Dataspaces are a key pillar of the European Union's digital strategy, designed to facilitate secure, interoperable, and sovereign data sharing across various sectors while ensuring regulatory compliance. These ecosystems enable seamless data exchange among businesses, governments, and individuals, fostering advancements in artificial intelligence, big data analytics, and digital services. However, despite their potential, it remains unclear whether they provide the necessary tools to fully support eficient and user-friendly exploitation. This paper presents a preliminary vision aimed at identifying usability challenges within dataspaces and exploring how human users can better interact with them. In particular, we investigate whether Large Language Models (LLMs) can bridge the gap between raw, machine-oriented datasets and intuitive, human-centered data interaction.
The concept of dataspaces is a cornerstone of the European Union’s digital strategy, designed to enable secure, trusted, and sovereign data sharing across various sectors while ensuring compliance with regulations. By facilitating seamless data exchange among businesses, governments, and individuals, dataspaces drive innovation in fields such as artificial intelligence, big data analytics, and digital services. Additionally, they promote interoperability and sector-specific advancements in areas like healthcare, energy, and agriculture, enhancing eficiency, competitiveness, and sustainability. Dataspaces aim to establish a fair and open data economy, balancing economic growth with ethical and legal safeguards.
The Internet is often referred to as the ”network of networks” because it serves as a global infrastructure that connects multiple computer networks worldwide. Building on this foundation, the World Wide Web (WWW) emerged as a global information system designed to provide access to documents, or web pages. To interact with this information, users rely on web browsers, which interpret HTML code and present the data in a human-readable format.
In the context of dataspaces, the ultimate goal is to achieve seamless interoperability between diferent spaces, creating a ”single market for data” or a ”space of spaces,” a concept akin to the Internet’s ”network of networks.” Within a dataspace, data is typically organized into catalogs, accessible via web browsers, allowing users to explore available datasets, their formats, metadata, and other relevant details. However, this data is often presented as large, complex collections of numbers, labels, and other elements in formats such as CSV and JSON. While these formats are useful for machine processing, they may not always be intuitive, easily interpretable, or manageable for human users.
This raises several important questions: Are machines the primary users of dataspaces? How can humans efectively interact with and extract value from these vast datasets? What is the equivalent of an Internet web browser for seamless interaction within dataspaces? This paper explores these questions and others related to the practical exploitation of dataspaces, examining whether recent advancements in Large Language Models (LLMs) can improve usability, accessibility, and overall human-data interaction within these ecosystems.
CEUR
ceur-ws.org
In this section, we analyze several dataspace projects that are currently being developed with support from EU and national funding. By examining these initiatives, we aim to identify common patterns and challenges related to usability, data accessibility, and practical exploitation. Understanding these key aspects will provide valuable insights into how dataspaces are being implemented, the obstacles they face, and the potential strategies to enhance their efectiveness and adoption across diferent sectors.
Next, we present six ongoing projects in which Libelium is actively involved, each focused on implementing dataspaces to enhance data sharing, interoperability, and digital transformation across various sectors. The selected projects are DEPLOYTOUR, BeatTheHeat, SENSE, Geo4Water, LDT-DS, and SC-DIXAE.
DEPLOYTOUR [
The BeatTheHeat project [
The project fosters collaboration with the EU Citiverse industry, including SMEs, to integrate Virtual
Reality, and metaverse technologies, enabling citizens to navigate and engage with digital urban spaces.
However, SENSE goes beyond technological advancements—it seeks to enhance the social, architectural,
environmental, and cultural dimensions of urban living. By creating immersive and interconnected
virtual spaces, the project strengthens citizens’ sense of belonging, ultimately enriching urban life
across Europe.
2.1.4. Geo4Water
The Geo4Water project [
night-time temperatures. This data will guide decisions on urban planning, such as selecting pavement types, building materials, and implementing infrastructure improvements like green roofs and other heat mitigation strategies.
Additionally, the Geo4Water project will enable city managers to assess urban resilience to extreme rainfall events, helping them make informed decisions on nature-based solutions like rain gardens and Sustainable Urban Drainage Systems (SUDs). These specific use cases provide user interfaces tailored for city managers, ofering functionalities for: (1) configuration of settings, (2) customized visualization of results, and (3) generation of detailed reports, ensuring transparency and easy interaction with the underlying dataspaces.
