalactic Catalogue (HECATE) into RDF using two alternative mapping approaches, X3ML and RML, whose characteristics and trade-offs are discussed and compared. Both the mapping rules and the resulting Knowledge Graphs are openly available to foster reuse, interoperability and community feedback. This work, which is part of the interdisciplinary PARSEC project, constitutes a first proof-of-concept effort towards a more generic methodology for representing astrophysical knowledge of various subdomains (e.g., Supernova Remnants and transients) using semantic technologies, thereby forming and automatically populating semantically-enriched KGs based on appropriately defined ontologies.
Datasets in the energy domain often encounter significant challenges in achieving data interoperability due to semantically equivalent concepts being inconsistently named. Semantic annotation addresses this issue by standardizing concepts to enhance data interoperability and facilitate efficient data integration. Traditional annotation methods, however, rely heavily on static resources, which limit their ability to dynamically understand cross-domain context. Large Language Models (LLMs) have shown promise in capturing semantic similarities between terms, and their capabilities can be enhanced by incorporating ontological reasoning. In this paper, we propose a novel framework that leverages LLMs for ensemble decision-making to automate and improve semantic annotation. By combining LLM-based semantic understanding with structural reasoning from ontologies, our approach enhances reliability and annotation performance. We evaluate its effectiveness using both proprietary and open-source models on real data from industrial companies in the energy sector. The experimental results demonstrate superior annotation accuracy compared to baseline methods. This study highlights the potential of integrating ontological reasoning, LLMs, and ensemble decision-making to advance semantic data integration in the energy domain.
The Virtual Knowledge Graph (VKG) paradigm enables querying heterogeneous relational databases through a unified semantic layer comprising an ontology and declarative mappings (typically in R2RML). While querying is well-supported, propagating updates from the virtual RDF graph (ABox) back to the source databases remains a challenge. This paper addresses the problem of translating updates (expressed in SPARQL Update) applied over the ABox into equivalent SQL updates over the underlying databases, specifically within the Ontop VKG system. Key difficulties arise from the non-injectivity inherent in R2RML mappings, where a single ABox update can correspond to multiple source update possibilities, and the potential for these source updates to cause unintended side effects—additional insertions or deletions in the ABox beyond the user’s original intent. While relying on Ontop’s query rewriting engine, our method employs lineage computation to identify source tuples for deletion and a strategy for handling existential variables during insertion. It generates candidate SQL translations, analyzes their potential side effects on the virtual ABox, and selects the ones that minimize these unintended consequences. This work represents a step toward closing the Linked Data Life Cycle loop, allowing changes in the knowledge graph to be reflected in the corresponding source.
The text-to-SQL task translates natural language questions into SQL queries, simplifying database access. While large language models (LLMs) have shown strong performance, they often struggle with complex reasoning, such as commonsense and numerical reasoning, required for more challenging SQL generation. We propose a new pipeline that enhances SQL generation by incorporating advanced reasoning skills, alongside techniques like entity linking and self-correction. Tested on the Archer dataset, which requires more complex reasoning, our approach improves performance by 32.8% over the baseline, demonstrating its effectiveness in handling challenging queries.
Interoperability between heterogeneous graph data models remains a significant challenge in knowledge graph construction. While both Property Graphs (PG) and RDF models are widely used for data representation, their structural and semantic differences hinder seamless integration and data exchange. In this work, we present PG2RDF, a fully automated and schema-aware pipeline for transforming PG data into RDF. Our method relies on an XML Schema Definition (XSD) to formalize the PG structure and can generate both X3ML or RML mapping rules for deterministic RDF production. By integrating schema discovery, XML annotation, and RDF materialization through the X3ML Engine and RML Mapper, we demonstrate the feasibility of our approach across heterogeneous datasets and their integration into an interoperable Knowledge Graph. PG2RDF ensures URI consistency and accurately represents the PG information.
Virtual Knowledge Graph (VKG) is a well-established framework in which users can access a relational data source through an ontology and declarative mappings. VKG systems traditionally assume uniform access rights for all users, an assumption that does not always hold in real-world scenarios involving diverse user roles and sensitive information requiring protection. Controlled Query Evaluation (CQE) provides a privacy-preserving framework by enforcing policies that define confidential information and implementing censors to prevent policy violations. However, it does not account for differences in user privileges during query answering. To address this gap, we extend the Policy-Protected VKG (PPVKG) framework, which embeds CQE policies into VKG mappings, by enabling role-sensitive query answering. Specifically, we incorporate Role-Based Access Control (RBAC) into PPVKG, by associating to each user role a specific set of policies, and ensuring that during query evaluation, only the policies relevant to the user’s role are applied. We validate our RBAC enhanced PPVKG approach using the MIMIC-III critical-care database, mapped to the Fast Healthcare Interoperability Resources (FHIR) ontology. Our experiments, conducted with the open-source VKG system Ontop, demonstrate effective policy enforcement with RBAC.
Although Quantum Computing (QC) is subject of intensive research, the data/knowledge management perspective is largely unexplored. Here we investigate how QC could be used for various tasks related to Knowledge Graphs (KG). We discuss the placement of QC in KG-related workflows, and then we (a) propose encodings of RDF (Redource Description Framework) data appropriate for QC, (b) show how a KG could be stored as a quantum state, (c) introduce a method for building indexes that is based on quantum circuits, (d) discuss how several important KG-related tasks, like reasoning, can be reduced to satisfiability, and thus gain from the quantum speedup. Subsequently, we report experimental results regarding encoding, decoding, superposition, as well as the limitations of the current QC hardware.
