Property graphs are widely used to model complex, interconnected data, but flexible property annotations often undermine consistency due to weak schema support. We introduce PG-Schema-PC, an extension of PG-Schema for constraints over property sets, covering structural, cardinality, and range constraints (including alternative groupings, multiplicity control, and value conditions). We formalize its abstract grammar and illustrate its use with practical examples.
The rapid emergence of novel psychoactive substances and evolving slang presents ongoing challenges for effective drug surveillance and public health intervention. Traditional keyword-based methods often fail to capture informal, misspelled, or newly coined terminology prevalent in online communities. In this study, we explore a data-driven approach to slang detection by analyzing Reddit posts from drug-related subreddits. Beginning with a curated list of known drug names, we collect Reddit data using PRAW, a Python-based Reddit API wrapper, and focus on high-engagement posts from sections like "hot," "new," and "top." After preprocessing the text via SpaCy for lemmatization and normalization, we identify frequent out-of-vocabulary terms not present in the seed list. To assess their potential as drug slang, we evaluate the semantic similarity between these candidate terms and known drugs using two NLP models: Word2Vec and BERT-based transformers. While Word2Vec captures surface-level morphological similarities, BERT models leverage deeper contextual understanding for phrase-level comparisons. We compare the outputs of both methods by analyzing their overlap with known drug variants and inspecting topscoring matches. Our results indicate that BERT models generally identify more semantically valid yet lexically distinct slang terms than Word2Vec. This comparative analysis highlights the relative strengths of each embedding approach and suggests transformer-based models are more effective for scalable and adaptive slang discovery from social media.
We investigate the use of Web agents that use planning to achieve goals involving preconditions on the dynamic Web. We implement our approach, AEPIS, based on Semantic Web technologies and rely on environmental information evaluated exclusively at run-time to take dynamic changes into account. We evaluate AEPIS in static and dynamic environments, based on the Web of Things, and compare its performance to a rule-based and an agent using large language models. We find that AEPIS combines adaptivity with rapid planning at scale, even in a dynamic environment.
Extracting structured knowledge from text in the form of (subject, predicate, object) triples is a key task for many artificial intelligence applications, in particular for knowledge graphs (KGs) construction. In closed information extraction (cIE) where the extracted triples are constrained by a predefined KG schema, most existing approaches rely on Wikidata for entities and relation extraction. As a result, they often lack the flexibility to adapt to other KGs without prior retraining and costly data annotation. In this paper, to address these limitations, we propose FlexCIE, an approach for cIE which leverages Large Language Models (LLMs) and a KG completeness analysis tool. Given an input text, it identifies a list of entity mentions, links them to entities in the target KG using embedding techniques combined with LLMs, and constructs triples from these extracted entities using relevant properties retrieved from the KG. This enables the use of LLMs for cIE, while ensuring that the generated triples are accurate and compliant with the KG schema. Therefore, our work contributes to making cIE more practical and flexible, particularly for domain specific or enterprise KGs.
Cultural heritage data, particularly from colonial contexts, frequently presents an incomplete and biased view, reflecting historical institutional priorities more than contemporary knowledge requirements. Consequently, knowledge graphs derived from these records often contain incomplete, fragmented, and skewed data, including absent attributes or values, missing semantic links, and under-represented perspectives. This work addresses the challenge of knowledge discovery under such limitations, presenting a real-world case study on the provenance research of colonial cultural heritage. We present a task-aware design method for building a tool to facilitate this process. The design approach of this application is rooted in a Knowledge Discovery in Database (KDD) framework and is particularly novel due to its formalisation and operationalisation of three distinct types of semantic association: explicit, abstract, and implicit. These semantic associations, grounded in domain interpretation, are crucial for bridging data gaps where user information needs cannot be directly met by existing data. We further demonstrate how these associations can be effectively communicated through user interface components, enabling users to infer new knowledge. We evaluated the resultant application through a user study among domain experts to assess its efficacy. The evaluation confirms the effectiveness of the tool in enabling new knowledge discovery and reveals opportunities to improve the representation of the underlying data, as users could successfully infer insights even when information was missing or poorly captured in the original data sets.