On the other hand, SENSE will ofer citizens immersive 3D experiences, enabling them to explore their cities and interact with data such as public transport systems, points of interest, and more. The interfaces provided by these services will be designed for the general public, ensuring that users of all backgrounds can engage with the data and derive meaningful insights from their surroundings.
Therefore, usability and data accessibility are integral aspects considered in the development of services, and are typically addressed using the same principles applied in any software development project. These principles are rooted in human-centered design, which aims to ensure that the final product is intuitive, efective, and user-friendly. To achieve these goals, developers usually adopt well-established guidelines such as the Web Content Accessibility Guidelines (WCAG) and implement best practices.
In general, direct user interaction with dataspaces primarily involves data catalogs, which allow users to discover and access datasets from various sources. However, this approach presents several challenges. First, data catalogs often lack uniformity, making it dificult to assess the quality, relevance, and usability of datasets quickly. Second, interoperability issues arise when data is stored in diferent formats or systems, complicating seamless integration across platforms. Additionally, the sheer volume of data can overwhelm users, especially if there are no efective tools to filter or prioritize information based on specific needs. Finally, ensuring the accessibility and usability of these catalogs for a wide range of users, including non-technical stakeholders, remains a significant hurdle, requiring interfaces that are both intuitive and flexible.
This issue stems from Open Data portals, which promote the free distribution of data for public
consumption and republishing [
MICE industry Weather forecasts, Earth obser- Decision-support services for city vation data, IoT devices, build- managers, providing high-value ing database, urban tree cata- datasets like shadow maps, thermal logs, trafic mobility data comfort maps, UHI maps, pollution maps, and Health Impact metrics City virtualization for managers and citizens, focusing on culture, environmental awareness, infrastructure resilience, city planning, and smart mobility Weather forecasts, Earth obser- Decision-support services for city vation data, IoT devices, build- managers, including infrastructure ing database, city infrastructure, monitoring, water pollution maps, geophysical data damage assessment, urban re
silience, and event history Data on water and energy con- Environmental management to aid sumption, transportation, waste decision-making for city oficials in management, parking, IoT de- urban planning and mobility vices, and more Inventory of IT equipment and Services for city managers, includnetworks, access and authenti- ing risk assessments, vulnerability cation logs, system vulnerabili- evaluations, and disaster recovery ties, security incidents, IoT sen- and resilience planning sor network data, network traffic monitoring, security policies and protocols, software update and patch status and more Open Data, making their datasets available online to enhance transparency and empower the public to monitor and hold government actions accountable. However, despite the abundance of available data, few individuals have the technical expertise to fully leverage it. For those who do, the process can still be daunting. Finding relevant datasets, understanding their formats, and validating their usefulness through basic tests or visualizations requires significant time and efort.
As a result, based on all the above, the answer to the question posed in the introduction—”Are machines the primary users of dataspaces?”—is no. Dataspaces have the potential to benefit all types of users, regardless of their technical expertise. Enhancing human-centered interaction not only makes data more accessible to the general public but also empowers technical users to develop data-driven services more eficiently and rapidly, ultimately strengthening the capabilities of dataspaces. Figure 2 presents a collaboration diagram that broadly illustrates the interactions and relationships between users and dataspaces.
Large Language Models (LLMs) are advanced AI systems trained on extensive text data to understand, generate, and process human-like language. Models such as OpenAI’s GPT series, Google’s Gemini, and Meta’s LLaMA leverage deep learning techniques, particularly transformer architectures, to deliver context-aware and coherent responses. Through conversational interactions, LLMs can assist users with a wide range of tasks, improving accessibility, eficiency, and automation. These capabilities make LLMs invaluable for applications in digital ecosystems, including chatbots, virtual assistants, and knowledge management systems, ultimately improving user experiences across sectors.
Figure 3 presents a conceptual illustration of how the catalog interface could be structured with new capabilities powered by LLMs. This is merely an example, so the specific details displayed, such as the query, the datasets or the resulting graph, are not intended to be relevant.
The interface would be primarily organized into three panels, from top to bottom:
• Query Input Panel: Users can enter natural language queries to search for data, apply filters,
and request basic processing or visualizations. For example, ”Retrieve data on the number of
visits to archaeology museums in Andalusian cities throughout 2024 and visualize the average
daily visits over the last 25 days of June”. Additionally, predefined query templates for common
tasks such as “Search & Filter” or “Data Visualization” can be provided, following the template
structures proposed by [
metadata, and access the data if needed, just as they would in a standard catalog. • Analysis Results Panel: This panel presents the generated visualizations and basic data processing outputs, such as graphs, statistical summaries, or tables with a limited set of transformed data samples. It provides users with an intuitive way to interpret the results derived from the selected datasets.
