Posts | Sven Lieber

ENDORSE 2025

Wed, 29 Oct 2025 09:00:00 +0100

The European Economic Area is one of the best use cases for (semantic) data integration: the many different languages, legislations, rules and systems must be or become interoperable! On October 8 and 9, 2025, I attended the 3rd European Data Conference on Reference Data and Semantics (ENDORSE) in the European Commission’s building Charlemagne in Brussels, Belgium. In this blog post I will reflect on the conference, provide details from some of the presentations, and talk about my own presentation in the Cultural Heritage session.

Context

Several hundred people participated in this year’s ENDORSE conference in person and many watched the sessions via an online livestream. A total of 85 speakers participated in 45 accepted presentations (out of 61 submissions), 8 posters, 2 panels and 3 keynotes.

The Publications Office of the European Union, which organized the conference already highlighted several key messages in a LinkedIn post.

(1) Data, semantics and AI must work hand in hand to build services that are truly citizen- and business-centric.

(2) The power of AI depends on the quality, precision and multilingual nature of the data it learns from, a uniquely European strength.

(3) Semantic interoperability used to be a niche field. Today it’s a strategic enabler of Europe’s digital transformation.

(4) Collaboration, co-creation and trust remain the foundations of an interoperable and inclusive digital ecosystem.

(5) Execution matters! Turning insights and knowledge into concrete actions is how we deliver meaningful impact.

(6) A key takeaway? Reference data is the glue that binds all these areas, a message that resonated strongly throughout our vibrant networking sessions.

I can only agree! This conference was a great networking event for the GovTech sector where Semantic Web standards and technologies are put into practice.

I keep this post short and only focus on my own takeaways, a few details from some of the presentations and my own contribution related to the use of Wikibase in the MetaBelgica project.

My main takeaways

From vocabularies to application profiles: governance matters
Linked Data meets Generative AI meets data spaces: existing structured knowledge can be offered via the MCP protocol to generative AI models, but also well-documented specifications make it more human and machine-friendly
The professional conference setup really made a difference: professional moderators, time keepers and online streaming all made the conference visit a nice experience

One downside of the conference: the speaker’s presentations, one of the main outcomes of the conference, are only available via links on the ENDORSE website and hence are not FAIR. I would have loved to cite all the presentations that I mention with a DOI. Providing all presentations via e.g. a Zenodo community (like done for the CoRDI conference) would have been a great solution to make the outcome of ENDORSE FAIR.

Data stewardship in the Age of AI

In his keynote about the future of data stewardship, Stefaan Verhulst from the Govlab covered many interesting topics around what he called the emergence of open data winter while having an AI summer.

According to Stefaan we witness the emergence of an open data winter: access to data is backsliding, so is the excitement for open gov data and even Creative Commons sees a decline in licensing. A general GenAI anxiety is noticeable, yet there is hope: we are able to ride the fourth wave of Open Data.

In a nutshell, this inspiring keynote introduced the following six shifts in data stewardship, to become more strategic and prevent an open data winter.

supply => demand (become demand-centric, key questions regarding a problem and need for evidence for decision makers)
datasets => Knowledge Graphs for interoperability
structured data => unstructured data (Large Language Models like unstructured data, but it was never part of data management)
consent => social license (to engage with communities, current consent is flawed in multiple ways)
FAIR => FAIR-R (AI-ready, authoritative provenance, model context protocol (connection from AI to external systems))
Open Data => data commons (GenAI anxiety, common-crawl + wikipedia fuels model training to a large extent, following the theories of Elenor Ostrom: govern resources in ways that are not extractive and align with principles and expectations of those that created the data or make it accessible, data commons prices)

The recording of Stefaan’s presentation is available here

Stefaan’s talk looked into problems at the horizon and how we could deal with them. Data governance and openness are not static goals but evolving practices. The next presentations showed how these principles are already being applied in concrete public sector contexts.

Interoperable Public Services and Data Spaces

In her keynote, Kjersti Steien from the Norwegian Digitalisation Agency (Digdir) talked about the implementation of Public Services in Norway.

Similar to other European countries, Norway also have a catalog of public services. Even though Norway is not a member of the EU, they use the Core Public Service Vocabulary Application Profile (CPSV-AP) of the EU.

More specifically, their service catalog is described in the Norwegian extension CPSV-AP-NO and can also be connected to relevant concepts described with SKOS-AP-NO, and datasets according to the DCAT-AP-NO model. A nice detail: the Norwegian CPSV extension also includes services delivered by the private sector.

This example shows how EU application profiles can provide a shared basis even beyond EU borders. At the same time, national approaches still vary quite a lot.

During the conference I have learned, that also Germany has a catalog of public services: the FIM portal (presentation by Felicitats Löffler and Marianne Mauch from the state of Thuringia, the green heart of Germany). However, it is not derived from CPSV-AP and I don’t think it is mapped to it yet. This brings me to an interesting point mentioned by Kjersti:

It would be great if the EU would recommend CPSV-AP as one of the specifications to use with the Single Digital Gateway

These different implementations underline that even with good standards, alignment still depends on manual curation and coordination between actors.

Overall the work on public service descriptions makes me think of two things: first, how important manual data curation is. Secondly, on some work I did a few years ago in the FAST project, on providing personalized workflows for life events to improve the customer journey (presentation)

The recording of Kjersti’s presentation is available here and the recording of Felicitas and Marianne here

After several talks on interoperability across administrations, the next keynote (which was the first of the conference actually) focused on Knowledge systems in general and what it means to make data understandable for both humans and machines.

Provenance and Agreement

In his keynote, Paul Groth from VU Amsterdam (co-author of the W3C PROV standard and the FAIR data principles) covered several topics related to Knowledge Engineering and Large Language Models (LLMs).

One interesting thing that I’ve learned about in Paul’s talk is C2PA, a global standard for content authenticity, that recently was integrated into Android. Basically it adds Content Credentials via cryptographical methods to digital media and thus the necessary provenance to verify that a certain image or video was created and not generated by AI. Something similar I’ve seen at the poster of the Piveau-X catalog, uploaded content needs to be cryptographically verified.

Paul also talked about how important, but also how time consuming, it is to work on standards and find agreement. He gave the example of the W3C PROV specification, where he and others were involved in in 8820 public E-mails, 666 issues, 600 Wiki pages, 6000 mercurial commits and 152 teleconferences.

What I take home from this talk:

Make standards more human readable, so they are more machine (LLM) readable

Adding more documentation to your specifications and data models is never a bad idea :-)

The recording of Paul’s presentation is available here

Paul’s talk highlighted the effort that goes into developing common standards and the importance of documenting them well. The following presentation by Bert Van Nuffelen from Digitaal Vlaanderen gave a concrete example about Knowledge Engineering of the Open Standards for Linked Organizations (OSLO) in Flanders, Belgium.

He highlighted the fact that reuse of knowledge is important, but also that it can lead to questions regarding data governance. For example, certain constraints should not be defined directly on the concepts in the namespace of a vocabulary, but rather in the application profile that reuses terms of this vocabulary.

However, how does the governance of an application profile actually works in practice? After all, the reused terms are governed in vocabularies of different namespaces. The emergence of data spaces aggravates the situation, because several data spaces can make use of a mix of different application profiles and vocabulary terms.

The OSLO presentation did not stop at data governance. It also touched on how AI systems can interact directly with these well-defined semantics.

OSLO allows language models to query its structured descriptions through a Model Context Protocol (MCP) server, which is an interesting step towards connecting Linked Data with generative AI. I first learned about MCP at this conference. While Paul Groth advocated for more machine-readable standards, the MCP approach demonstrates how machines can use them in practice.

The recording of Bert’s presentation is available here

Talking about data spaces, they were another recurring topic at this year’s ENDORSE. In one of the panels it was said that, in theory, you could even run a data space on paper. This highlights how much the concept still depends on governance rather than technology. So far there is not one single standard, but several coexisting initiatives and implementations.

During the discussion, moderated by Pieter Colpaert, one speaker mentioned the browser wars of the 1990s: many competing implementations, not all interoperable yet. The International Data Spaces Association (IDSA) was mentioned as one of the key players, and the European Commission already recognizes more than 70 common European data spaces even if it is not always clear to me what qualifies as one.

What I found particularly interesting was the perspective on procurement. Instead of one large platform where only big companies can compete, data spaces could allow smaller providers to offer interoperable services. The sovereignty of participants and the absence of monopolies were highlighted as essential principles. The main recommendation from the panel: don’t make the definition of a data space too strict, it’s better to have working examples than theoretical perfection.

The recording of the data spaces panel is available here

These discussions about interoperability, governance and shared infrastructures also resonated strongly in the cultural heritage session, where I presented our own approach to data management in the MetaBelgica project.

MetaBelgica: central vs decentralized data management

This year’s ENDORSE also featured a session on cultural heritage. Besides presentations about DE-BIAS from Orfeas Menis Mastromichalakis, the Dutch data space for cultural heritage from Enno Meijers and a case study for historical encyclopedias from Thora Hagen, I presented work of our MetaBelgica project.

In particular, my presentation focused about our choices on data management. On the one hand, due to existing legacy systems and duplicate efforts in data curation, we opted to not have a fully decentralized solution.

On the other hand, we also did not choose a fully centralized solution, to avoid having a single-point of failure system in which partners from several institutions have to maintain all their data.

Instead we have chosen for a Wikibase system that I situate in-between with a best-of-both worlds approach: Loosely coupled FAIR data.

We rely on the mature data management software Wikibase (with Technology Readiness Level 9), while keeping a balance of control and collaboration. For example, we keep local autonomy and possible additional institution-specific data, while providing de-duplicated high quality FAIR reference data in a single trustworthy system with Persistent Identifiers for the public.

The recording of my presentation is available here and my slides via the following DOI: 10.5281/zenodo.17344191.

Final remarks

ENDORSE 2025 once again showed how Europe combines technical precision with collaboration. From open data stewardship and interoperable public services to AI-ready standards, the conference made clear that the future of digital government depends on connecting structured knowledge and the people maintaining it. This post only covered a few talks that I found particularly interesting, there are many more presentations (and recordings) to be discovered on the ENDORSE website.

References

Lieber, S. (2019, June 25). The FAST project: Criterion and Evidence from the Public Service UX perspective. Zenodo. https://doi.org/10.5281/zenodo.8001639
Lieber, S. (2025, October 8). Opening up Belgian Cultural Heritage Reference Data with Wikibase. The European Data Conference on Reference Data And Semantics (ENDORSE), Brussels, Belgium. Zenodo. https://doi.org/10.5281/zenodo.17344191

CoRDI 2025

Sun, 07 Sep 2025 09:00:00 +0100

In the old days, research was done by famous individuals. Nowadays, research is a team effort! Working together requires making agreements about how to collaborate and using common tools. These kinds of agreements (or standards) and tools are especially crucial given the amount of data nowadays. Me and around 700 other people attended the 2nd Conference on Research Data Infrastructure, in Aachen, Germany from August 26 to 28, 2025. In this blog post I will reflect on the various presentations and posters I saw as well as the discussions I had.

Context

In August 2025, the German National Research Data Infrastructure (NFDI) organized the second edition of the Conference on Research Data Infrastructure (CoRDI). Around 700 people attended this international multi-track conference. This year’s edition did not follow a standard peer-review, but was curated based on the conference organizers and a balance of the various disciplines and topics in the field of research data infrastructure.

Even though my contribution was not selected, I anyway attended the conference as this is a great chance to network. Next to the small things, such as live music from the RWTH Aachen big band, the daily buffet with default vegetarian/vegan options, and exciting robot demos at the social event, the conference featured Research Data Management (RDM) in various disciplines.

My main takeaways

Standards and identifiers are the hidden champions: Whether in workflows, Knowledge Graphs, or infrastructures, persistent identifiers and authority data are the backbone that allow collaboration across disciplines.
Infrastructure is both technical and human: Behind cables, servers, and workflows, it’s people, governance, and communities that keep infrastructures alive.
Bridging scales is key: The inspiring keynote reminded me how crucial large-scale infrastructures and global governance are. But in the sessions and posters, I also saw how much progress depends on practical tools, workflows, and community initiatives that researchers can actually use today. The real challenge is linking both scales together.

Even though in the rest of this post I will only focus on my personal main highlights it is still a quite lengthy post. Feel free to immediately jump to the last section on my reflections or other sections that interest you by using the table of contents below.

Table of Contents

Keynotes

This year’s conference featured two keynotes: Rob Finn talked about the complex data ecosystem of life science data and Cathrin Stöver provided insights into the physical networks that make all of our work possible in the first place.

Rob Finn - Life Science data

On the first conference day, Rob Finn from the European Bioinformatics Institute talked about the kind of data that they process and publish. As well as where he sees potential for AI.

According to Rob, and hinting towards work of Carole Goble, there are ad-hoc workflows which can be made FAIR by describing them with Ro-Crate and uploading those descriptions to the WorkflowHub.

It’s not really my domain, but what I take home from Rob’s talk is the evolution of workflows. Because I recognized my work for the BELTRANS project, where we create a corpus of book metadata of contemporary translations between Dutch and French. It’s a completely different area, but in our case we also start with a bunch of Python scripts, glued together with bash scripts.

Cathrin Stöver - Research and Education Infrastructure

In her impressive keynote on the last conference day, Cathrin Stöver the Chief Communication Officer at GÉANT, discussed how GÉANT provides global connectivity for researchers.

She wasn’t referring to a few online services or cloud solutions. She discussed real-world 800 GBit traffic via GÉANT’s own submarine cables, as well as other success stories, such as the well-known eduroam Wi-Fi internet access service.

GÉANT provides research and education network services by interconnecting 38 National Research and Education Networks (NRENs) such as Deutsches Forschungsnetz (DFN) in Germany or BELNET in Belgium and serves around 50 million users across Europe. In the last 15 years, other parts of the world have been connected as well, such as East Africa extensively and West Africa to a lesser extent.

In addition to these interesting facts, the keynote also focused on the conference organizing organization NFDI. It must undergo evaluations and needs to become operational within a few years. Based on GÉANT’s experiences, Cathrin identified three important aspects for success: validation, people, and funding. GÉANT receives top-down and bottom-up validation from both its users and recognition on policy level, including an alignment with the Sustainable Development Goals. Evenly important is the people aspect with a strong community and regular community conferences. But also having a value conversation to keep people with passion. And last but not least, the aspect of secure and sufficient funding, which due to the multiplier effect pays off.

If you ask me, I think NFDI is on a good way! There is a large user base and recognition from the policy level, both from the German government as well as via links to European research. I have also met many passionate people at the conference. And at least from an outsider perspective, I think lots of money is invested in Germany for RDM at the moment.

Persistent identifiers, Authority data & Knowledge Graphs

If there’s one area at CoRDI that resonated most with my own work, it was the trio of persistent identifiers, authority data, and knowledge graphs. They are deeply interconnected: identifiers give us stable references, authority files give meaning to those references, and knowledge graphs weave them together into broader ecosystems.

PIDs: Small links, big impact

Every Knowledge Graph or authority file starts with the simplest building block: a persistent identifier. Several posters and discussions at CoRDI showed just how central PIDs have become.