National intelligence agencies have the complex task of investigating threats to the national security within strict legal and policy frameworks. Reconstructing the context of investigative decisions for post-analysis and compliance checks is prone to error and labour-intensive. To address this, we propose to capture decision-making processes and their rationale directly using an OWL-based ontology. This approach overcomes the limitations of traditional data management and existing decision ontologies in handling the intricate data dependencies within threat intelligence (TI) decision-making. The result is the Threat Intelligence Decision Ontology (TIDO), which structures analysts' decision-making while incrementally capturing a decision trace for post-analysis as investigations unfold. The ontology was developed under the complex constraints of safeguarding threat intelligence practices and case information, and validated through competency questions from intelligence experts from the Dutch Defence Intelligence and Security Service (DISS). TIDO offers a novel solution for capturing and understanding decision processes within the sensitive domain of threat intelligence, and evidence-based decision-making in general.
Link Prediction (LP) approaches based on Language Models (LMs) operate over the labels and descriptions of entities and relations in a Knowledge Graph (KG). Recent approaches have shown that incorporating a local graph neighborhood can improve the LP capabilities of LMs. These approaches usually sample a context from the neighborhood around a query triple randomly, thereby incorporating noise that might hinder the model in making correct predictions. In this work, we derive an approximately optimal context for a given query under the assumption that we know the correct answer. This allows us to investigate the characteristics of such contexts, thereby providing an approximate upper bound on the achievable performance when using optimal contexts.
Business process models are widely used in enterprises to orchestrate workflows involving events, tasks, and agents. BPMN provides a standardized syntax to capture structural semantics, yet domain-level semantics often remain implicit, typically expressed through descriptive annotations. Thus, integrating BPMN processes with external data is challenging, limiting semantic interoperability. Building on the principles of Linked Data, several approaches have been proposed to embed BPMN processes within ontologies and knowledge graphs to overcome this limitation. However, current approaches are restricted to abstract process models, with limited or no support for dynamic runtime executions. Therefore, we introduce semantic BPMN (sBPMN), a framework for representing BPMN process models in a modular ontology directly derived from the official BPMN 2.0 XML schema, and their execution traces in a dynamic knowledge graph. Both static and dynamic fragments can then be easily linked to external sources. We demonstrate a system architecture compatible with state-of-the-art BPMN orchestration engines, showing that sBPMN enables semantic linking of process elements and external data as well as validation of process instances against formal constraints. Finally, we showcase the applicability of sBPMN in an industrial use case, where human-AI processes are modeled via BPMN and semantically enriched with domain data to facilitate standardization and semantic downstream analysis.
Contemporary Arabic narrative generation faces a critical cultural authenticity crisis, where neural models systematically violate classical adab principles despite sophisticated linguistic capabilities. We introduce CST-Onto, a hybrid neuro-symbolic framework addressing this challenge through the first domain-specific OWL/SWRL ontology for Arabic cultural narratives. Our four-tier ontology encodes 87 core cultural archetypes from expert annotations of Kalila wa Dimna and One Thousand and One Nights [13], expanded to 847 specialized concepts through systematic derivation protocols. The framework formalizes 234 SWRL rules modeling virtue dependencies (e.g., hakim -> hilm,'adl, fasaha) that directly enhance symbolic-neural integration via constraint-aware neural decoding. Our constraint-injection algorithm maintains sub-28 ms/token latency on NVIDIA V100 GPUs across 1,000 narrative segments, enabling real-time cultural schema enforcement. Evaluation using our novel CRAFT benchmark demonstrates substantial improvements: across 120 native Arabic speakers, we report role coherence (kappa = 0.89), moral-causality consistency (kappa = 0.83), and dialectal authenticity (kappa = 0.92) using validated classifier methodologies. CST-Onto outputs achieved 78% human preference in pairwise comparisons (p < 0.001) against baseline models, representing a 21% improvement in cultural fidelity.
We introduce AnNER, an open-source, lightweight, web-based tool for annotating scientific and other domain-specific entities-including named entities-in text corpora. AnNER offers a full expert review mode that allows reviewers to accept, reject, or revise annotations created by humans or machines. The full provenance of all annotation and review actions is captured in a JSON-based semantic model to ensure utmost transparency during initial annotation and review. AnNER also supports exporting annotations and their histories as RDF documents compatible with the Ontology for Named Entity Representation (OnNER), enabling easy knowledge graph construction to facilitate querying, analysis, and downstream machine learning tasks involving the annotated entities.