Regarding the functionalities that the conversational interface could ofer, they may align with the
features suggested by Distefano et al. in [
Figure 4 presents the proposed approach for integrating LLMs into the dataspace architecture. When
a user accesses the catalog, an LLM agent is instantiated to manage the conversational interaction.
As the user engages with the agent, the LLM may need to interact with the catalog API to retrieve
metadata about datasets and perform searches. In alignment with the IDS-RAM [
When a query requires accessing data for processing or visualization, this operation must be executed
via the user connector. This ensures that, firstly, the user has the necessary authorization to access the
data, and secondly, that the traceability of transactions is preserved. Furthermore, if it is necessary
to download data for processing or temporarily store intermediate data or results, this will be carried
out in a sandbox environment within the LLM agent. This approach optimizes integration with the
LLM while also enabling the implementation of secure processing environments in compliance with
the European Data Governance Act [
Recent research has highlighted the potential of LLMs for semantic labeling tasks. For instance, studies
like [
The IDS-RAM introduces the Vocabulary Hub as a component for managing domain-specific vocabularies based on RDF schemas. This functionality enables the assignment of meaningful attributes to datasets, ensuring that the data can be properly categorized and interpreted in relation to its context. By using well-defined vocabularies, the component improves the semantics of the data, facilitating more accurate and relevant searches. However, when LLMs are employed, their interpretation of user queries depends on the general-language semantics embedded during their training. A few key points arise in this context: 1. LLMs could directly process natural language descriptions of datasets, reducing the need to create complex ontologies or annotate datasets manually. This approach would not only save eforts but also enable non-technical users to contribute to the process of describing semantics. Moreover, providing LLMs with contextual information can enhance their understanding of domain-specific descriptions, improving their ability to interpret and interact with the data. 2. Despite advancements in Ethics and Explainable AI, LLMs remain black-box models that lack transparency. For instance, the assumptions they make to interpret user queries may not be traceable. This lack of transparency can introduce biases and inaccuracies, which may afect data searches. The potential risks of this are similar to those posed by internet search engines, where results are influenced by opaque algorithms and may favor certain perspectives over others. In dataspaces, this could lead to some datasets being more visible than others, depending on the LLM interpretation. 3. The Vocabulary Hub could play a relevant role in addressing the previous issue. We could extract semantic terms from user queries automatically according to a specific ontology. Then, the consistency of these terms could be validated against the datasets resulting from the search performed by the LLM, ensuring alignment between the query and the selected data. This process could help mitigate potential biases and enhance the relevance of search results.
This paper has explored the usability challenges that human users face in interacting with dataspaces, particularly in the context of accessing and leveraging data efectively. Through the analysis of multiple projects, we identified that while dataspaces provide a robust framework for secure and interoperable data sharing, they often fall short in ofering intuitive, human-centered access mechanisms, making it dificult for non-technical users to extract value from available datasets.
To address this gap, we examined the potential of Large Language Models (LLMs) as a bridge between complex, raw data and more accessible, human-friendly interfaces. Our analysis suggests that LLMs can enhance usability in several ways, including facilitating natural language search queries, improving data discovery, enabling contextual visualizations, and assisting with semantic interpretation. By integrating LLMs into dataspace architectures, we can lower the technical barriers for users, making data-driven decision-making more inclusive and efective.
However, challenges remain. Issues such as data quality, bias in AI-generated responses, and the need for secure and transparent interactions with dataspaces must be carefully managed. Additionally, integrating LLMs requires adherence to ethical and regulatory guidelines to ensure fair and unbiased data access. Future research should focus on refining LLM capabilities within dataspaces, optimizing human-data interaction models, and developing approaches to improve accessibility while maintaining data sovereignty and security.
This research was supported by the European Union’s Digital Europe Programme through the following grant agreements: 101167948 (SENSE), 101123342 (BeatTheHeat and Geo4Water cascade funding via DS4SSCC-DEP), 101084071 (DOME), and 101100728 (Citcom.ai).
During the writing of this paper, the authors used GPT-4o in order to: Grammar, translation 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.