Unfortunately I missed the PID panel, because it was scheduled in parallel to the humanities and authority files session But thanks to the poster sessions, I could speak with experts directly. For example at the poster of Barbara Fischer from the German National Library (DNB) and Paul Vierkant from DataCite about the PID network Germany.

We don’t have something similar on national level in Belgium yet, (probably also due to the special federalism in Belgium), but I’m glad that my colleagues from the FedOSC project are also looking into the PID topic, at least for the Federal Scientific Institutions! While talking about PIDs, we were also joined by Daniel Mietchen, who argued that Wikibase already comes with PIDs out of the box.

But identifiers on their own are just strings (or ideally Uniform Resource Identifiers (URIs)). Their real value comes when we use them to describe entities such as people, places, or works, which is where authority data comes in.

Authority data in the humanities

Authority files give identity to PIDs, turning abstract identifiers into meaningful references for people, places, and works. The humanities session at CoRDI highlighted how this plays out in practice.

The humanities session featured talks about the Gemeinsame Normdatei (GND), which is the authority file of German-speaking countries, and the Knowledge Graph software Wikibase.

In the presentation, Authority Files and the Text+ Data Space (10.5281/zenodo.16736146), Barbara Fischer highlighted that

“standards are not dropping from the sky”

Furthermore, she argued that we should not lose diversity when using standards. Barbara used English as an example of a language used for understanding each other in science. However, it does not replace one’s mother tongue, with which one can express much more diverse concepts. Barbara further argued that we have to give people the edit button, or at least a button to suggest changes, this will also enhance collaboration.

Open Knowledge Graphs are one way to enable (large-scale) collaboration. While most people are familiar with Wikidata, there are also lesser-known, more domain-specific Knowledge Bases. One such example is the FactGrid Wikibase instance, used by many historians. Because the historians share a single Wikibase instance, they are also forced to work together.

FactGrid’s creator, Olaf Simons gave the presentation Wikibase - the best software to communicate with the upcoming knowledge graphs? (10.5281/zenodo.16892677). In it, he states a very concrete issue: the 100 million item limit of Wikidata’s (and Wikibase’s) Blazegraph triple store! Apparently the Wikidata graph was split as a workaround. Yet, Olaf argues that these kind of workarounds do not solve the underlying issues. He calls for the simple but effective solution:

“The NFDI should develop a joint interest with Wikimedia Germany to break the present 100 million item limit”

I can very much relate to Olaf’s statement. As a computer scientist I am used to developing custom software solutions when I cannot find standard software that addresses the problem. I often see software developers in different (engineering) fields develop software even when standard software is already available. One example is software to generate RDF from other data sources. Standard software is important! In the humanities, for example, where tools with a user interface and a high Technology Readiness Level (TRL) are needed, Wikibase plays an important role. This fundamental software should be maintained and further developed to overcome its current limitations.

Authority files create structure, but once we start connecting them across domains and datasets, we step into the world of Knowledge Graphs and ontologies.

Knowledge Graphs and Ontologies

Knowledge Graphs are where identifiers and authority files come together, enabling connections across datasets and disciplines. At CoRDI, I saw many examples of both the opportunities and the challenges of building such ecosystems.

In the humanities session, Florian Thiery and Lozana Rossenova presented challenges of federated queries using the Wikiverse and OpenStreetMap within the NFDI Knowledge Graph Ecosystem (10.5281/zenodo.16736048). While doing federated SPARQL qeries across Wikibase instances, including FactGrid, they noticed technical and conceptual hurdles, such as namespace collisions, property mappings or allow lists. The latter specifying which other Wikibase instances can be accessed by federated queries.

On the last day of the main CoRDI conference, Lozana Rossenova also presented the results of a survey about how NFDI consortia are using Knowledge Graphs (10.5281/zenodo.16736078). The survey was conducted between March and May 2025 and covered 27 Knowledge Graphs from 11 consortia.

For their presentation, the challenges were grouped into four larger categories. To me, it looks very familiar. Anyone who starts working with Knowledge Graphs will eventually face one of these challenges. We cannot overcome every challenge in a generic fashion. I believe we should curate best practices related to these challenges and create teaching materials from them.

Volker Hofmann presented the Helmholtz Knowledge Graph (10.5281/zenodo.16736336). I liked the used images of the presentation, such as the following about separated islands of data. Apparently only a small percentage of all the entities.

This made me think about the many authority records in MetaBelgica for which we only have a name and no linked birth/death dates or birth/death places. In our source data we have linked publications though, instead of leaving them out, we maybe should add some linked PIDs, ISBNs or publication places to MetaBelgica to not create authority islands…

In Sarah Rebecca Ondraszek’s presentation on the NFDI4Memory Knowledge Graph (10.5281/zenodo.16736125), I found two things interesting. First, they do not yet have a link with PROV-O, because of the more difficult alignment with BFO (yet there could be mappings). Second, they mentioned Sampo, a Linked Data publishing tool from Finland, that I have recently seen at various conferences. We are testing it for the BELTRANS project as well, and I know that the Sampo team is collaborating with the CLARIAH-VL+ project in Flanders/Belgium to improve the software (Sampo fork from the Ghent-Centre for Digital Humanities that will be rebased to the original Sampo).

During the Q&A of Julia Thoennissen’s presentation about FAIR and scalable access to large image data in the range of petabytes (among others in neuro science) (10.5281/zenodo.16736220) I wondered if their workflow, which contains 3D images, could be extended with the International Image Interoperability Framework (IIIF) for data publishing. This would provide a standard software solution instead of a custom one. The authors were not familiar with IIIF, it’s probably less well-known outside of the cultural heritage domain. Afterwards I had a look myself, there is a 3D community working group at IIIF. However, based on the community group’s current description it does not seem that this feature already exists.

At the poster of the NFDIcore 3.0 ontology, I also had a small chat with Jörg Waitelonis about the BFO basis of the ontology (10.5281/zenodo.16736251). We talked about the possibilities of the ontology to also model temporal aspects, such as names that change over time. Which comes with a lot of extra triples (in case this is done for many entities).

In the presentation of Leyla Jael Castro about Bioschemas and Schemas.science at NFDI (10.5281/zenodo.16735850), I have learned about Bioschemas and Schemas.science. The first is a is a community-driven initiative to build upon Schema.org for the life science domain. The latter is a sister website in which they broaden the Bioschemas efforts to a more general research-related extension for Schema.org. In both cases to enhance findability, as Schema.org is mainly used by search engines.

AI

It’s 2025, of course there has to be at least one session about AI! Sandra Geisler introduced the AI session, by mentioning current barriers for FAIR Research Data Management, and for which barriers AI can help.

For me it’s also not only about how AI can help RDM, but also how RDM can help AI. I forgot who said it, but someone at CoRDI remarked that Ontologies and the Semantic Web are the ground truth to make AI creative. I find this a very spot-on statement, because machine learning without annotated data is just pattern-finding. To move beyond that, we need humans and semantics: structured knowledge that gives AI something meaningful to work with.

Other AI-related talks and posters at the conference focused on different aspects to annotate machine learning models. To this regard I have learned about things like FAIR4ML, a vocabulary to describe Machine/Deep Learning models (10.5281/zenodo.16735334), presented by Leyla Jael Castro, or MLentory, a registry for Machine/Deep Learning models making use of FAIR Digital Objects (10.5281/zenodo.16735316).

One presentation also mentioned an AI product, that can generate a conversational podcast out of a research paper. So you hear artificial voices having a chat about the topic. The presenter said that after listening to the podcast, one interesting thing was that the AI podcast contained examples that were not in the paper, but which the author deemed interesting examples as well. That sounds amazing to me, I have the feeling that this can revolutionize scientific communication.

Infrastructure

One part that is utterly important, and leans towards the impressing keynote, is actual physical infrastructure. With all the existing clouds, data spaces and services, I’m often wondering where the stuff is actually hosted. Who pays for it and for how long will the services stay online?

At the very basis we have the NRENs from GÉANT, thus for example BELNET in Belgium. And via procurement one can get access to virtual machines for example via the OCRE programme at data centers of partners.

In the context of the NFDI, base4nfdi provides higher level services such as jupyter4nfdi. At the poster of base4nfdi, Martin Reinhardt explained to me how the jupyter4nfdi works. It is essentially a JupyterHub hosted at Forschungszentrum Jülich and another data center. Research institutions can host their own Jupyter services, creating a decentralized network that researchers can access via the central JupyterHub and Single-Sign-On. This way, research institutions can configure their settings so that only their own researchers can use the Jupyter resources from the institution’s servers, for example. However in theory, researchers from other institutions could also use the provided resources. According to Martin, this is also in the interest of cloud providers, because they usually want at least a certain percentage of server load; otherwise the infrastructure just sits there and costs money.

Eventually, research institutions still provide the actual infrastructure, or at least pay for virtual machines in the cloud. This brings us back to the initial question: Who pays for it, and for how long will the services stay online? This obviously depends on the research institutions, but I think it will likely be project-based funding after all.

While not ideal, there are real-life examples of successful infrastructures that prove such funding can work. For example, historians can use FactGrid to store their data even after their funding period ends. In Flanders, I came across the ODIS platform, which also offers a home for research data within its scope. For the latter I also have heard that researchers already started to mention this platform (or repository) in their mandatory data management plans.

Technical infrastructures are impressive, but they rely on more than cables and servers. Agreements, policies, and shared responsibilities - in short, governance - are what make these systems sustainable.

Governance

When we talk about infrastructure, we inevitably end up talking about governance: the agreements, rules, and decisions that make collaboration possible. Networks don’t just run on cables and servers, they run on policies, contracts, and trust. This theme surfaced repeatedly, from the inspiring keynote to lively panel debates and practical project presentations.

Global governance

They keynote of Cathrin about global connectivity obviously is also related to global governance, which is not an easy topic in today’s geopolitical reality. This was the main subject of the keynote’s Q&A, with questions relating to why certain parts of the world are not connected (GÉANT’s headquarters are in Amsterdam, so it falls under Dutch and European law and is forced to follow sanctions) and why satellite networks are not used instead of terrestrial networks (there is no low-orbit capacity yet owned by Europe, and high orbit does not provide enough bandwidth).

A related perspective came up in the resilience panel: the reminder that researchers are citizens of science, not citizens of a country. Infrastructure cannot be shielded from geopolitics, but the research community aspires to build connections that transcend national borders.

This perspective was underlined by the example of PANGAEA during the resilience panel. It stepped in to safeguard research data that risked being lost due to shifts in U.S. policy. Here governance became a matter of resilience: ensuring that research data remain accessible for the global community, even when political frameworks change.

Governance in practice

Governance is also a topic on a much more practical level, related to data stewardship.

Robert Herrenbrück presented a poster about the establishment of a central helpdesk for KonsortSWD (NFDI4Society), (https://doi.org/10.5281/zenodo.16735332) illustrating governance through support: coordination and service structures that help researchers navigate responsibilities without losing sight of their own work.

Somewhere between the crowded poster stands, I found myself in a memorable discussion that perfectly captured some practical governance challenges. The discussion was aboutt templates for data sharing agreements. Someone mentioned that when adapting the template to their needs and referencing the original template, they often get feedback such as “the template is different, why did you deviate from it?”. I think this highlights that while templates provide a much-needed baseline, they shouldn’t become rigid dogma. Real collaborations need room for adaptation, and governance should support flexibility rather than punish it.

If governance defines the rules and structures that make infrastructures sustainable, communities are the people who put these rules into practice, train new researchers, and ensure that collaboration thrives.

Training & Communities

While governance sets the framework, it’s the communities of researchers, data stewards, and trainers who bring it to life. From workshops to collaborative projects, these people ensure that research infrastructures are not just functional, but actually usable and sustainable.

As an outsider I must say that I am impressed about all the different initiatives represented at CoRDI. Through discussions at the conference I have learned that there are also other non-NFDI initiatives in Germany, some of them older than NFDI. For example the different federal state RDM initiatives like bwFDM (shout out to The Länd) which also had posters this time. They jointly presented themselves on the last CoRDI in 2023 (10.52825/cordi.v1i.242, see also my previous conference report 10.59350/pg3xj-4z449).

Other initiatives are the Fachinformationsdienste (FIDs), that in its current form receive funding since 2014. I personally know the thematic FIDBenelux as it relates to Belgium and because they are also in the follow-up committee of the MetaBelgica project. At CoRDI, Reinhard Altenhöner and Jana Fabrizius presented about the Role of FIDs Between Disciplinary Research and National Infrastructure (10.5281/zenodo.16736302). I have learned that since 2018 FIDs are self organized and currently have long-term funding via FIDPlus.

Regarding RDM training, two posters stood out to me as memorable and insightful. One covered the general aspects of collaboration, and the other covered the CARE principles.

The poster Lost in Collaboration really hit the nail with the identified difficulty for effective collaboration:

“Turning Collaboration into Concrete Outcomes.”

At the second poster I joined an interesting discussion about the CARE principles. Josef Jeschke explained that apparently there is no clear definition in CARE about indigenous, for example, it does not have to be a minority. In general I find the ethics aspect of CARE interesting, so I am wondering if it could be applied even in non-indigenous use cases. More info on their poster The Informed CARE Data Steward. Ways to CAREification: https://doi.org/10.5281/zenodo.16736131.

Reflections & Takeaways

Looking back at CoRDI 2025, I was struck by how often the same themes surfaced across very different tracks. Whether we were talking about workflows in the life sciences, authority files in the humanities, or large-scale infrastructures, the same questions kept coming back: how do we connect, how do we sustain, and how do we collaborate?

I’m glad that this community exists. PIDs and standards are probably not the most exciting topics, but they are key and enable so much more! So far we are preaching to the choir, but according to me, research infrastructures must enter main stream discussions. It’s not just about some tools or consortia, it’s nothing less than a shift in how research and research training works. And due to the scaling this also should be a shift in how funding schemes work and how we train new generations of students and researchers.

If there is one thing missing at CoRDI, it is more European and international contributions. But I think we are on a good way. During the open panel discussion about how to measure the success of NFDI, Joanne Yeomans from the Dutch Thematic Digital Compentence Centre, publicly brought up something that I was thinking as well: the success of NFDI is that NFDI is there in the first place!

We, from other European institutions can reference to the NFDI/CoRDI, simply because it exists. Even if it would be only a small initiative it would already be great, but the scale of NFDI makes it even better: all applied research and lessons learned stem from collaborations of all major universities and research institutions of the most populous country of the EU.

Or in the words of the keynote speaker Cathrin Stöver

a community needs it conference.

I hope CoRDI will become the research infrastructure conference, beyond NFDI. Hopefully see you all at the next CoRDI!