The automatic generation of BPMN models from textual descriptions remains a major challenge, particularly when dealing with long and complex texts. Traditional natural language processing (NLP) approaches struggle with linguistic disambiguation, often resulting in incorrect BPMN models characterized by overloaded or missing relevant process elements. On the other hand, large language models (LLMs) such as GPT-4, while effective at capturing semantic nuances, tend to omit key elements required for complete modeling. The objective of this study is to improve the quality of automatically generated BPMN models from textual descriptions by overcoming the respective limitations of traditional approaches and GPT-4. To this end, we propose a prompt-guided fusion method that combines a literature-based NLP approach with the reasoning capabilities of GPT-4. This fusion helps reduce the overload of activities and exclusive gateways, particularly for the NLP approach, while partially correcting errors related to the generation of parallel gateways for both approaches.
Knowledge Graphs (KGs) are widely used to represent structured knowledge. However, their incompleteness under the Open World Assumption (OWA) limits their effectiveness for reasoning and inference. Neural link prediction models can recover missing links. Yet, these models often overlook the distinction between semantically valid and invalid inferences and lack mechanisms to validate predictions against domain-specific constraints. In sensitive domains such as healthcare, predicting a plausible but contraindicated relation can have harmful consequences. This work addresses semantically grounded KG completion by extending the Partial Completeness Assumption (PCA) with two metrics-PCA(valid) and PCA(invalid). These metrics distinguish constraint-compliant from constraint-violating predictions using SHACL validation. They guide the selection of symbolic rules and the generation of labeled training data for neural models. As a result, link prediction systems can assess both plausibility and semantic validity. Experiments on 315 testbeds demonstrate that incorporating constraint-aware symbolic knowledge enhances MRR and Hits@K across multiple KG embedding models, including TransE and TransH. Thus, this approach supports interpretable and trustworthy KG completion.
This work presents experiences and lessons learned in the development of ontologies using a combination of OWL for domain modeling and SHACL for the validation of RDF knowledge graphs that follow the OWL model. In our work, through a real-world use case in the domain of railway transport, we propose a joint OWL+SHACL development approach, including the use of SKOS for the representation of reference data, that is coupled to the definition of the ontology. In this approach, classes and properties are represented in both languages, and constraints may be represented in one or both languages according to the need for ontology reasoning and knowledge graph correctness requirements. During our development process we encountered in the first place that there are differences between SHACL engines with regard to support of language features such as reusable components, which imposes constraints on how SHACL shapes can be developed right now, beyond what is proposed in the W3C Recommendation. In the second place, several tools have been developed to automate the process of generating SHACL from OWL; however, although we may profit from these tools from the development point of view, a significant amount of manual effort is still necessary. Our experience has led us to propose a set of best practices that include manual and automatic approaches to the definition of constraints, the management of changes in both semantic artifacts, ontology and shapes, and the definition of "parameterized" shapes.
We propose CONDOR, a weighted knowledge distillation approach that distills the information from a common-sense KG to a domain-specific KG by intelligently balancing domain relevance and generalization. We use a teacher-student distillation framework where a teacher model trained on a large common-sense knowledge graph guides a student model trained on a smaller domain-specific graph, using a combination of distillation loss and domain-specific triple completion loss to improve downstream task performance. We benchmark CONDOR on offensive speech and mental health domains across classification and generation tasks. Using COMET, pre-trained on common-sense commonsense KGs like ATOMIC and ConceptNet-as the teacher, and domain-specific KGs such as StereoKG (offensive speech) and HealKG (mental health) as student inputs, we evaluate on LatentHatred, SBIC, EDOS, and HOPE datasets. Results show that enriching domain-specific knowledge with distilled commonsense significantly improves downstream performance.
The construction of a knowledge graph (KG) can be performed manually. Nevertheless, ensuring minimal coverage of a KG often requires the automatic data extraction from multiple sources. However, sources and extraction algorithms often vary in quality, may provide conflicting data with different levels of specificity or even contradict each other for the same entity. To reconcile these conflicting data and integrate them consistently within the KG, numerous fusion models can be adopted that simultaneously evaluate both the quality of the sources and the data provided. However, most of these models are usually evaluated on datasets that do not specifically represent differences in specificity, the heterogeneity of data types, or the presence of long-tail entities. These three challenges are frequently encountered in KG construction, making the data fusion process more complex. In this paper, we propose to overcome these limitations by introducing WikiConflict, a dataset built from the Wikidata revision history and designed for KG construction.
Soil health is fundamental to environmental sustainability and food security, yet relevant knowledge remains fragmented across diverse sources, hindering its effective application. Knowledge graphs (KGs) offer a robust solution by integrating disparate information into a structured, semantically rich format. Addressing this need, this paper presents an ontology-compliant soil health KG derived from domain literature, and the semi-automated, human-in-the-loop pipeline developed to construct it. Our pipeline leverages large language models to accelerate knowledge extraction, while incorporating expert oversight to ensure ontological compliance and accuracy. The resulting KG integrates unstructured knowledge into 11,715 RDF triples representing 2,018 entities, including 1,786 soil-related concepts. The KG’s fidelity was confirmed by soil scientists through a validation process involving competency questions. The KG, supporting ontology, and the source code of the pipeline are made publicly available.
