To reduce inaccuracies or a lack of context, AI applications may heavily benefit from structured knowledge modelled relying on reference ontologies. We present rapid-triples, a customizable and dynamic interface that facilitates human-in-the-loop collection of structured knowledge in RDF format. The system allows domain experts to manually input knowledge or semi-automatically enhance an extraction by an automated system from existing data sources. By utilizing a common schema mapped to a target ontology, rapid-triples ensures that the knowledge collected is semantically interoperable and machine-readable. This tool supports various use cases, including expert-guided knowledge creation, validating and refining outputs from automated extractors, and generating high-quality training data for AI systems. Finally, we discuss the adoption of rapid-triples to support a use case in the industrial domain through the collection and exploitation of procedures as knowledge graphs.
The adoption of Large Language Models (LLMs) is reshaping the landscape of AI applications across
diferent domains, yet their black-box nature often limits their ability to access and utilize factual
knowledge, highlighting the need for a complementary integration with structured knowledge systems
such as Knowledge Graphs (KGs) [
The remainder of this paper is organised as follows. Section 2 discusses the human-AI workflows envisioned and enabled by the proposed solution. Section 3 describes rapid-triples and its functionalities, while Section 4 exemplifies its usage in the industrial domain to collect procedural knowledge. Section 5 presents related work on interfaces to facilitate the collection of structured knowledge in
RDF format and compares them with rapid-triples. Finally, Section 6 draws the conclusions and suggests future work.
Constructing high-quality Knowledge Graphs (KGs) aligned to a reference ontology remains a complex and multifaceted challenge. A critical bottleneck lies in the collection and curation of knowledge, which often requires human expertise, contextual understanding, and iterative refinement. In many real-world scenarios, the human actors involved in the process are not experts in ontologies, RDF, or semantic technologies. Therefore, it is essential to support them through intuitive workflows that abstract away technical complexities while still enabling the creation of semantically rich and ontology-compliant knowledge.
To address this problem, we envision a hybrid intelligence approach that combines automated techniques with human-in-the-loop workflows to improve accuracy, completeness, and semantic alignment of the collected knowledge. We identify three representative workflows to be supported: W1 Tacit Knowledge Collection: In this case, the relevant knowledge resides in the minds of domain experts and is not yet documented. The user must be guided in articulating and formalising this tacit knowledge in a structured form that conforms to the target ontology.
W2 Knowledge Completion from Unstructured Sources: Here, initial knowledge is automatically extracted from unstructured content (e.g., documents, reports) using AI-based tools. The extracted information serves as a draft that the user must validate, correct, and enrich. This human-inthe-loop process ensures that the resulting knowledge is accurate, complete and semantically consistent with the reference ontology, overcoming possible mistakes of the automatic extraction. W3 Enhance Automatic Knowledge Extraction Systems: The structured knowledge, validated and reviewed by users, is used as training or contextual data to enhance the accuracy of automatic knowledge extraction solutions.
To address the challenges of efectively modelling structured knowledge while minimising domain user exposure to RDF, we propose a form-based solution supporting the three knowledge collection workflows identified in Section 2.
Figure 1 shows the main components envisioned. A Data Catalogue represents the point of access for users to structured knowledge represented using a target ontology in a knowledge graph. The rapid-triples tool is designed to support: (i) the manual insertion of tacit knowledge via a formbased interface that guides the user and generates a proper RDF output to be added to the knowledge graph (W1), (ii) the possibility for domain experts to use the same form to review and update knowledge extracted by Automatic Extraction components from existing documents (W2), (iii) the provision of expert-curated data collected via the form to the Automatic Extraction components to be used as additional training data to improve existing models or as contextual data to implement on-the-fly enhancements to the automatically extracted output. The last two workflows should be mediated by a Human-in-the-loop Validation component that manages assignment of tasks to users and handles interactions among components. An Intermediate Format representation (i.e., JSON document) is leveraged to facilitate interactions with components that do not support RDF natively.
The rapid-triples tool is based on the process in Figure 2 and can be declaratively customised,
providing: (i) a compliant JSON Schema [
The implementation of the rapid-triples tool is composed of two main components: a Web
interface in Vue.js2 for creating, editing, and visualising procedures and a corresponding backend component
written in Python for storage, processing and interaction with other software components. The frontend
component enables an interface for manual insertion by leveraging the vuetify-json-schema-form3
library. The web interface can be used as a standalone component and is made available open source on
GitHub. The frontend allows defining a template 4 to generate the target RDF directly client-side, while
the backend can execute fully declarative RML mapping rules [
The tool accepts and renders all valid JSON Schemas, but the schema can be enriched with specific directives5 to personalise the interface, i.e., how the form is rendered to the user. As an example, the 2https://vuejs.org/ 3https://github.com/koumoul-dev/vuetify-jsonschema-form 4https://mozilla.github.io/nunjucks/ 5https://koumoul-dev.github.io/vuetify-jsonschema-form/latest/ interface can be customised to use specific terminology that is more familiar to the user, but, at the same time, can be mapped properly to the target ontology. Similarly, a field of the form can be rendered using custom visual components (e.g., a dropdown with image selections) or dynamically retrieved values (e.g., from APIs) through special directives. While this is not mandatory, it is preferable to define a JSON Schema aligned to the semantics of the target ontology. If possible for the considered use case, a proper JSON-LD context can be used to convert a JSON compliant with the JSON Schema to the target ontology.