References

Fischer, B. K., Eckart, T., Körner, E., Buddenbohm, S., Al-Eryani, S., & Gradl, T. (2025). Authority Files and the Text+ Data Space. Zenodo. https://doi.org/10.5281/ZENODO.16736146
Simons, O., & Moeller, K. (2025). Wikibase - the best software to communicate with the upcoming knowledge graphs? Zenodo. https://doi.org/10.5281/ZENODO.16892677
Thiery, F., Rossenova, L., Mietchen, D., Homburg, T., & Thiery, P. (2025). Distributed Research Data Knowledge Graphs - Challenges of federated queries using the Wikiverse and OpenStreetMap within the NFDI Knowledge Graph Ecosystem. Zenodo. https://doi.org/10.5281/ZENODO.16736048
Rossenova, L., Limani, F., Fortmann-Grote, C., Kaplan, A., Thiery, F., Latif, A., Shigapov, R., Zapilko, B., Schubotz, M., Fliegl, H., & Tietz, T. (2025). How are NFDI consortia using Knowledge Graphs?: An overview of common functions and challenges by the Working Group “Knowledge Graphs”. Zenodo. https://doi.org/10.5281/ZENODO.16736078
Hofmann, V., Soylu, M., Preuß, G., Fathalla, S., Kulla, L., Dmello, F., & Sandfeld, S. (2025). The Helmholtz Knowledge Graph - towards a Helmholtz FAIR data space. Zenodo. https://doi.org/10.5281/ZENODO.16736336
Ondraszek, S. R., Tietz, T., Posthumus, E., Fliegl, H., Waitelonis, J., & Sack, H. (2025). Indexing Historical Research Data: MemO and the NFDI4Memory Knowledge Graph. Zenodo. https://doi.org/10.5281/ZENODO.16736125
Thönnißen, J., Oliveira, S., Oberstrass, A., Kropp, J.-O., Gui, X., Schiffer, C., & Dickscheid, T. (2025). A Perspective on FAIR and Scalable Access to Large Image Data. Zenodo. https://doi.org/10.5281/ZENODO.16736220
Waitelonis, J., Tietz, T., Beygi Nasrabadi, H., Bruns, O., Posthumus, E., Fliegl, H., & Sack, H. (2025). NFDIcore 3.0: A Mid-Level BFO2020-Compliant Ontology for Sustainable Research Data Interoperability Across Consortia. Zenodo. https://doi.org/10.5281/ZENODO.16736251
Castro, L. J., Gaignard, A., Juty, N., Schnitzer, H., Reed, P., & Goble, C. (2025). Bioschemas and Schemas.science at NFDI. Zenodo. https://doi.org/10.5281/ZENODO.16735850
Solanki, D., Ciuciu-Kiss, J. T., Quiñones, N., Ravinder, R., Venkatesh, S., Rebholz-Schuhmann, D., Garijo, D., & Castro, L. J. (2025). FAIR4ML, a vocabulary to describe Machine/Deep Learning models. Zenodo. https://doi.org/10.5281/ZENODO.16735334
Quiñones, N., Geist, L., Ravinder, R., Solanki, D., Venkatesh, S., Rebholz-Schuhmann, D., & Castro, L. J. (2025). MLentory: NFDI4DS registry for machine learning models and related artifacts. Zenodo. https://doi.org/10.5281/ZENODO.16735316
Herrenbrück, R., & Bayer, S. (2025). Establishing a Central Helpdesk for KonsortSWD - NFDI4Society: Goals, Challenges, and Solutions. Zenodo. https://doi.org/10.5281/ZENODO.16735332
Brand, O., Bruhn, K., Cyra, M., Fingerhuth, M., Gerlach, R., Jacob, B., … Weiner, B. (2023). The Federal State Initiatives for RDM as Intermediaries in a Dynamic Landscape of RDM Infrastructures and Services. Proceedings of the Conference on Research Data Infrastructure , 1. https://doi.org/10.52825/cordi.v1i.242
Lieber, S. (2023, September 17). CoRDI 2023. Posts | Sven Lieber. https://doi.org/10.59350/pg3xj-4z449
Altenhöner, R., & Fabrizius, J. (2025, August 4). Bridging Communities: The Role of FIDs Between Disciplinary Research and National Infrastructure. 2nd Conference on Research Data Infrastructure (CoRDI), Aachen, Germany. https://doi.org/10.5281/zenodo.16736302

Clustering Book editions

Mon, 16 Oct 2023 19:00:00 +0100

What do the books “The invention of Nature” and “De uitvinder van de natuur” have in common? Well, they are both different versions of the same work “The invention of nature” by Andrea Wulf, whether it is in a different format or a different language. In this blog post I will briefly introduce the advantages of keeping work-level records in library catalogs. Furthermore, I will introduce a fast Python implementation (DOI: 10.5281/zenodo.10011416) which we used in the BELTRANS project to identify the works in a corpus of book translations #FRBRization.

Context

Library catalogs usually contain records about different versions of a book. In technical terms, these records are so-called manifestations, but for simplicity I will refer to them simply as book versions or book editions. Such a conceptual division between different levels of description is part of the Functional Requirements for Bibliographic Records (FRBR) or how it is referred to nowadays, the IFLA Library Reference Model (IFLA-LRM).

Better user experience and better data quality due to identified works

Imagine that you are searching for a specific book in a library catalog. You remember that it had the word “invention” in its title. Instead of seeing two search results “The invention of nature” by Andrea Wulf and “Invention: A Life of Learning through Failure” by James Dyson, you are overwhelmed with hundreds of versions of both books and others such as ebooks, paperbacks, reprints, etc. Actually you are interested in finding out which works the library has that contain the title. Which specific version maybe is of less priority to you. In such a scenario it might be better to only show works in the search results. After clicking on the work you looked for, you still can be presented with all the different versions of the book.

There is also an advantage for the library personnel that works with the book data. Creating catalog records often is a manual process, or at least it was for a long time. Some of the records might contain more information than others, for example the genre according to a specific categorization. Identifying all the different versions of a book in your own catalog or in a collection of (other) catalogs can help to fill gaps in the data. If one book record is classified as a Comic and mentions a specific person as its illustrator, another record of a different version of that book which maybe has no such information, can be completed/enriched with this information.

Alright, you convinced me, how can I do it?

There are different ways to do this. I have read a bit of literature and it seems that there is one very common way to do this. Let me introduce you to the work-set clustering algorithm, initially devised by OCLC. The idea is simple: create several possible descriptive keys of a book version based on a combination of work-level information such as title and author, then check if two book versions have at least one key in common. Sounds complicated? Check out the image below which hopefully makes it clear.

In this example you can see two book records (red), each describing a book version. For each book version descriptive keys (blue) are generated to assist in finding matches. As you can see, both records have at least one key in common. We found a match! In this case only because we also considered a possible original title as part of the key. Similarly we would find a match between the shown Dutch version of the printed book and a hypothetical ebook version: the ISBN would be different, but there would be a match based on the title-author or title-translator combination.

Nice idea, how do I write a program to do this?

That’s the one million dollar question. Different scientific papers talk about the algorithm, a few share code. But a code implementation always depends on the context it was written for, for example by using XSLT rules for XML records. I tried to write a program in Python that is as generic as possible and came up with three solutions. Why three? Well, it turned out the first two were not fast enough for the amount of translations we have in the BELTRANS project.

First attempt: reusing the Python package sklearn

Most of the papers I have read are already a few years old, they mainly use software stacks that are less common nowadays. I needed a solution that I can quickly run without a large setup. Currently a lot of code libraries for common languages such as Python exist. So why not reuse an existing one?! The class AgglomerativeClustering from the Python package sklearn implements a hierarchical clustering algorithm.

Actually I am not interested in different levels of clusters, only the highest level with the least number of clusters. So I did what a lot of people do nowadays, I had a chat with ChatGPT to brainstorm possible solutions with AgglomerativeClustering. And it came up with a smart idea!

The input for the clustering is a so-called distance matrix: basically a table listing all elements as rows and all elements as columns, where the value in a cell indicates how similar an element in row X is with the element in column Y. We don’t need a sophisticated similarity measure, a simple “yes” for an overlap would suffice.

ChatGPTs idea was to use negative distance values. Like this, the whole clustering just takes two lines of code where I instruct the AgglomerativeClustering library to return clusters with distance threshold zero.

from sklearn.cluster import AgglomerativeClustering

model = AgglomerativeClustering(n_clusters=None, \
                                affinity='precomputed', \
                                linkage='single', \
                                distance_threshold=0)

# get a list where each index 
# corresponds to an element in elementIDs
# e.g. [0,2,0,1] the first element is in cluster 0, 
# the second in cluster 2, 
# the third in cluster 0 and the 4th in cluster 1
clusterLabels = model.fit_predict(distanceMatrix)

This went well for a relevant subset of roughly 20 thousand records in our BELTRANS project. After a few minutes we got correct and directly reusable results! However, when we wanted to achieve the same for more than 150 thousand records, this implementation of the clustering was not performant enough. Even on a powerful server with more than 32 GB of RAM, the computation of the distance matrix takes too much time and space.

One element likely has very few or no matches with all the other elements. A sparse matrix can be used, where not every value is stored, only not-zero values. Unfortunately the AgglomerativeClustering module does not support this.

Second attempt: implementing the algorithm myself

A little disappointed from the last solution, I thought I will try implementing it myself, all I need: a few data structures and loops. The idea of the algorithm is to start with every element in its own cluster. Then comparing it to other clusters until we find a match. When a match is found the loop is stopped, the clusters are merged and we loop again over the (now updated) list of clusters. We do all this until no more merges are possible.

This algorithm does not need a large distance matrix to start with. Testing if there is an overlap between two clusters can be reduced to a single intersection operation between two Python sets that represent the descriptive keys of the respective clusters.

Unfortunately all that takes too long. With a few example data I could verify that it worked as expected, but as soon as there are a few thousand elements the algorithm keeps running …

I tried to remedy the situation by putting parts of the computation in a reusable function that I can call in parallel by using ThreadPoolExecutor or ProcessPoolExecutor, but it did not help much.

Third attempt: inverted index for clustering in no time

Even more frustrated I went back to the drawing board, or pen and paper in my case. What I need is a solution that ideally only runs once over all elements that need to be clustered. But how do I find out possible overlaps without comparing everything?

The disadvantages of the previous solution were that I had to perform set intersections to detect overlaps while iterating over all clusters, stop the iteration to update the clusters and start iterating again.

Actually the data already includes the matches between elements implicitly, and I used this to compute clusters in no time. Every element has one or more descriptive keys. But instead of only storing the mapping element->descriptive keys in a Python dictionary, I also store descriptive key->elements in another dictionary. This is my inverted index which I can use to look up overlaps.

Eventually I start with zero clusters and I simply iterate over the inverted index. If one of the elements from a key already is part of a cluster, I add the other elements to it. If no cluster was found, I create a new one and add the elements. Basically all elements of a descriptive key I check, either go to an existing cluster or end up in a new cluster. If different elements for a descriptive key are already in different clusters, a new merged cluster is created and the old ones deleted.

This solution works as well, and wow, it is fast too. Interestingly it takes longer to upload the computed cluster information to our database than performing the clustering itself. Inverted indexes can do a lot of heavy lifting!

You can find the source code on GitHub (DOI: 10.5281/zenodo.10011416).

Why is it so fast?

It all has to do with scalability. I only iterate once over all elements to build the inverted index dictionary. Then I iterate once over all elements in the inverted index to compute the clusters. For 100 input elements this roughly means iterating 200 times, for 100,000 input elements roughly 200,000 times. An algorithm that scales linear to its input.

Compare this to compute a distance matrix of 100*100 elements (compute and store 10 thousand results) and 100,000*100,000 elements (compute and store 10 billion results). An algorithm that scales quadratic to its input.

Final remarks

Having work-level information brings advantages for both librarians and users. The presented solution is just one possible implementation of one possible algorithm. Which data fields you use to create descriptive keys and how you extract them from your bibliographic data explicitly was not covered in this post. I also did not cover how you could use the information of the clusters to actually achieve the advantages. The main purpose of this blog post was to introduce and share the generic Python implementation, because I had the feeling there is not much out there that can be reused easily.

For the BELTRANS project we extracted descriptive keys with a SPARQL query from an Resource Description Framework (RDF) representation of bibliographic information. Similarly, we used SPARQL INSERT queries to explicitly create RDF representations for each work cluster and link its manifestations to it using the fabio ontology. Here you can have a look at our SPARQL queries.

Curious about more content like this or some behind-the-scenes? Consider subscribing to my bi-weekly Newsletter FAIR Data Digest to receive more interesting content about Linked Data every other Tuesday!

CoRDI 2023

Sun, 17 Sep 2023 09:00:00 +0100

It is difficult to foster innovation and perform high quality research when the underlying data is not accessible and usable. Data and data services should be offered via so-called research data infrastructures, think of user-friendly web platforms, repositories, and databases. But how does it work in practice, and how important is the human aspect of it? I have been at the 1st Conference on Research Data Infrastructure, 12 – 14 September 2023, and got some pretty good insights. In this blog post I will reflect on the conference and cover the keynotes, the poster sessions, our own work around Knowledge Graphs and Wikibase, as well as a selection of other interesting talks.

Context

(Research) data is produced in various fields, ranging from cultural heritage and agriculture to physics and biology for example. Such data should be managed and preserved in a systematic way to accelerate scientific progress and promote collaboration, while enhancing public trust and fostering innovation.

Several large research (data) infrastructure initiatives emerged on European Level in the past decade. Those initiatives, including the European Strategy Forum on Research Infrastructures (ESFRI) or the European Open Science Cloud (EOSC), have brought hundreds of institutional partners together to build and maintain research infrastructure within numerous research projects. National initiatives have emerged as well.

Under the theme Connecting Communities, the association related to the German National Research Data Infrastructure (NFDI) organized the 1st Conference on Research Data Infrastructure (CoRDI). More than 680 persons attended this international multi-track conference which hosted 84 presentations and more than a 100 posters. Next to inspiring keynotes, there were thematic tracks at the first day and tracks related to enabling, connecting, linking, harmonizing, securing, and spreading of Research Data Management (RDM) on the remaining days.

My main takeaways

It’s mostly about human problems that need to be fixed (connecting communities, data governance, etc)
Knowledge Graphs and especially Wikibase gain attention in the context of research (data) infrastructures
There is an increasing need for sustainably funded data management careers

In the rest of this post I will focus on my main highlights, not on all the presentations that I have seen. Also, since the conference was organized in the frame of NFDI, a lot of networking and promotion was devoted to the different NFDI consortia. I am not affiliated with any of these consortia and because I haven’t worked with any of those consortia in the past, this blog post is limited to my own personal (outsider) experiences at the conference. I am glad to update the post and link relevant other blog posts that provide other insides.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

Table of Contents

Keynotes

Three excellent keynotes were given which gave different insights into data: Christine Borgman zoomed into the bigger picture of Knowledge Infrastructures, Julia Janssen presented how data and scientific insights can be translated into art installations, and Mark D. Wilkinson talked about his regrets when co-designing the first FAIR principles and the road ahead.

Christine Borgman – Knowledge Infrastructures

“May all your problems be technical” – Jim Gray

Christine Borgman ended her keynote (DOI: 10.5281/zenodo.8344854) with this quote from Jim Gray. And after listening to her presentation one really grasps how much truth lies in this statement. From a research data infrastructure point of view we have already a lot of platforms and initiatives. But what about data governance? To which extent do different communities agree on data management practices? According to her, we should focus on Knowledge Infrastructures, including also the social component!