Attention Deficit Hyperactivity Disorder is a neurodevelopmental disorder whose manifestations vary significantly depending on the context. This situational variability poses major challenges for assessing and understanding symptoms, particularly outside clinical environments. In this work, we propose a framework that integrates a contextual vision to enrich medical information. The framework is composed of three main components: a modular ontology that formalizes both medical and contextual dimensions of ADHD; a multi-agent system powered by large language models for automatically extracting and populating knowledge from heterogeneous data sources; and a clinical rule-based reasoning mechanism capable of inferring high-level interpretations from instantiated data. Experimental results demonstrate the framework's ability to generate accurate, context-sensitive interpretations of symptom manifestations. This approach lays the groundwork for more personalized, explainable, and context-aware patient monitoring, with promising applications in intelligent healthcare systems.
Integrated knowledge graphs form the foundation of numerous data-driven applications, including search engines, conversational agents, and e-commerce solutions. Declarative link discovery frameworks utilize link specifications to define the conditions necessary for establishing a link between knowledge graphs' resources. Despite domain expertise, defining such link specifications remains challenging due to their intricate syntax, threshold tuning, and the need to precisely express complex linking logic. To address this challenge, we propose NL2LS, a novel language-driven approach that leverages large language models to automatically translate natural language (NL) into link specifications (LSs), enabling domain experts and practitioners to express correct and complex linking rules more effectively. NL2LS employs three distinct training paradigms to handle the complexity of link specifications: zero-shot learning, one-shot learning and supervised fine-tuning. We evaluated NL2LS using different large language model architectures in comparison with a rule-based baseline model on different multi-lingual datasets. Our evaluation using BLEU, METEOR, ChrF++, and TER metrics demonstrates that NL2LS effectively translates natural language into link specifications, lowering the technical barrier and assisting users in specifying link rules more intuitively.
Arabic text adaptation confronts severe computational barriers from morphological complexity and dialectal fragmentation. Current graph-based methods employ static representations that fail to capture evolving morphological dependencies, causing substantial semantic degradation (28.0% cross-dialectal error rate) across dialectal boundaries. We introduce MM-GTA (Morphological Modeling via Graph-based Text Adaptation), the first dynamic-graph framework for Arabic style transfer that jointly optimizes morphological dynamism, dialect-conditioned attention, and cultural-adaptive resource fusion (dialect-aware embedding compression). MM-GTA introduces dynamic graph reconstruction and dialect-conditioned attention, reducing cross-dialect errors by 55.8% versus AraT5 (from 28.0% to 12.3%). Evaluation on ArabicStyleBank (10,000 parallel texts across five dialects) demonstrates substantial gains: BLEU 41.2 +/- 0.8 (+6.5 over AraBERT's 34.7 +/- 0.9), ROUGE-L 71.8 +/- 1.0 (+6.1 points), BERTScore 78.2 +/- 1.0 (+4.3 points). The architecture achieves 51% memory reduction with 89ms average inference latency per sentence (vs. AraT5's 142ms), enabling efficient deployment within 8.0GB memory constraints. Human evaluation by 25 native speakers (5 per dialect, avg. 10 years NLP experience) confirms superior preservation across style authenticity (4.2/5), semantic fidelity (4.3/5), and grammatical accuracy (4.1/5). Our model achieves +17.2% improvement in cultural alignment compared to AraGPT2 across three major dialect families.
Automated Scene Graph Generation (SGG) is challenged by the "Global-Local Dilemma"-the tension between capturing holistic scene context and granular object details. This paper presents a preliminary investigation into a new generative SGG paradigm designed to address this challenge. We introduce a novel, two-track generative framework (GLIIDE) that creates a rich textual intermediary for a Large Language Model (LLM) to extract knowledge. Our hybrid pipeline fuses a YOLO-based track for extracting factual attributes with a DETR-guided track for generating contextual descriptions of object interactions. Our preliminary results on 1,000 images from the Visual Genome dataset demonstrate the potential of this fusion, which achieves a higher Relation Density than relative baselines. This work provides a preliminary proof-of-concept for a generative SGG paradigm that integrates local and global visual information.