This section discusses the adoption of the rapid-triples tool within the PERKS project6 for the
collection of procedural knowledge relying on the Procedural Knowledge Ontology (PKO) [
Within the project, the rapid-triples tool has been customised to address the specificities of each use case (e.g., considering custom ontology modules or language) and adapted to enhance the user experience (e.g., developing custom visual components). Moreover, it has been integrated with solutions for AI-based extraction of procedures from documents. The Beko use case in PERKS targets Lock-Out Tag-Out (LOTO) safety procedures, which define a precise list of tasks to be mandatorily executed to ensure that dangerous equipment is properly shut of and not able to be started up again prior to the completion of maintenance or repair work. Figure 3 shows the customisation of the tool for this use case.
The form has been initialised by automatically processing the PDF document shown in the background. The user is now leveraging the tool to further edit and enrich the procedure. While the user sees terms like energy point and custom visualisations (e.g., images for the selection of locks and personal protective equipment), the underlying JSON Schema and the declarative mapping rules guarantee an interoperable output using the classes and properties from PKO (e.g., pko:Procedure) and from the custom ontology module defined for Beko (e.g., pko-beko:MachineEnergyPoint).
Within the PERKS project, we run a preliminary qualitative evaluation of the rapid-triples tool with Beko users, testing both the manual knowledge collection and the completion of automatically extracted procedures. The users were able to generate a valid RDF representation of LOTO procedures using PKO for diferent machines in the factory, and we then compared the existing LOTO procedures with the ones collected via the rapid-triples tool. The domain experts appreciated the manual collection through the adaptive form, as it efectively guided them in documenting additional tacit knowledge while providing a better user experience with respect to paper or tabular-based approaches. Moreover, they positively evaluated the implemented workflow for the human-in-the-loop collection of procedural knowledge from existing documents. In particular, they liked the seamless experience compared to the manual approach. Finally, safety managers highlighted the higher quality of the generated procedures, and LOTO operators valued the possibility of having access to more details during the procedure execution.
The challenge of enabling high-quality RDF authoring by users less familiar with Semantic Web
standards has spurred the development of numerous tools that balance customizability and reusability
of the solutions. One common strategy is the use of schema-aware assistance to dynamically customise
the interfaces based, for example, on the target ontology or the definition of SHACL shapes. ActiveRaUL
takes a web form-based approach, automatically generating RDF-editing interfaces from ontologies
to separate schema modelling from data creation [
The proposed rapid-triples tool also employs a form-based approach but leverages a JSON
Schema definition to automatically instantiate the interface and validate the content. Unlike the other
tools mentioned, we chose to decouple the serialization of the form content from the target RDF
representation. Indeed, the JSON intermediate form facilitates the integration of rapid-triples with
other AI-based solutions and human-in-the-loop interfaces not capable of directly processing RDF.
Declarative mapping rules for knowledge graph construction based on RML [
The FORMULIS system [
Other approaches for the collection of structured knowledge in RDF are based on the annotation of existing documents. OntoPawls [17] is an interface to annotate specific portions of a PDF document according to the classes and properties of a given ontology. However, expecting users to annotate documents directly is often overly demanding, so we identified the need to design a tool that guides users in the process and integrates with automatic methods for knowledge extraction from existing data sources.
We presented the rapid-triples tool, a flexible form-based interface for domain users to perform knowledge graph construction and editing. The form can be configured via a JSON Schema and easily customised to address specific requirements. The intermediate JSON representation handled by the tool facilitates the integration of AI-based solutions for human-in-the-loop knowledge collection and enables a coherent validation and lifting process to the RDF representation according to the target ontology. The frontend component of the rapid-triples tool is available open source and can be used as a standalone solution or integrated with AI-based solutions to implement complex workflows through a dedicated backend component. The tool has been preliminarily validated in the PERKS project, considering the collection of procedural knowledge in industrial settings, and a new development and evaluation phase is planned in the next months.
The future steps will focus on refining the designed interface and further smoothing the user experience in order to reduce the perceived overload [17]. Additionally, we aim to further explore and investigate diferent human-in-the-loop interaction patterns (e.g., extending those introduced in [ 18, 19]), thus exploiting the combination of the tool with various AI-based solutions. Finally, considering the lifting mapping rules, we plan to investigate solutions combining declarative mapping rules and automatic methods to handle parametric procedures that may benefit from the natural language generation capabilities of language models in the KG construction process [20].
This work is partially supported by the PERKS project, co-funded by the European Commission (Grant id 101120323).
During the preparation of this work, the author(s) used Grammarly in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [17] A. Rula, G. Re Calegari, A. Azzini, I. Baroni, I. Celino, Annotation and extraction of industrial procedural knowledge from textual documents, in: Proceedings of the 12th Knowledge Capture Conference 2023, 2023, pp. 1–8. [18] H. F. Witschel, C. Pande, A. Martin, E. Laurenzi, K. Hinkelmann, Visualization of Patterns for Hybrid Learning and Reasoning with Human Involvement, Springer International Publishing, Cham, 2021, pp. 193–204. URL: https://doi.org/10.1007/978-3-030-48332-6_13. doi:10.1007/ 978-3-030-48332-6_13. [19] S. Tsaneva, D. Dessì, F. Osborne, M. Sabou, Knowledge graph validation by integrating LLMs and human-in-the-loop, Information Processing & Management 62 (2025) 104145. URL: https://www. sciencedirect.com/science/article/pii/S030645732500086X. doi:https://doi.org/10.1016/j. ipm.2025.104145. [20] X. Liang, Z. Wang, M. Li, Z. Yan, A survey of llm-augmented knowledge graph construction and application in complex product design, Procedia CIRP 128 (2024) 870–875. URL: https://www. sciencedirect.com/science/article/pii/S2212827124007911. doi:https://doi.org/10.1016/j. procir.2024.07.069, 34th CIRP Design Conference.