One really interesting point she made is the difference between knowledge production as done by researchers and data production for others as done by data creators. Christine Borgman pledges for a data management workforce: domain data scientists, data librarians and data archivists who cover the middle ground between knowledge and data production. I fully agree! More such careers are needed (and sustainably funded) at different research institutions.

Julia Janssen – data art installations

Data, .. this anyway is something really abstract. The researcher and artist Julia Janssen has shown us in her keynote how data can be made tangible. She introduced us to some of her art installations and why and how she came up with it.

For example, according to her data from Meta (the company behind Facebook, Instagram and WhatApp), she likes the color green. She built a whole art installation with more than 4000 small balls where each ball stands for one of her interests, according to Meta. Just to try figuring out why she might like the color green. Other people literally crawled under it to get to know her better, looking at her interests trying to find out why she might like the color green.

Mark D. Wilkinson – FAIR principles

A common theme at the conference were of course the FAIR principles. The third and last keynote about those principles was given by the main author of the original FAIR principles paper, Mark D. Wilkinson (DOI: 10.5281/zenodo.8353153). Besides presenting his most recent work, he also talked about his regrets after almost 10 years since the initial version of the principles from a 2014 workshop. One of his regrets is the term data, apparently the term FAIR Digital Objects or FAIR Research Objects seems to capture their ideas better with less confusion. Also the term license turned out to be an issue, access or usage policy would have been better, avoiding the heavy legislative meaning.

But looking forward, Mark D. Wilkinson talked about (not so) autonomous agents and FAIRification in different use cases. It was nice to see that they are also using the RDF Mapping language (RML) and the YARRRML serialization format to accomplish FAIRification. Those specifications and implementations for Knowledge Graph Construction were developed by ex-colleagues of mine and I am also still using it a lot.

Poster session(s)

One of the most engaging events during a conference is of course the poster session. And CoRDI had two of it! Many, many posters and as much motivated researchers to present it. Topics covered among others anything around data management plans, terminologies, persistent identifiers, FAIR initiatives and of course posters presenting the different NFDI consortia. From the plenty of posters I will focus on two in this blog post: one more technical poster and one highlighting international collaboration.

There have been posters from various domains, and it became apparent that there is an overlap between the domains in terms of problems and solutions. For example, there was the poster from Steffen Neumann and colleagues about repository federations in the chemical domain (DOI: 10.52825/cordi.v1i.202). Among others, they work with the Open Archives Initiatives Protocol for Metadata Harvesting (OAI-PMH), which is also a common interface in the library domain where I am currently in. There I have learned that it is possible to embed JSON-LD in OAI-PMH and retrieve it via Xpath. A very handy way to publish Linked Data using already available interfaces.

Considering the Humanities domain, there was an interesting poster from Nanette Rißler-Pipka and colleagues, including my colleague Sally Chambers. (DOI: 10.52825/cordi.v1i.274). The poster focused on synergies between national and European research (data) infrastructures based on the example of the SSHOC market place. In the 8th edition of the FAIR Data Digest, I have covered with an example how differently national ERIC nodes might be linked with national initiatives (what is an ERIC? check here). Nice to see it and talk about it in real life.

MetaBelgica, Wikibase and NFDI4Culture

I participated at CoRDI to present our envisioned platform MetaBelgica which aims to become the single source of truth for cultural heritage metadata about Belgian entities of type persons, organizations, time/events and locations.

Abstract: https://doi.org/10.52825/cordi.v1i.381
Presentation: https://doi.org/10.5281/zenodo.8337301

For those of you subscribed to my FAIR Data Digest newsletter, you already read about the main idea behind MetaBelgica in the last edition. The following image summarizes the problem we aim to solve.

Each participating institution (left in the image) manages metadata about their own collections using domain-specific data standards. This is totally fine, because usually there aren’t any conflicts, because there is no overlap in the collections. But as part of the cataloging process, the same related data about entities such as persons or events are created by each institution separately. Over and over again. This results in duplicate efforts for the institutions and for the users.

We want to integrate the data of the partnering institutions by using the Resource Description Framework (RDF) and use the Wikibase software as a platform to manage the integrated data in a Knowledge Graph.

Knowledge Graphs at NFDI

With our submission related to Knowledge Graphs we are not alone!

Under high pressure of continuously ringing alarm bells during the national-wide sirens and (phone) alarm system test, Lozana Rossenova presented work about Knowledge Graph Infrastructure based on Wikibase (DOI: 10.52825/cordi.v1i.266). Wikibase instances are already deployed by the German National Library of Science and Technology (TIB) in the context of NFDI4Culture.

Speaking of which, Harald Sack also presented work that originated from NFDI4Culture (DOI: 10.52825/cordi.v1i.371). Their original NFDI4Culture ontology was extended to cover other NFDI consortia. In that adaptation process, the ontology was modularized into a core and several extensions.

This is great! Plenty of lessons learned and possible networking and collaboration ahead with respect to MetaBelgica.

Selection of presentations

There were tons of interesting presentations, but unfortunately I could not split myself to see all. Here is a small selection of some presentations from various domains such that you will get an idea of research data infrastructure solutions at CoRDI.

I have learned about the term microdata: information at the level of individual response of surveys (used in social sciences or official statistics). Dana Müller presented work about geographically separated safe rooms as part of the International Data Access Network (IDAN). In those rooms, data can be accessed without the having to leave the host institution (DOI: 10.52825/cordi.v1i.269). When looking at the following picture I am remembered of my MetaBelgica presentation, I am wondering if with respect to safe rooms there is a Belgian gap in social sciences too :-)

Data spaces

Peter Wittenburg presented the idea of FAIR Data Objects (FDOs), which according to his vision should become similarly to TCP/IP a standard protocol for data exchange (DOI: 10.52825/cordi.v1i.263). Like this, FAIR data can flow through the emerging data spaces similar to how TCP/IP packages flow through the internet.

Physics

Not my cup of tea, but apparently there are a lot of physical sciences data infrastructures in the UK. And as we have learned further in the presentation given by Nicola Knight, there are also alignments with the NFDI (DOI: 10.52825/cordi.v1i.339).

Scholarly Communication

Even though we have all the open and FAIR data and code repositories, the actual way of scholarly communication via research papers has not change much over the centuries. Sören Auer presented the Open Research Knowledge Graph (ORKG) and how they plan to change that (DOI: 10.52825/cordi.v1i.272). One adds structured information about the contributions of a specific paper to the Knowledge Graph via a web interface. Comparisons and visualizations can be created on the fly, everything gets DOIs and is thus citable. They also experimented already to promote the ORKG earlier in the publication process, for example as preparation for the submission to the SEMANTiCS conference according to Sören Auer.

Research Data Management

A whole team of people presented the state initiatives for Research Data Management (DOI: 10.52825/cordi.v1i.242). The following image shows some German states that are already participating in this initiative. They do some necessary field-work, providing support to researchers directly, or support smaller members such as Universities of Applied Sciences in grant applications.

Networking

Last but not least, Nina Leonie Weisweiler presented the engagement of the Helmholtz association in NFDI (DOI: 10.52825/cordi.v1i.402). The following image nicely shows which Helmholtz centre is active in which NFDI consortium. Internally, the different NFDI representatives of Helmholtz regularly sit together to discuss their involvement, great initiative!

Connecting the dots as the title mentions, or connecting communities, the theme of the conference!

Final remarks

This is it, I hope you enjoyed reading this blog post about CoRDI 2023. Let me know if you have any questions, feedback or remarks.

Thanks to the whole CoRDI team for making this edition happen! I am looking forward to a possible 2025 edition of CoRDI to learn about the progress of the community.

In the meantime, there will be another NFDI event in 2024, but apparently slightly smaller, because the combined NFDI and EOSC communities are too big for a suitable venue.

Don’t forget to share this post on social media and in your networks! Please also consider subscribing to my bi-weekly Newsletter FAIR Data Digest to receive more interesting content every other Tuesday!

RDF named graphs

Mon, 26 Jun 2023 19:00:00 +0100

With the Resource Description Framework (RDF) you can represent Linked Data as subject-predicate-object triples. But what if you have to represent additional context of those triples? In this Blog post I will briefly introduce the advanced data management feature “named graphs” with two examples: different data source named graphs in the BELTRANS project and different story-contexts for Digital Humanities research. Furthermore I will graphically illustrate the different ways you can query named graphs with SPARQL!

Context

Linked Data represented using the Resource Description Framework (RDF) consists of subject, predicate and object triples! This is what you will learn in any introduction to Linked Data.

However, did you know that with RDF you can also have Quads? Statements with four components: subject, predicate, object and context. Continue reading for some practical examples and graphical illustrations.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter FAIR Data Digest.

Example of named graphs for provenance information

According to the Data Management chapter of the Linked Data Patterns website, named graphs can be used for different purposes such as expressing the provenance of triples, use the named graph for access control, or in general for data management.

In the BELTRANS project where we collect data from various sources, we store the data of each data source in a separated named graph. One the one hand, this allows us to update a specific data source independent of other data sources, and on the other hand, we can refer to a triple by its data source as the following example illustrates.

We have a named graph containing RDF triples that describe book translations in our project database (<http://project-db>). Each of this book descriptions might be available in data that we obtained from different national libraries (our data sources <http://kbr-data> and <http://bnf-data>). This is represented by having a schema:sameAs link from the book description in the project named graph to the different data source named graphs.

Without named graphs, the generic schema:sameAs link would not allow us to distinguish the descriptions of the book in one particular data source by using SPARQL.

A side note: it might be possible by using a FILTER expression on the string representation of the URI. For example, if URI starts with http://data.bnf.fr it probably is a book from the National Library of France. However, this is a bit against the idea of the Semantic Web, because now we assume semantics based on the name of the thing and not via explicitly expressed semantics with RDF terms. This is considered a bad practice.

With named graphs, however, we have a way to semantically express the context of a book description using explicit RDF statements. In a SPARQL query we could do the following if named graphs are used for data of each data source:

SELECT ?bookURI ?kbrURI ?bnfURI
WHERE {

  # we are interested in book descriptions from our project database
  graph <http://project-db> { ?bookURI a schema:CreativeWork . }

  # based on the named graph context, identify the KBR book description
  OPTIONAL { 
    graph <http://project-db> { ?bookURI schema:sameAs ?kbrURI . }
    graph <http://kbr-data> { ?kbrURI a schema:CreativeWork . }
  }
 
  # based on the named graph context, identify the BnF book description 
  OPTIONAL {  
    graph <http://project-db> { ?bookURI schema:sameAs ?bnfURI . } 
    graph <http://bnf-data> { ?bnfURI a schema:CreativeWork . } 
  }
}

In this example, the specific URI linked via schema:sameAs also has to be a book description in a specific named graph. Like this, we can distinguish in the SPARQL query based on the named graph context, in which schema:sameAs we are interested in.

Illustrations of different ways to query named graph data

You have already seen that I have used the graph keyword in the SPARQL query above. But there are more possibilities! To illustrate these possibilities, namely the keywords FROM and FROM NAMED, I will consider a slightly different example, inspired from a presentation at the DHBenelux conference 2023 about Greek Mythology.

In the following example, we are interested in the fictional character john and his place of birth according to different novels. Each novel is represented as a different named graph, i.e. <http://story-1>, <http://story-2>, and <http://story-3>.

Somewhere in the first novel, it is mentioned that john was born in ex:town1. The second novel even specifies that ex:town1 is located in country ex:country2. However, the second novel does also state that john was born in ex:town2. Last but not least, according to the third novel, the city ex:town1 is in country ex:country1.

This example should illustrate that each novel provides its own context. The john of novel 1 might not be the same of the john in novel 2. Similarly, one could say that ex:town1 is the same as ex:town2 or that ex:country1 is the same as ex:country2.

But, are these entities actually the same? This might be a complex and even philosophical discussion. Instead of explicitly indicating that something is the same as something else, we can simply keep the context of each story. What we know is that ex:john appears in all three novels.

Let’s assume, that the fact that john is a (fictional) Person (e.g. by indicating a schema:Person) is specified without any specific context (named graph). The tricky part is now how to query the data. Depending on your use case you have different options which are visualized below.

In the illustration we see that each query returns different results. Let’s have a look what each of the queries does. One remark: The values in the output that are strikethrough will not be part of the result, because not every variable of the triple pattern will be bound. If you would wrap the triple pattern to fetch the country into an OPTIONAL, you would get those results too.

(1) In the first case no graph is explicitly specified. Thus the default graph will be queried. Depending on the triple store (RDF database) you are using, the default graph might be yet another named graph which in our example is almost empty. Remember, it only contains the triple that ex:john is a person. However, in this example we assume that the default graph is a union of all named graphs: an inclusive strategy.

(2) In the second case, we explicitly specify with the FROM clause, which named graphs the query should consider. If you do not specify any context in the triple patterns of the query, the query looks into the union of the specified named graphs. As expected, data of <http://story-3> is not queried and will thus not be part of the output.

(3) In the third case, we similarly specify named graphs explicitly, this case with the FROM NAMED clause. Like this the query loops over each of the mentioned named graphs. This means, that when producing one result set (one row of the output), the variable for the named graph is bound to only one named graph. Therefore it is not possible to query across named graphs. And this is exactly why for the given data, no results will be provided.

As you can see, a named graph is also just a URI! Therefore we can use properties to describe the context as well. For example

<http://story-1> a schema:Book ; 
                 schema:about ex:john ;
                 schema:author ex:author .

And where you store this information is up to you. Now you know how you can retrieve it with SPARQL!

Final remarks

Named graphs can be a powerful data management technique. Yet you have to be aware of different implementations, especially if you change your software stack at some point. Additionally, it makes SPARQL queries slightly more complex which may look scary for beginners.

Besides the RDF and SPARQL specifications, you can find more information about named graphs at this Blog post from the company metaphacts, or at this Blog post from the company Stardog.

Knowledge Graphs for Data Integration - UHasselt 2023

Mon, 12 Jun 2023 09:00:00 +0100

In June 2023 I attended a workshop in Hasselt, Belgium, organized by a research network on Knowledge Graphs and Data Integration. Linked Data-based systems and declarative SPARQL queries somehow magically do all the work efficiently, but how? For many people, presentations about using Linked Data and SPARQL already seem very technical. But this event went further, it covered the fundamental algorithms that power the Linked Data systems! In this post I try my best to briefly summarize some of the talks and highlight the added value of those fundamental and often behind-the-scenes work.

Context

The FWO-founded research network Knowledge Graphs and Data Integration organized a workshop at the Data Science Institute of Hasselt University in Belgium. As part of this workshop, a keynote about a logical approach to model interpretability was given by Marcelo Arenas from the Pontificia Universidad Catolica de Chile. Afterwards, several papers were presented in three different sessions: foundations, experience and exemplars, and processing and mining. I will group my summaries in these categories as well.

My overall takeaway is that there is still plenty of research for Knowledge Graphs! Not only for how to use them or how to encode knowledge in them, but also for how to actually implement Knowledge Graph software based on formally sound techniques.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

Table of Contents

Keynote

Before I focus on the keynote itself I would like to start with an example. Recently in a show of the Flemish comedian Lieven Scheire, I heard about a Machine Learning (ML) algorithm that was trained to distinguish images from wolves and from huskies. The algorithm seemed to perform very well, but not always. Later, by trying out a few things, it turned out the algorithm apparently “learned” that the distinguishing factor is a white background (snow). So if there was white background, the animal on the image was classified as a husky image by the ML model.