Database triggers have been introduced since the 1980s, and their value has been widely recognized. They allow for automating several tasks and enforcing business rules at the database level, ensuring data integrity, auditing, and consistent logic across applications. However, notwithstanding their high value, the use of triggers in mainstream triple stores has been mostly overlooked. The dynamic nature of semantic data, the increasing scale of RDF datasets, and the growing demand for real-time processing in graph-based applications demand efficient methods for real-time data validation and integrity enforcement, change tracking and auditing, and automated inference and reasoning. In this demonstration, we recognize triggers as first-class citizens in the semantic web and introduce a new interpreted language to describe them. We present 3T, an integrated development environment (IDE) and a graphical user interface (GUI) enabling triple-store-independent triggers, allowing their real-time, effective, and efficient execution.
Public institutions seldom document the informal, undocumented processes that shape everyday work. As institutions digitize and modernize, such tacit knowledge is easily lost, particularly amid demographic change and limited process modeling expertise. To address this problem, we present Process Copilot, a generative AI system that converts natural language process descriptions into BPMN 2.0 models using large language models (LLMs). To improve the quality and reliability of generated process models, we incorporate domain-specific ontologies into the prompting process. We evaluate two prompting strategies, with and without ontologies, across ten real-world processes, using syntactic and structural similarity metrics alongside stability analysis based on output variance. Results indicate that ontology integration enhances semantic and structural performance, though at the cost of stability.
The proposed work presents a new approach to craft education that combines semantic knowledge graphs (KGs) with real-time 3D interactive simulations using the Finite Element Method (FEM). Crafting processes are modelled as KGs, outlining each procedural step, possible errors, and their consequences. Learners practise tasks in a virtual environment where their actions are checked against the KG, enabling real-time feedback and correction. After each task, the workpiece is rendered with physically based visualisation to show realistic outcomes. This method enhances learning by providing predictive guidance, error correction, and high-quality visual feedback.
In the information age, making informed decisions about food should be straightforward. Consumers increasingly seek trustworthy information on sustainability, nutrition, provenance, and the risks and benefits associated with the products they consume. However, in the case of aquafood, the intersecting complexities of health considerations, environmental sustainability, and climate impacts create a challenging landscape for both consumers and industry stakeholders. To address this, the VeriFish project introduces a robust framework of verifiable indicators designed to support informed and responsible choices among retailers, producers, fishers and citizens. This paper presents the data-driven methodology for constructing a semantic knowledge base that underpins the VeriFish indicator framework by integrating heterogeneous data sources. We outline the key challenges and requirements in building this knowledge base, describe its technical implementation, and detail the mechanisms developed to support discovery and access. Finally, we showcase dedicated applications built on top of the knowledge base that enable practical use of the indicator framework.
The exponential growth of knowledge graphs necessitates effective and efficient methods for their exploration and understanding. Frequently Asked Questions (FAQ) is a list of questions and answers related to a specific topic intended to help people understand a particular subject. In this demonstration, we present ULYSSES, the first system for automatically constructing a FAQ for large Knowledge Graphs. Our method consists of three key steps. First, we select the most frequent queries by exploiting available query logs. Next, we answer the selected queries using the original graph and finally, we construct textual descriptions of both the queries and the corresponding answers by exploring state-of-the-art transformer models, i.e., ChatGPT and Gemini. In the demonstration, we explain the various steps and allow conference participants to test the system in real time and identify its usefulness.
Knowledge Graphs (KGs) are increasingly leveraged in Cultural Heritage (CH) domains to enable structured, interoperable representations of artefacts, people, events, and places. The ISO CIDOC CRM, as a core reference ontology, facilitates semantic alignment and interoperability across diverse CH datasets. However, despite the CIDOC CRM, linking independently developed KGs remains a complex challenge. This paper investigates the process of integrating two KGs collecting knowledge about ancient manuscripts: the Mapping Manuscript Migrations (MMM) and the IMAGO knowledge graphs, both of which adopt CIDOC CRM as their underlying schema. Through this case study, we explore methodological and technical hurdles in aligning entities, reconciling modelling decisions, and preserving provenance and semantics during integration. We discuss strategies for identity resolution, semantic approximation, and metadata enrichment, and propose reusable patterns for future KG linking efforts in the CH domain.
Knowledge graphs about statistical data, such as the Eurostat Knowledge Graph (EurostatKG), present unique challenges for relation prediction due to their rich ontological schemas and strict semantic constraints. To address this, we reformulate relation prediction as a triple classification task and propose a unified, ontology-guided neural framework. Our approach integrates embedding-based models with graph-based semantic encoders (GCN, GAT), and introduces a hybrid loss function alongside a constraint-aware negative sampling strategy that respects ontology-defined domain and range axioms. Evaluated on EurostatKG, our method achieves strong results, including an F1-score of 0.96, PR-AUC of 0.99, Hits@5 of 0.618, and a Soft Hits@1 of 0.97. This work provides a principled integration of symbolic knowledge into neural models, while establishing a reproducible baseline for ontology-aware KG completion in complex, schema-rich domains.
To evaluate the potential of Quantum Computing (QC) on real data management tasks in this paper we elaborate on the task of encoding large in size Knowledge Graphs as numerical data for using them for creating the corresponding quantum state. This is a task that cannot be bypassed, and has to be performed using a classical computer. We report experimental results for datasets up to size 800 million triples, in particular for the entire DBpedia. We also report times for creating the corresponding quantum states by superimposing the encoded datasets. Finally, since a significant number of tasks are reduced to Grover’s algorithm, we report efficiency results on running this algorithm on datasets with size up to 10^8 bitstrings on a simulator and on IBM Quantum Computer. Even if the output of Grover is not stable on current quantum hardware, this analysis is useful for estimating the efficiency and thus identifying to what tasks the application of QC will be beneficial.