This is just an example to illustrate that we don’t actually know what a ML algorithm/model has learned and thus why a certain output is generated. That’s why such models are usually called a black box. But in a world in which more and more ML enters our daily lives, it is absolutely fundamental that we are able to explain why the model behaves in a certain way.

Explaining predictions of ML models

In simple terms, the keynote of Marcelo Arenas presented some fundamental research in early stages that tries to specify a declarative query language (think of SQL or SPARQL) for ML models. The keynote lecture had some very nice examples to illustrate the graph algorithms. But why graphs actually?

Machine Learning models are usually based on so-called “neural networks” that are inspired by neurons in the human brain. These neural networks that may consist of several layers of neurons “learn” patterns by adjusting how information flows through the connections between the nodes. This is not a Knowledge Graph in a Linked Data-based sense in which you explicitly encode information that you can query later. But it is a graph and hence one can perform graph algorithms that include some sort of queries.

A logic-based query language for ML models

In his talk, Marcelo explained how they define theoretical logics for operations on such graphs. Their overarching goal is to develop a declarative query language, such as SPARQL, but for neural networks. This is not purely theoretical as they also implemented their “FOIL” logic using a special logic programming language that is different to languages such as Python or Java that you may know. Like this the theories can be tested with real life data and it can be observed how effective they are with specific data (besides the theoretical complexity). If you are interested in more, here you can find a preprint of an earlier paper: https://arxiv.org/abs/2110.02376

Foundations

From this session, I will focus on three of the talks. Two that focus on effective computations and one about reliable query languages.

Graph compression

The work of Jeroen Bollen is motivated by the fact that memory is often a bottleneck for so-called Graph Neural Networks. There are ways to compress the graphs. But so far there haven’t been any formal proofs showing that the accuracy of the algorithm remains stable with differently compressed graphs.

In his talk, Jeroen presented his work where he showed that they could minimize computational and memory requirements while keeping a similar accuracy. Very interesting work that demonstrates that fundamental research can make a difference!

Effective reasoning

Other more formal work inspired by a real world use case was presented by Robin De Vogelaere. You probably have used glue (a cohesive) at least once in you life. Selecting the right adhesive in the right situation is not trivial. It depends on the material and in which conditions the product will be used (heat, shock, etc).

Robin presented the FO(.) first-order logic that they use in their IDP/Z3 Knowledge Base system to perform reasoning (on cohesives). In his presentation he highlighted why the Knowledge Graph Community should care (see image below), A perfect example of the knowledge transfer the workshop aimed for!

Reliable query languages

Staying in the Linked Data RDF world: You might have used SPARQL already to query something from Wikidata via the Wikidata Query Service. But using some baked-in SPARQL functionality, you can actually write a federated SPARQL Query that fetches different parts of the data from different servers.

But what if one crucial FILTER part of the query relies on the data of a server that is currently offline? Should the whole query fail or would you accept data from the other servers that could not have been filtered (and which are wrong for your use case)?

Tim Baccaert presented some early work on exactly those problems. He is looking into different ways to make query languages more reliable. For example by providing better machine-understandable error messages such that clients can interpret partial results better.

This is one way of dealing better with the reality where servers are down sometimes. Another way is Linked Data Fragments (LDF): Instead of processing the whole possibly complex SPARQL query on the server, the LDF server only accepts simple patterns and the computation of the result is done on the server and the client.

Experience and exemplars

This session is about experiences in setting up Knowledge Graphs and in using the data.

Data dependencies

Understanding relationships between attributes of relational data is important to integrate data in a meaningful way. Marcel Parciak presented work on detecting (approximate) functional dependencies in data.

By performing a literature review before their experiments, they could identify 12 measures to detect possible dependencies. They performed an evaluation of the different measures on benchmark datasets. Before doing that they also had to annotate the benchmark data because there were no labels for when there is a functional dependency and when not.

Larissa C. Shimomura is researching another kind of dependency: Graph Generating Dependencies. Such dependencies are helpful to understand the data and possible correlations between different attributes. In particular, Larissa investigates reasoning on property graphs.

Rapid Prototyping

Declarative solutions following common standards are the goal when working with Knowledge Graphs in RDF. However, especially in the beginning of a project the requirements are often vague and different stakeholders need to build a common understanding.

In his talk, Christophe Debruyne presented work about the TOXIN Knowledge Graph. The aim of this Knowledge Graph is to describe existing safety data about cosmetic ingredients. It therefore contributes to non-animal systemic toxicity assessments.

In the beginning phase of the project many things were set up with Python. This helped in getting a shared understanding between the stakeholders before setting up a complete infrastructure. In later stages, R2RML and other declarative tools such as SHACL for quality assessments were used.

Technology stacks Knowledge Graph

While talking about infrastructure, the technology stacks of cloud platforms are quite complex. This means that switching between providers can become really cumbersome.

In his presentation, Johannes Härtel talked about how they used RDF to describe technology stacks. Furthermore, they have implemented a tool that harvests data about the software stacks from GitHub repositories.

For example, if you have a docker file, the script detects this and will add to the Knowledge Graph that you at least use YAML files and that you have at least one docker setup provided.

The goal of their work is to allow for a smoother switch between technology stacks. For example, to identify Open Source technologies that fit the current (proprietary) technology stack you are using. At the core, they used Apache Jena to store RDF. Johannes listed four reasons for why to use Knowledge Graphs (see image below).

Law as code

The transparency of legislation and the accountability of governments are very important for a democracy. In his presentation, Tom De Nies presented the tools Themis and Kaleidos which are used to support the legislation process in Flanders, Belgium with Linked Data.

The clue? There is no data conversion needed, their system works with RDF at its core! Data is stored, accessed and published from a RDF triple store. Input forms and other website components are created by using an EmberJS frontend that relies on a rich microservice infrastructure in the backend.

Processing and mining

Last but not least, the last session focused on mining information and effective query execution.

Mining software engineering patterns

Yunior Pacheco Correa presented work on mining software engineering patterns from code repositories. This can help to support a developer with auto-completion when writing code, such as GitHub Copilot. Another possibility would be to query the mined patterns to for example look for security vulnerabilities.

More efficient SPARQL queries

Three presentations from the IDLab KNoWS group focused on the behind-the-scenes of querying with SPARQL. SPARQL is a declarative query language: the user specifies what to retrieve and not how. Internally, the used SPARQL query engine processes the query and builds a query plan. For example, it determines how to join result patterns.

Bryan-Elliott Tam presented work related to guided link traversal query processing. So basically, how a RDF query interface can already provide hints about its data via the TREE hypermedia specification, such that the query engine can make better choices in query planning. Similar research was presented by Jonni Hansi.

Bryan’s research has shown that with SPARQL FILTER expressions and the TREE specification, they were able to prune links to Linked Data fragments that will certainly not contribute to the SPARQL query. This means the query will be faster!

Ruben H. Eschauzier presented early work on a ML model to learn effective joins in query planning. As a modular part of the Comunica query engine, the presented model can be used to predict the execution time of a join and tries to select the most efficient one. Currently this model only outperforms the Comunica templates in 7 out of 18 templates. There is thus still more research needed!

More on query optimization

The research of Wilco van Leeuwen centers around cardinality estimation: a technique used to estimate the number of items returned by a query or subquery. He is interested in obtaining the most accurate answers using the least amount of resources (time and space).

Compared to relational data, graph data introduce additional challenges for cardinality estimation. His research helps understanding different cardinality estimation techniques and their trade-offs. This could for example be applied to improve query planning on graph data, such as for SPARQL query planning.

Final remarks

Thanks to the organizers for planning this event and to the researchers to present their work! I think this has been a fruitful exchange of knowledge. I am looking forward to the following events.

Don’t forget to share this post on social media and in your networks!

Please also consider subscribing to my bi-weekly newsletter FAIR Data Digest to receive more interesting content every other Tuesday!

DHBenelux 2023

Mon, 05 Jun 2023 09:00:00 +0100

This year we had the honor to host the 10th anniversary edition of the DHBenelux conference at the Royal Library of Belgium (KBR). In this blog post I will reflect on the different sessions I’ve attended at this year’s multitrack edition of DHBenelux 2023. I will group my notes around the following four recurring topics of the talks that I have seen: (1) data platforms and their usability, (2) how digital methods such as Linked Open Data are used for DH research, the other way around, (3) how DH research studies digital methods and tools, and finally (4) research with (historical) geospatial data in cultural heritage.

Context

Digital Humanities (DH) is a discipline in which computational methods are combined with humanities research, both by systematically using digital methods and artifacts for humanities research and by studying the application of said methods.

To foster DH research in Belgium, the Netherlands and Luxembourg (Benelux), the DHBenelux conference was founded in 2014. Although the focus lies on these three countries, the conference is open to anyone interested and the conference language is English.

This year’s edition had several tracks and because I have a technical background and mainly work with data and research data infrastructure, I mainly followed more technical sessions. However, if you are interested in other talks, the abstracts of the conference are openly available in the DHBenelux 2023 community on the research repository Zenodo (persistent), or on the DHBenelux 2023 website.

My main takeaways are

Digital methods support the humanities research, but humanities methodologies can also positively influence digital tools and methods
the use cases of the diverse humanities research provide excellent food for thought for data management practices, for example what kind of biases exist and what good practices are
the conference aims at community-building and experience exchange which I find is the greatest distinction to the more competitive Computer Science conferences I’ve attended in the past (which have a much lower acceptance rate of for example 25% compared to the 90% of DHBenelux)

Because the conference took place at my workplace, I did not travel to the conference. This also meant that I commuted home and did not join most of the social programme (#partypooper). Therefore, and also because I volunteered during the conference for the organizers that are my colleagues, I will focus in this blog post mainly on the scientific presentations and discussions and less on the venue.

Instead of listing the talks chronologically, I will group my notes under the following recurring topics I have encountered in the talks that I have attended.

Data platforms and usability
Using humanities methodologies for digital resources
Using digital methods for humanities research
Geospatial research

There were also a few workshop the day before the conference, but I did not participate in any of the workshops as non of the workshops sparked my immediate interest and I anyway had some work to finish 🙂 This was only my second DHBenelux conference, but it seems that a combination of workshops about DH tools and Linked Open Data are common. This is great!

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

Table of Contents

Data platforms and usability

Humanities research revolves around a lot of valuable and super interesting data such as Libraries’ collections, digitized documents, or other media data. Different platforms and tools exist to work with this data in a user-friendly way. I would say even more than in Computer Science, user-friendly interfaces are a must for less-technical users in the Humanities. Therefore it is nice to see different efforts around usability.

There have been talks about usability of the CLARIAH Media Suite in the Netherlands, the upcoming DATA-KBR platform in Belgium, as well as the LuxTIME and the Tropy projects in Luxembourg. On top of that, other talks focused on data visualization.

Stakeholder Readiness Level in the CLARIAH Media Suite

Roeland Ordelman presented User-related research around the CLARIAH Media Suite. What I really liked is the idea to build something similar to the Technology Readiness Level (TRL) but for usability by different stakeholders: the so-called Stakeholder Readiness Level (SRL). I find this a very interesting and holistic approach that, on the one hand, can improve the usability of systems, and on the other hand could be used as a benchmark to compare the usability of systems.

My own research on usability has been a while, but I can clearly see how it could be connected or extended. For example, how a Stakeholder Readiness Level could be used to measure the customer journey when using public services (DOI: 10.5281/zenodo.8001638, FAST project) or to define requirements for a user interface that uses visual notations to visualize constraints of data constraint languages (DOI: 10.3233/SW-210450).

Difference between scholar and journalist users

Besides the presentation of Roeland about the general Stakeholder Readiness Level, Willemien Sanders talked about differences between scholars and journalists when creating data stories with the media suite.

There are differences with respect to speed, transparency, reflection, data/service provider involvement and sensitivity/copyright issues. More details about the differences are in the abstract. In my opinion this kind of research is very important. Not only to understand the user needs better, but also to improve the digital product and to come up with suitable tools and business models.

Lessons learned when building a consortium of stakeholders - LuxTIME project

Petros Apostolopoulos and Sean Takats shared their experiences from the creation of a consortium to analyze and interpret historical big data within the Luxembourg Time machine (LuxTIME) project DOI: 10.5281/zenodo.7986361.

They performed interviews that focused on the following four aspects: historical data, research initiatives, collaborations with other institutions, and future collaboration with the project and participation in the consortium.

These interviews were necessary, because detailed information about what historical data is held at the institutions or in which research initiatives the institutions are involved is often not available on their websites. The objective is to map the state of the art of historical data in Luxembourg.

Interestingly, some institutions do not consider their existing digitization initiatives as research projects, because, according to these institutions, these initiatives do not involve research.

One objective is not to complicate collaboration even further by needing to select representatives, but rather create a peer-to-peer collaboration. Additionally, regular meetings should be organized alternately at the different consortium members locations to strengthen the collaboration.

Turn photos into items

Anita Lucchesi and Sean Takats presented the Tropy platform. It helps in organizing research collections: imagine that you have a bunch of images in a folder, with Tropy you can annotate and organize the images. Among others, this allows at a later stage to export the items in standardized data formats.

Collection-centric approach

Sally Chambers presented updates on the DATA-KBR project that aims to optimize KBRs ICT infrastructure to stimulate sustainable data access. This contribution is therefore not only about an upcoming platform (for now internally a Nextcloud instance is used), but also about the collections themselves: collections as data.

The updates she presented focus on a researcher-centered and iterative approach to gather requirements. If you are interested to read more about how to publish collections as data, she referred to the pre-print A Checklist to Publish Collections as Data in GLAM Institutions DOI: 10.48550/arXiv.2304.02603

Focussing on collections and data instead of being too application-centric made me think of the amazing book Software Wasteland from Dave McComb. After all we want to study cultural heritage collections!

I got a similar vibe from a question that was asked during the talk of Loren Verreyen about her previous work on Radio’s ephemerality in DH. There she used textual data from the BBC programme index to study BBC producers of the Third Programme (1946 - 1967). Digitized documents from a broadcast listing magazine were used as a basis.

A question from the audience, whether corrected OCR data of the research project was fed back to the BBC portal, got my attention. According to Loren, this has not yet happened but would be very useful. I like this idea of a more collection-centric approach where different smaller or larger research projects continuously improve or extend a collection.

This made me also think of the Researcher-in-residence programme of the Royal Library of The Netherlands. In this programme, researchers have to use existing digital collections to perform research and improve or extend these collections (or create new collections/datasets). Like this, the collections get richer over time. By the way, as of writing this post, they are looking for a new researcher-in-residence, check out their website!

Another very interesting talk from Douwe Zeldenrust about the Records Continuum Model highlighted the evolution of collections.

In his talk, he presented meta research about the Dutch language in the United States that originally was performed by Jo Daan. More precisely, Douwe focused on the archiving process of research data. He did so based on the example of Jo Daan’s collection and several other collections that were added later during other PhD projects which partially built their collection on top of the one from Jo or criticized the original collection.

Visualizations and user interactions

When talking about users and usability, visualizations are always important!

Aida Horaniet Ibañez investigated the current situation of data visualization in DH (DOI: 10.5281/zenodo.7963466). She collected a corpus of more than 200 figures from research papers. After noticing that there was a predominance of statistical graphs, they extended the corpus with 500 extra figures from online magazines, social media, etc.

Basically the presented research differentiates between the following three types of visualizations:

Statistical graphics (reducing cognitive load by visualizing reliable data sources)
Data humanism (in-depth details about context where design and topic stimulates the user’s interest)
Humanistic interpretation (visualizations with user interaction)

Apparently, only a few tools support the latter two of the types. One interesting finding was that many visualizations from personal blogs are unavailable after a while. In my opinion, yet another good reason for heritage institutions to invest in web and social media harvesting (like we did at KBR with the BESOCIAL project, or like the International Internet Preservation Consortium (IIPC) does on a large scale).

Using humanities methodologies for digital resources

One fascinating thing about DH is that digital methods are not only used within DH, but humanities methodologies are also used to study digital resources. I think this provides new and interesting ways to look at the technology we are building and using!

There have been presentations about post colonial heritage, Wikidata’s data model and inclusivity, and best practices when annotating human characters in children’s literature according to ethnicity and gender.

Keynote about Cosmovision

This keynote from Patricia Murrieta-Flores focused on decolonial praxis in the Digital Humanities. She talked about languages in the Mesoamerican area. Apparently, there existed more than 68 languages before the “Conquest of America”! Yet most we know was written by Europeans.

My main takeaway is the following paraphrased quote:

“Be aware of the role of our work as DH researchers, there are also other perspectives in the world that are equally wonderful and which should be taken into account.” - Patricia Murrieta-Flores (paraphrased)

Most of the existing texts in her field are written from one perspective only: written by the colonizers and not the indigenous people. Partricia made the following comparison: Imagine all books about France and its culture would have been written by the British and in English. This would probably give a different view on things.

According to Patricia, this is also a crucial aspect for machine learning: there is a lack of training material for certain topics and definitely from different worldviews, which raises the question if AI is safe to use!

And if you wonder what Cosmovision is (I have to admit I did not know it): According to Law Insider,

“Cosmovision means the conception that indigenous peoples have, both collectively and individually, of the physical and spiritual world and the environment in which they conduct their lives.” - Law Insider

So in simple terms, it is about different worldviews.

The expressiveness of Wikidata

Ammandeep K Mahal and James Baker presented their work on histories of women active in archaeology, history and heritage in the BeyondNotability project. They curated and publish a dataset in a Wikibase instance.

Their presentation focuses on the free and open knowledge base Wikidata. Mainly they criticize the following three points (summarized in my understanding):

Wikidata statements do not accommodate for complex historical phenomena, such as (historical) women life’s combining paid work and care work
Ethnicity and citizenship are likewise not represented in a sufficient detail, even worse, bots often assume citizenship based on other properties and change the values
Sex and gender in Wikidata are sexOrGender, (a single property)

My opinion: Without considering any context about the meaning of the property itself, I have to say that the third point seems already questionable: a property with the word “or” in the name should, according to me, be split into several properties!

A key message of their talk was that Wikidata has a clean and simple interface which suggests simple and authoritative facts. However, the talk pages behind the interface may provide more controversial discussions that the average user might not be aware of.

I think: Data models can be made as complex as necessary. In the vision of the Semantic Web and Linked Open Data, anyone can create their own ontology to accommodate any world view in detail. The practice, however, has taught us that a few broad and simplified solutions often get the most attention. For data models, for example the generic schema.org vocabulary and for data, apparently, the Wikidata platform that does not even have a formal ontology.

This simplicity makes such solutions appealing but indeed amplifies certain world views and leaves out important context. I think that setting up a domain-specific Wikibase with its own vocabulary, as the authors did, is for now the best solution. A solution that resonates the most with the original Semantic Web vision of many decentralized ontologies (even though in this case it is not a formal ontology).

Uncertainty when labeling images

The talk of Paavo van der Eecken about the use of digital tools for uncertain humanities data was very interesting. In a nutshell, the talk covered considerations when creating labels used to annotate and categorize human characters on images in children’s books. I have learned that it is very important to use labels that are as specific as possible when dealing with ethnicity and gender to avoid biased generalizations.

But I also learned that, when labeling images, there is a difference between ontological doubt and epistemic doubt. The former refers to unclear boundaries and the latter to incomplete knowledge. Knowing these theoretical foundations helps, on a very practical level, to plan, schedule and execute projects with human annotators and to achieve the best possible outcome.

Using digital methods for humanities research

Many talks actually covered this kind of topics as it is a core part of DH research. However, since it was a multi track conference, I only attended a few talks that applied digital methods for humanities research.

I have attended talks that apply Linked Data for Greek mythology narratives or the Dutch East India Company, and persistent identifiers for library catalogs.

Linked Open Data for Greek mythology

Daria Kondakova and Jakob Kohler talked about their use of Linked Open Data to study methodological narratives.

They have built a tool with a user interface to create narratives in the form of RDF triples, with the aim to compare mythological narratives. Currently they are investigating how they can express relationships between the narratives. For example, Zeus in story A is not the same as Zeus in story B.

Thinking about it, one part of the solution could be to use the W3C PROV ontology and specify that the different Zeus instances are a specialization of a more broad Zeus concept (PROV agents can also be PROV entities I believe). Yet this would not solve the initial issue as there are still no relationships between the Zeus specializations.

Maybe named-graphs might be a solution to encapsulate the different narratives. In such a scenario there should be only one Zeus URI, but it would be used in different named-graphs. One would not have to explicitly specify the relations between different Zeus URIs, each is embedded in the context of the narrative named-graph.

Linked Data thesauri to study the Dutch East India Company

Brecht Nijman and Kay Pepping presented their work in building a SKOS taxonomy DOI: 10.5281/zenodo.7973694.

They want to unlock a key resource of the Dutch East India Company archive: the Overgekomen Brieven en Papieren which are roughly 5mio digitized documents. These handwritten documents went through a Handwritten Text Recognition (HTR) step.

However, due to text recognition errors or different spellings as well as synonyms, a simple text search for a concept is not sufficient. Their example was the herb Clove (Dutch: kruidnagel) that could be written as kruidnagel, kruytnagel, garioffelnagel, moernagel, etc.

Because existing thesauri are not detailed enough they created their own SKOS and SKOS-XL taxonomy with the tool PoolParty.

Persistent ISNI identifiers for research and legal deposit

My colleague Ann Van Camp and our volunteer Sergio Alonso Mislata presented the benefits of using a persistent ISNI identifier for contributors to creative works in the catalog of the Royal Library of Belgium.

This helps not only in the research projects BELTRANS and CAMille, but also to identify relevant works for the legal deposit in Belgium.

A side note: Sergio addressed us after a conference last year, saying that he would like to know more about Linked Open Data. We are always open to host volunteers or job students at KBR!

Geospatial research

One particular type of DH data is geospatial data, think of digitized old maps or location information in metadata. This again is something where DH shines, beautiful old maps are digitized and used.

There was a keynote about geo metadata and large language models, and talks around historical location data in Belgium and in Greece as well as location data in written works.

Closing keynote about Geographic Information Systems

The closing keynote from Piraye Hacıgüzeller provided thoughts on typological thinking and traditional Geographic Information System (GIS). It mainly focused on the complexity of encoding geographic information on archaeological sites.

They tried to encode this kind of information using CIDOC-CRM, but it was not expressive enough. Furthermore, some geographical information is difficult to express on maps, but more easily in text due too ambiguity and fuzziness such as for directions.

Piraye highlighted that with ArcheoBERTje, a sophisticated language model in Dutch language exists (based on the Dutch language BERT model BERTje). According to research, this language model could not further be improved with domain-specific thesauri as it already covered all the relevant information.

She ended the talk with the open question if there is still a need for metadata. Unfortunately there was not much time for the QA. But after the conference I checked again the presentation with the title The future of Knowledge Graphs in the World of Language Models by one of Wikidata’s founders Denny Vrandečić. And I think there are a few good points why there still is a need for curated metadata. As Denny phrased it:

“In a world of infinite content, knowledge becomes valuable.” - Denny Vrandečić

However, I get her point, especially after the explanation of the issue with fuzziness in geographic data in text. Thus who knows, for certain applications and use cases a language model might still be helpful and even preferable to metadata following a (too) strict schema.

Historical Belgian places

Léa Hermenault presented the Belgian Historical Gazetteer (DOI: 10.5281/zenodo.7997173. Their work will create a database of historical place names that will facilitate the mapping of place names found in historical documents.

They are using a PostgreSQL database with PostGIS extension to store the data. Furthermore, they are using the tool OpenRefine to enrich the data and create RDF. Eventually the data will reside in a Druid database hosted by Triply.

This is a great project which also facilitates (upcoming) work at KBR. In the recently approved project MetaBelgica, KBR and three other Federal Scientific Institutes will combine their person, organization, events/time and location information into a Wikibase instance. The presented historical gazetteer is a great reference dataset.

Vincent Ducatteeuw presented another project that focuses on historical Belgian geo data (DOI: 10.5281/zenodo.7945196), more precisely on historical street names in the city of Ghent. Street names change over time and there is an ontological issue: if a street name has changed, is it still the same street or a new one?

In his presentation, Vincent explained that they are using an identification methodology to distinguish streets by their proper name and their location or type of mereology (DOI: 10.3233/AO-200235). Furthermore, they are applying this methodology on the Wegwijzer der stad Gent, a city directory with scanned documents from 1875-1932.

Historical Greek places

Laura Soffiantini is pursuing the research question How Greece is described in Natural History? How, why, and in relation to what does Pliny refer to Greek places?

Therefore, she is using a multi-step workflow to extract place names from TOPOSText, enriches is with the tool flair and disambiguates places with the help of an authoritative list from Trismegistos and the tool TheFuzz (formerly known as FuzzyWuzzy).

Visualizing location data in books

Dan Costa Baciu presented a prototype to visualize location data of well-known institutions, infrastructures, etc that are found in books.

He argues that geospatial discovery such as in GoogleFlights or AirBnB does not yet exist for library catalogs. But also that it would be useful! However with their platform it is possible. They are using different color to distinguish between location data of the book metadata (such as publisher location) and location data extracted from the content of the book.

This is a very interesting and cool idea. For the BELTRANS project I am working on, which covers FR-NL / NL-FR translations, this would be a really interesting way to visualize data. Especially, since for certain comics, locations in the book are culturally translated as well. For example, Belgian cities of the Wallonian region in Belgium that got translated to Dutch cities for the translation for the Dutch-language market (translations that mainly target the larger Netherlands rather than Flanders).

However, this would require extracting the location information of the books. Hence this is a bit out of scope of the quantitative study in BELTRANS. But it would be an interesting addition to refine the BELTRANS corpus in the future (see also my notes on collection-centric thinking here in this blog post).

Poster session

There was of course also a poster session with many interesting posters.

One poster (DOI: 10.5281/zenodo.7976748) was about the research in the BELTRANS project for which I also work. It was presented by two of the project’s PhDs students: Clara Folie and Timothy Sirjacobs.

Another interesting poster (DOI: 10.5281/zenodo.7986819) was about the organization of a Hackathon at KU Leuven, presented by Leah Budke.

My ex BESOCIAL colleague Fien Messens presented a poster (DOI: 10.5281/zenodo.7956417) on web archiving of religious websites at KU Leuven.

Final remarks

This is it, I hope you enjoyed reading this rather lengthy blog post about DHBenelux 2023.

Thanks to my colleague Julie Birkholz and all the other organizers and volunteers for making this DHBenelux edition happen! The presentations gave me a lot of food-for-thought and I enjoyed catching up with the community. See you at the next year’s edition at KU Leuven.

Don’t forget to share this post on social media and in your networks!

Please also consider subscribing to my bi-weekly Newsletter FAIR Data Digest to receive more interesting content every other Tuesday!

My Phd explained

Sun, 21 Jun 2020 16:28:14 +0200

A side note: This post describes my PhD work from the perspective of June 2020. The actual finished PhD in 2022 was slightly different in comparison.

It is particularly challenging to explain PhD research to family and friends, especially in an abstract field such as Computer Science. In the following blog post I will try to explain my PhD, freed from academic precision and more from the big picture and with plenty of examples.

The bullet points form one coherent story, but each paragraph marked with an arrow can be clicked to read more details.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

The context

Did you ever wonder how an assistance system like Alexa or Siri can answer arbitrarily questions? Therefore such a system needs to search data of different disciplines also referred to as domains.

To assist that, it makes sense to combine these data in a natural way, you have to imagine it like a graph: nodes are things like concrete persons and connections between the nodes are the relationships between those things. As text such a graph could look like this:

Sven hasBirthday 25.04.1988 .
Sven worksFor IDLab .
IDLab headQuartersIn Belgium .
Belgium belongsTo Europe .

But first it has to be determined **what** can be connected **how** under which **restrictions** and this is called a **data model**. In this example my employer IDLab, which is an organization cannot have a date of birth and similarly Sven cannot have headquarter as he is a person. Generally speaking I'm researching how such data models can be created.

Data modeling is nothing new, but in contrast to the past where data was modeled for a single database or a single computer program we make use of the world wide web. Each concept and every relationship such as `person` and `headQuartersIn` but also concrete data itself like `Sven` or `Belgium` gets an own web address! Thereby they are globally unique identifiable and computer programs as well as users can look up the concept or the information! This could look like on the following page: https://sven-lieber.org/profile

My research - part 1: the data modeling process

Well, everyone can create such a data model in one afternoon, this is not a big deal. The challenge, however, is to make it professional so in the end a qualitatively high data model pops out. Ideally the data model should be constructed in a way that it solves the problems it was created for in the first place! For the question “How is the weather today in Ghent?” the data model for example at least needs to represent the concepts weather and town as well as temporal information.

Now a data model can be developed being easy and reusable, or complex and tailored to solve a very concrete problem.

An example of an easy and reusable data model: if you google any business, Google will show you on the right within an infobox what the opening hours or the founding year of that business is. Among others, Google can do this because the website owners used a standardized data model to mark information in their website. Not much precision is needed as the information is mainly shown to humans. In contrast to this, in *biomedical* science very complex data models following logical rules are created which are very precise such that also computer programs can "understand it".

Usually within the creation of a data model one has to talk to officials and experts, analyze existing data and have a clue about existing data models. After all the data model should follow standards and support seamless data exchange with other data models, called interoperability. **All that is also a complex process which can be split into steps.**

By the way, this is the same for professional software engineering. A measurable process subject to optimization distinguishes software engineering from the simple act of *programming*.

The first part of my research revolves around this process. I’m pursuing the question what is the best way to represent restrictions in a data model, so how one can create a data model which is not too simple but also not too complex for a given problem. Restrictions could be on the one hand logical axioms like all life forms have a birth date, or on the other hand testable constraints important for data exchange and data quality such as the birth date has to be smaller than the death date.

Thereby I do not reinvent the wheel, I carefully read the literature of knowledge modeling and I still do! The goal is to reuse existing methods as much as possible but to adapt them to the modeling of constraints. Part of this is also to analyze which measurable pros and cons arise if certain restrictions are either expressed as logical axioms in the data model or as constraints. One use case in my PhD are privacy regulations. To check compliance with privacy regulations on the one hand the data model needs to provide relevant information and on the other hand constraints are possibly needed.

My research - part 2: user-friendly visualizations

In my lab we mostly work in different projects together with industry, the government or institutions. The experience of these projects revealed concrete problems regarding user-friendliness of modeling languages: A 2017 newly introduced language to represent constraints in graph data models is text-based, user have to first learn it which might be tricky!

Pictures are worth a thousand words. Humans possess an astonishing information processing system though, .. no, I do not talk about the smartphone in the pocket but about the brain! From cognitive science we know that text is processed slowly as we have to concentrate more. In contrast to that shapes and colors are almost automatically processes with no visible effort. We work with knowledge graphs which are tangible, we take advantage of that!

We develop graphical languages to describe constraints on our knowledge graphs. Based on the assumption that such languages are more intuitive compared to an alien text syntax we examine in user studies which graphical languages are suited best.

And why all that?

Data without description is useless. Imagine you find a piece of paper full with numbers. Without context you have no clue what these numbers mean. Maybe somewhere a computer program exists which knows how to read and interpret this data. However, for you or other programs this piece of paper is useless!

Adding descriptions provides context and makes the data self-describing and hence interpretable. If you note on the piece of paper that the first column is weight measured in kilogram and that the second column represents birth dates and that both are characteristics of a person whose name is noted in the third column you performed data modeling! Additionally the data are now self-described and independent from specific computer programs. By the way: This piece of paper could be representative for a file or a website.

If data are properly described they can be combined in a meaningful way and such combined data can be searched quicker, easier and most importantly in a uniform way which is then utilized by systems such as Alexa or Siri! A not-meaningful combination is for example the addition of a temperature in degree celsius on my age in years.

Summarized

Based on what was read we can summarize: I deal with data modeling. I pursue the question how we can build knowledge graphs which also comply with requirements towards data quality and privacy. In particular I investigate on the one hand methods of knowledge engineering, so which steps need to be performed to create such knowledge graphs, and on the other hand the user-friendly visualization of data constraints such that also non-experts can create such constraints.

So that’s it! I hope I cleared some things out :-)

Below is a picture of the ISWC 2019 conference where I presented my PhD plan both as presentation and as a poster (my trip report of this conference held in New Zealand is here).

K-Cap 2019

Sun, 24 Nov 2019 21:09:40 +0100

At the end of November 2019 I could visit the sunny California to attend the Conference on Knowledge Capture (K-Cap). A lot of interesting workshops, tutorials and talks covered topics such as the representation of knowledge in Wikidata, approaches to transform tabular data to Linked Data, or the representation of scientific literature and processes as Knowledge Graphs. Additionally I could present our work on MontoloStats. Continue reading for the full trip report!

The International Conference on Knowledge Capture (K-Cap) takes place every two years alternating with the EKAW conference. This year it took place in Marina Del Rey, CA in the United States. I traveled directly from New Zealand were I attended the ISWC conference a few weeks earlier.

What I liked the most K-Cap is a single track conference and thus you can’t miss any talk because of another one.

What I didn’t like It is interesting that K-Cap is not only about Semantic Web stuff but knowledge capturing from a general perspective also including NLP and machine learning. However, for me some of these talks were hard to grasp and were not directly related to my research.

Talks regarding “Concept embeddings” seemed to be very present, but since this topic is not too relevant for me I’ll skip those in this report 😃. Instead in this post I will mainly focus on talks from the following topics which also seemed to be very prominent:

Wikidata
Languages and approaches to map tabular data to Linked Data
Knowledge Graphs of scientific practices or literature

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

SciKnow workshop

On the first day I attended the Third International Workshop on Capturing Scientific Knowledge (Sciknow 2019). Yolanda Gil gave a very interesting keynote about the capturing of hypotheses and scenarios in scientific research. “Hypotheses are not finished after a paper is accepted” she said, the confidence value of the hypothesis changes over time. New and/or more test subjects might be available later. Additionally there are also other methods available to evaluate a hypothesis even though some labs might use the same or similar methods just because they were always used. I think we should definitely keep track of our hypotheses, methods and results in a systematic fashion: We are scientists in the end, future research from different angles might bring new insights on the current state of knowledge in our community!

Allard Oelen talked about the annotation of scientific literature using crowdsourcing to easier compare scientific contributions. It is difficult to find crowd workers qualified for such a task, additionally it is very time consuming as a scientific paper first has to be understood. Thus Allard and his colleagues aim for a system in which the authors annotate their papers while submitting. Ideally this should take less than 10 minutes of the authors’ time. It is a very intersting approach, but authors might submit only the bare minimum information when provided with a template during paper submission. I think a gamification approach could here be of help, i.e. showing additionall information based on the Knowledge Graph’s statistics, such as “x% of authors also filled in this form field”.

Oscar Corcho presented ongoing work from Ana Iglesias Molina about using spreadsheets to declaratively express Linked Data mappings which then can be translated to the most suitable language. Starting point of the presented work is the Bio2RDF dataset which contains biomedical knowledge; how is it generated?, is it still up to date?. Based on experience in working with different domain experts, spreadsheets have proven to be a good way to represent information. It is an interesting approach and distantly reminds me a bit of the tabOTTR representation of the ontology template language OTTR.

In classical Linked Open Data, Linked Data is generated and then linked. Daniel Garijo presented WDPlus which follows a link-first approach, different domain-specific satellites are created around the Wikidata core.

Some of these satellites might be based on tabular data. Different languages such as R2RDF and RML exist to transform structured data to RDF. However, especially in spreadsheets domain experts use a variety of layouts to represent the data which poses challenges. Therefore WDPlus has the T2WML framework, using YAML to express the transformation from tabular data to Wikidata. These mappings makes use of YAML which apparently is human-friendlier, something which I’ve seen in other approaches too. I like the idea of the “link-first” approach, it reminds me of Ontology Engineering where existing classes and properties should be reused as much as possible. In fact, also the challenges described by Daniel are similar, i.e. the identification of new properties which is currently performed manually by knowledge engineers, namespace issues and inter-satellite links.

Wikidata tutorial

In the afternoon of the first day I attended the Linking, Extending, Exploiting and Enhancing Tabular Data with Wikidata tutorial given by Daniel Garijo and Pedro Szekely. I finally learned the basics of Wikidata and its data model which differs from e.g. the one from DBpedia.

Wikidata is a collection of claims whereas each claim has related qualifiers and references.

Since Wikidata is using a role model it can represent information more accurrate as e.g. DBpedia. In DBpedia Pluto is still a planet, whereas in Wikidata it is stated that it was considered a planet in a certain period but now not anymore.

The following picture shows some stats about the concept hierarchy in Wikidata. It also shows the very nice output of a commandline tool showing the hierarchy with attached stats.

However, the mapping of the Wikidata data model to RDF looks complex in the beginning, i.e. each Wikidata statement is expressed using multiple triples as reification has to be used. Please note also the different namespaces.

According to GoogleTrends and statistics published by Wikidata it gained a lot of attention in the recent years. The role-based model with qualifiers and references looks very interesting, also the fact that Wikidata is a huge crowdsourcing project. I wonder how easy or hard it would be to adapt or develop certain Ontology Engineering activities suitable for Wikidata.

First conference day

Peter Clark from the Allen Institute of AI gave the keynote of the first conference day. He introduced the project Aristo which is about reasoning and explanation.

“Tables are a way to express systematic knowledge, e.g. in science” he said. Peter talked about so called table knowledge systems which are also used in the backend of the Aristo system. Knowledge is then extracted from these tables to execute table reasoning. It is interesting to see that there are more approaches than only node-link-based Knowledge Graphs. This table-based solution performs quite good in different challenges. The Aristo question answering system outperforms an average 8th grader.

Later that day I had the chance to present our work on MontoloStats which apparently sparked some discussions among the reviewers and thus was marked a spotlight paper because it was controversial.

Looking back at the first K-Cap 2001 I initially asked the question how we model knowledge nowadays were we have hundreds of ontologies and dozens of W3C recommendations available at our finger tips. I presented our approach to create a statistical dataset about the use of different OWL axioms in ontologies extracted from LOV and BioPortal. Concretely we are looking into how such statistical metadata can improve the process of ontology reuse, but beyond that, the dataset provides empirical data providing several insights and raises new questions to research. I am lucky about the feedback I got. Specific tools like Protégé or visualizations like VOWL might focus too much on certain restriction types and thus seduce users to use these more often than others. Although marked as controversial, the audience seemed to appreciate the fact that there are now openly available empirical data about the use of axioms in ontologies. For me such axioms are just one way to express restrictions, limiting conditions imposed by the real world or use case. Let’s see where my research towards better modeling guidelines to express restrictions will lead me 😃.

Town Hall meeting

In the town hall meeting it was announced that Oscar Corcho becomes the new chair. Furthermore the location and duration of the conference, as well as the review process and list of topics were discussed.

The audience seemed to prefer larger cities over isolated areas as the traveling becomes easier. Although it was pointed out that the sessions were quite packed with presentations, the 2-day duration is preferred over 2.5 days. Also given the fact that the last half day of a conference some people leave earlier anyway.

Whereas long papers were accompanied by a presentation, short papers had a poster. It was pointed out that a short paper is not shorter in research and that the posters should also be visible for the duration of the entire conference.

Since the review phase was in August (vacation period) it was difficult, but generally everybody was happy with the reviews. Something like a best reviewer award was also proposed to give further incentive for good reviews.

The shepherding process for certain papers was also appreciated although it gave an extra burden to reviewers and meta reviewers as an answer letter and resubmitted paper needed to be checked again.

For the future the list of topics might also undergo a small change.

The venue for the next K-Cap will be … announced.

Second conference day

Due to a project meeting I unfortunately missed the keynote and most of the talks of the morning. Even a single-track conference can’t help much in such a case 😃.

In the afternoon Cristina-Iulia Bucur talked about the review process in science which didn’t really change much the past few hundred years. Currently she and her colleagues are looking into ways to improve the reviewing process by using Semantic Web technologies. Reviews and concerned text passages are expressed as Linked Data, additionally reviews can be further annotated to e.g. say if it concerns grammar or content. A user study was performed to evaluate the proposed fine grained review data model. Looking at how the review process works, I wonder why nobody ever changed it. I think even in a pre-semantic-web era other tools could have been used to improve the review process. In the here proposed solution I especially like the UI in which the review is directly attached to the text, so researchers can review while reading.

Binh Vu talked about the language D-REPR to map tabular data with different layout to RDF. Based on openly available tabular data extracted from data.gov it was verified that the language meets stated requirements. Additionally the performance was evaluated by comparing the execution of the Rust-based implementation of D-REPR against two other solutions on data in the nested relational model layout. There are many languages and tools to transform tabular or any structured data to RDF. A discussion after the talk revealed that instead of creating new languages and tools the existing recommendations should be updated, taking current limitations into account.

I’m currently looking into visualizations for RDF data, thus I found the talk of Vitalis Wiens also very interesting. He presented GizMO, a methodology and tool to visualize ontologies. It is based on an ontology which addresses the graphical appearance of OWL constructs in one layer, and visual properties for conceptual elements in another layer. Great work! I am not aware of any ontology to describe the visual appearance of graph based models. Additionally nested visualizations like UML are also mentioned in the paper and can be generated in the demo. I’m pretty sure this work will be relevant for a lot of visualizations out there and can also serve as a common format to interchange visual information.

Closing event

Within the closing event, Ling Cai was awarded the best paper award for the paper TransGCN:Coupling Transformation Assumptions with Graph Convolutional Networks for Link Prediction. Congratulations!

This was K-Cap 2019! I enjoyed the many conversations I had. Additionally I also had my first main track presentation. I will definitely have a look into Wikidata and will also follow up on some of the presented papers. I’m looking forward to submit something to next year’s EKAW and then of course to K-Cap again in two years!

ISWC 2019

Tue, 05 Nov 2019 08:41:37 +0100

In the last week of October 2019 I attended the International Semantic Web Conference (ISWC) in Auckland, New Zealand. I had the honor to present my PhD plan at the ISWC Doctoral Consortium. In this blog post I talk a bit about the conference and focus on a few presentations I found particular interesting or relevant for my PhD regarding guidelines for the design and use of ontologies.

First things first: I will provide a bit of context and say what I liked/disliked.

The International Semantic Web Conference (ISWC) is THE conference of research regarding the Semantic Web also covering Linked Data, Knowledge Graphs and AI. This year it was held in Auckland New Zealand which on the one hand is an amazing venue, but on the other hand resulted in super long flights and visa issues for a lot of attendees.

What I liked the most The conference also had several sessions regarding ontologies and ontology design which is relevant for my PhD.

What I didn’t like I missed a few really good workshops because I had to attend the Doctoral Consortium a full day. Fully attending it is a good thing, just a pity it overlaps with other workshops.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

OTTR tutorial

For the first workshop day I’ve chosen to attend the tutorial on scalable construction of sustainable knowledge bases which was all about the Reasonaable Ontology Templates (OTTR) framework. It was given by Martin G. Skjæveland, Leif Harald Karlsen and Daniel Lupp from the University of Oslo, and Melinda Hodkiewicz from the University of Western Australia.

In a nutshell OTTR is a language and a framework to describe knowledge on a conceptual level, whereas it abstracts from the actual knowledge representation in e.g. OWL. Thus ontology engineers can communicate with domain experts in their language and talk about for example a Person without needing to know which concrete axioms are used to represent a person in OWL. Templates contain parameters representing the variable part of a template which then gets extended to RDF together with its static part; working with templates feels like working with a constructor in languages such as Java.

I was particular interested in this tutorial as I see OTTR as the missing building block to generate different ontology representations based on the same conceptual model which is relevant for my PhD. The tutorial introduced the basics, we could get our hands dirty on some examples and then two real-life use cases were introduced.

Melinda Hodkiewicz stressed the need for ontologies in the Engineering sector, which she did as well in each session I have seen her during the main conference 😃. Which is nice, we need more enthusiastic people such as her to promote our work and highlight its concrete added value for businesses or the end user.

Besides the content I really liked the usage of https://menti.com to interactively get feedback from the audience such as demographics or incentives to join the tutorial. And the end of the tutorial the organizers went back to the initially asked questions and incentives of the audience and it was discussed whether each point was addressed. Thumbs up, please more of that in the future!

Doctoral Consortium

A lot of workshops took also place at the second day but I didn’t attend as I has the great pleasure of participating in the ISWC Doctoral Consortium all day where I could present the plan for my PhD.

The doctoral consortium started with a great keynote given by Valentina Tamma. She started by defining what a PhD means according to its definition, i.e. doctor of philosophiae in Greek “love of knowledge”, and continued with further details regarding the scientific process. She recommended the book Science in Action from Bruno Latour (1987) and stressed that we should read a lot and talk about our research as much as possible, first to the broader community and then in lay terms also to the general public.

Each presenting PhD student got upfront a mentor assigned who gave feedback to the presentation, but of course everyone in the audience was invited to ask questions and provide feedback. The presentations varied in quality and depth, but I think no one was able to present in the given time of 12 minutes (me included). It actually is a challenging task to present your whole PhD research within 12 minutes, especially if you are in an early-stage and it is just a plan.

The presentations were supposed to be accompanied by a poster presented in poster sessions in the coffee breaks. However, due to the fact that no one finished in time with the presentation, the poster session was moved to the general poster session of the main conference on Monday. Thanks to the chairs and local organizers to find some spare walls and arrange that we could present next to the main poster session.

A panel session rounded up the doctoral consortium. Several senior researchers discussed questions regarding the role of the PhD in industry, I will just briefly mention the questions and the main points:

Which is your opinion with respect to this migration from academia to industry?”
- Higher salaries and more stability in industry
- Nowadays more opportunities in industry and a more research-influenced style of work
- On the negative side: whole research areas get confiscated by industry, e.g. information retrieval and NLP
From your experience, which is the actual need of industries in terms of innovation? Are they really believing in it or it is just a trend that they want to follow?
- The industry wants simple things which work
- Companies need to stay ahead of competition
May this new age of AI affect the programs of your PhD students?
- Putting PhDs on irrelevant topics makes also everyone unhappy
- We also benefit from it, e.g. knowledge embeddings, but one has to be careful because building a system by themselves is not research, how does it help our community?
Do you think that a strong collaboration with companies would have a detrimental effect on the quality of research?
- A lot of tools we use nowadays are built by industries, we wouldn’t have build them for ourselves
- Industry brings a lot of insights
- Of course also the risk in credibility if too much research is directly funded by industry

Monday

At the first official conference day I participated in the opening session, the keynote by Dougal Watt, a session with industry talks and two sessions regarding data quality.

Opening session

Within the opening session it was stated that from 194 valid research track submissions 42 papers were selected and thus the acceptance rate of the research track was 21.6%. According to the presented statistics of submissions per topic in the research track (topics with more than 10 submissions) almost 80 submissions concerned database, information retrieval, information extraction and NLP. Whereas Ontology Engineering and ontology design patterns for the web are the topics with the least submissions in this more-than-10 category. It was a pity to see that the privacy topic was ranked in the statistics of topics with less than 5 submissions.

Congratulations to Olaf Hartig, Christian Bizer, Johann-Christoph Freytag to win the SWSA Ten-year award with their work “Executing SPARQL Queries over the Web of Linked Data”. Also congratulations to Julius Volz, Christian Bizer, Martin Gaedke and Georgi Kobilarov to also win the SWSA Ten-year award with their work “Discovering and Maintaining Links on the Web of Data”.

Keynote by Dougal Watt

Monday’s keynote was given by Dougal Watt and was titled “Semantics: the business technology disruptor of the future”. He gave a small tour through the different eras of information processing: from pre-literate and literate via the printed word to modern computing, the web and context nowadays. Furthermore he talked about application centric thinking and about the different problems of relational databases, enterprise resource planning, enterprise data modelling, SOA and data warehousing. Summarized these problems mainly focus on limited semantics, massively complex models, interoperability and missing schema agreement leading to a 28 minute discussion about definitions within a 30 minutes meeting. Of course semantic technologies and ontologies offer a possible solution for these problems. But according to my opinion some of these problems still remain and are just shifted to how ontologies are modeled and how they are used, i.e. discussions of definitions and complex models.

Within the QA the topic data shapes and SHACL were brought up. Dougal said that SHACL shapes are fantastic things, as they can hide the complexity of ontologies which can then be used by businesses. Personally I find this a very interesting point which is also relevant to my PhD: Shapes become important components when working with ontologies and thus need also to be built systematically and/or considered when designing ontologies.

Presentations

Evgeny Kharlamov presented the work of colleagues regarding a knowledge-base for material science within the industry track. Working in material science involves the creation of new materials for which knowledge of thousands of existing papers and patents need to be searched. Queries such as which material X reacts with material Y and has its melting point below Z need to be asked. An ontology was designed and a knowledge graph was built to facilitate the task of this search scenario. Furthermore they are working on an end-to-end system including not only the ontology but also search interfaces for end-users. I was interested in how the ontology was designed to deal with that task: apparently a lot of subclass axioms were used. I like their use case in which they have to provide a user-friendly system powered by an ontology. As Dougal pointed out in his keynote: enterprise data modeling tend to create massively complex models. I would say that this enterprise modeling mindset still exists in the mind of some ontology engineers, a use case such as presented here shows that ontologies also have to be used by systems which also comes with extra requirements regarding their use.

Within one of the data quality sessions, I listened to Albert Meroño-Peñuela who presented the work of him and his colleagues about RDF-lists. MIDI applications (Musical Instrument Digital Interface) express songs as long lists of musical chords. In the context of this, the authors surveyed common list modeling practices taking e.g. different ontology design patterns for lists into account. The authors reported coherent results among triple stores when evaluating the triple stores’ performance when processing the different ways in which lists are currently represented. The rdf:List is the only expression which produces “closed” lists, but it is also the one with the poorest performance. It is interesting to see that certain design decisions have an impact on performance. It actually shows that guidelines of how to implement an ontology for a certain use case are needed to provide an optimal solution, i.e. in a use case in which performance is important the usage of rdf:List might not be optimal, whereas in a use case demanding more strict semantics the closeness of lists might be required.

Tuesday

For Tuesday I will discuss the keynote, panel session, PhD lunch, town hall meeting, gala dinner but also actual research sessions :-)

Keynote by Jérôme Euzenat

On Tuesday, Jérôme Euzenat presented the keynote For Knowledge. He quotes that data without knowledge is like a map without a legend and stresses that explicit knowledge is important. Also, a great idea remains a great idea. Sometimes the time for an invention has not yet come and reinventing the wheel sometimes makes sense respectively invent different variants of a wheel.

Another interesting idea he proposes concerns eScience. Nowadays scientific experiments can be described semantically: why not submitting a detailed experiment description to conferences/journals to receive reviews purely on the scientific process and not the outcome?!

Panel

After the keynote a panel session took place. It was titled “How much semantics goes a long way”, emphasizing on the how much with respect to the famous “a little semantics goes a long way” from James Hendler. Valentina Tamma, Jérôme Euzenat and James Hendler were the panel and the room was packed with senior SemWeb researchers. The audience lively participated in this panel discussion. I will briefly list some of the main comments.

The past 20 years demonstrated that different ontologies in the same domain can exist, people choose what they want and one ontology possibly becomes the lingua franca of that domain
There are obstacles; if technologies and languages are used which are too complex for people to use they won’t use it. That’s why expressivity might be reduced
A way to achieve consensus is maybe a more expressive language to express nuance and isolate which statement you agree with and to which not. So you might end up to agree
The problem is not about expressivity per se, but about the incentives to put the expressivity. E.g. in the research of genes, researchers put the expressivity in. We should ask, how to incentive the people to put the expressivity
How much semantics is perceived wrongly, choose either to define less or richer expressivity, but it depends! For certain tasks we don’t need much expressivity. Focus should be to make these different types to coexist with the same data

PhD lunch

On Tuesday also the PhD lunch took place. This program point belongs to the Doctoral Consortium; PhD students and their assigned mentors should have a more informal environment to talk. I liked the setup, I was on a table with a few other PhD students and a few mentors. We could ask questions and a lively discussion was sparked. There were certainly very helpful tips, however given the time constraint we couldn’t reach a certain depth in the discussion so everything stayed more high-level.

Town hall meeting

The ISWC conference provides the town hall meeting as general session in which all kind of feedback can be provided.

One of the first points brought up was the review process. It was the first time that the research track was double-blind. Some people liked it and some not; there are certainly pros and cons but it was mentioned that better guidelines in how to review should be provided.

Another point of discussion was the location of the conference and thus the visa problematic a lot of attendees faced. The visa regulations of New Zealand were just recently changed so it couldn’t be taken into account back then when the conference venue was chosen. Unfortunately the university also had zero power in influencing the government’s visa decisions. Comments from the audience were among others that we should be more inclusive as a community and choose venues more wisely.

Session on ontology design

Two papers in this session gathered my attention, a paper regarding the detection of ontology design patterns in biomedical ontologies, and a spotlight paper about the use of OWL and WebProtégé at Pinterest.

Christian Kindermann presented how the influence of ontology design patterns in the biomedical domain can be measured. The presented results indicate that there does not seem to be much use of ontology design patterns in the analyzed ontologies. However, regardless of the results this is very interesting research as it provides a baseline which allows further investigation. One introduced way to detect ontology design patterns is by checking first if a specific axiom used by a pattern is present in an ontology. This indirect way to conclude if an ontology design patterns is present in an ontology can be facilitated by our MontoloStats dataset/webservice on which we recently published a paper.

Rafael S Gonçalves presented how an OWL ontology and the use of WebProtégé drastically shortened the development cycle of the Pinterest taxonomy. The Pinterest taxonomy contained around 6,000 “interests” users can have which were arranged in 3 layers. This taxonomy was increased to 11,000 interests represented as OWL classes organized in 12 layers. It took less than 2 months to build the end-to-end system the Pinterest employees without any ontology experience can use. Furthermore it took less than 1 month to build the final ontology. I would say that OWL is often perceived as very complex, but this presentation shows that with the use of appropriate tooling such as WebProtégé, which offers collaborative features and complete provenance, even non Semantic Web experts can handle an ontology! A question asked by the audience further gathered my attention: why did you use OWL and not just SKOS? Apparently there are future needs which can be addressed by OWL, especially the implementation of existential restrictions. I find this particularly interesting as it shows that there exist specific needs for certain restriction types which is relevant for my PhD and the definition of guidelines in how to model restrictions.

Session on domain ontologies

Within this session I liked the talk from Karl Hammar about a Real Estate ontology and the talk about the W3C recommended SSN ontology.

Real estate data are not just about prices, there are also needs regarding energy efficiency, maintenance and sustainability. Karl Hammer presented a real estate ontology which was built following an adapted version of the eXtreme Design methodology. The ontology consists of a core and different modules and it deliberately renounces the reuse of foundational ontologies. I like that he emphasized that an ontology is useful but not enough! There must be also a way to interact with it, i.e. the ontology has to be used. Short after the conference, Karl released a first version of a tool to create a OpenAPI specification from an OWL ontology. Cool stuff 👍

The SSN ontology is a W3C recommendation to model sensor data. Kerry Taylor presented the work around the SSN ontology. Interestingly, not only the SSN ontology was built but several best practices were published under the umbrella of W3C. I have read a journal paper related to this talk, it discusses design decisions to make the initially defined data model easier. This was achieved by modularizing the ontology into a “lightweight” core and outsource alignments to other ontologies into specific modules. The core ontology mainly defines concepts and no axioms, even to determine domain and range schema.org’s domainIncludes and rangeIncludes are used. In my opinion the presented solution solves a lot of problems and demonstrates how a trade-off between strict semantics regarding interoperability to other standards and lightweight description of concepts can be achieved!

Gala dinner

Besides all the research the conference of course also provided a lot of room for social gathering :-) On Tuesday’s agenda was the gala dinner. A Māori ensemble performed several songs and traditional dances to welcome the ISWC participants from all over the world to New Zealand! The Māori are the indigenous people of New Zealand which inhabited the islands roughly 700 years ago. This is also interesting as it means that New Zealand was the last place on earth on which the Homo Sapiens settled down, just a few hundred years ago! After that we enjoyed our food to which also the ice-cream kiwi from the blog posts header image belongs.

Wednesday

The last day was shorter, my highlights are the keynote, a presentation about ontology ranking within a search scenario and the closing ceremony.

Melanie Johnston-Hollitt gave the keynote Extracting Knowledge from the Data Deluge to Reveal the Mysteries of the Universe. This keynote was very interesting as it revealed a lot of really cool scientific applications which could profit from explicit knowledge. I can only recommend to listen to the keynote which is available on YouTube.

Short before Lunch I listened to some talks regarding ontologies for recipes and food. This seems very interesting and I’m happy some efforts regarding this direction already exist.

In the last talk I’ve heard at ISWC 2019, Niklas Kolbe talked about the ranking of ontologies regarding relevancy when searching for concepts. He pointed out that people have a certain trust into the first few search results and hence a ranking plays a crucial role. In the already mentioned recently published paper we argue that for an ontology reuse scenario the search for ontologies containing certain axioms might be relevant. It was interesting to hear that something like our solution could be integrated into the presented solution based on qualitative features.

In the closing event best papers in several categories were awarded, I will not list all the winners, just a general “congratulations”! 😃

The next ISWC conference will take place in Athens, Greece. In 2 years ISWC will take place in Albany, USA. The audience seemed not that enthusiastic when this venue was announced, .. maybe because of the already mentioned issues with visas for certain countries such as the US.

That was the ISWC 2019 conference from my point of view, I hope you enjoyed reading it! I got a lot of feedback for my PhD and also saw the relevancy for it.

I was writing this blog post while traveling around in New Zealand, I will now enjoy the remaining time in New Zealand, see you at the next conference!

ISWS 2018

Mon, 09 Jul 2018 14:27:27 +0200

In the first week of July I participated the first edition of the International Semantic Web Research Summer School (ISWS) in Bertinoro, Italy. It was an exciting event, where I had a lot of fun and learned also a lot! Continue reading for the full travel report.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

About the summer school

This first edition of the International Semantic Web Research Summer School (ISWS) was held in July 1 until July 7 in the University Residency Center in Bertinoro, Italy. The Residency Center is basically an old castle used for conferences and training courses and is a very nice venue. The center also provides hotel-ish bedrooms with bathrooms. 113 people applied for the summer school, from which 60 students were selected for participation! 73% were PhD students, 17% master students and 10% postdocs and researchers from mostly France, Italy and Germany, but also a lot of other Countries. Among other, I met people from a university in Chile (which are originally from Cuba), the United States, the UK and Denmark. The gender balance could be better, but we at least reached a 63% to 37% ratio.

The summer school aimed to teach basics of the semantic web in the form of tutorials, keynotes and so called in-depth-pills. Furthermore there was a bigger group work, in which each team worked on a research topic with a paper and presentation as outcome. Overall the summer school aimed to model and represent a researcher’s life in one week, including team work, socializing and deadlines!

Overview

I’m gonna give now a short description of each program point with further details later on. After an introduction of the school and the tutors, two poster session were held, where each participant presented his/her research. In the course of the week different presentations (keynotes, tutorials, in-depth-pills) regarding the Semantic Web were given. There was also a bit of room in the schedule to work in so called Research Task Forces, small teams led by an tutor, to investigate in one perspective on the theme “Linked Open Data validity”. Besides socializing in the form of team work and coffee-break conversations, there was also a gala dinner with small party, an excursion to Rimini and things like slide karaoke 😉

For the lazy reader I can already conclude, that it was a lot of fun and that I learned a lot! You can make yourself a picture if you have a look at the Twitter hashtag #isws2018 to see what happened during the week and see some conclusions from the participants.

Poster Session

Well, it was a poster session, not much to say about that. I saw a lot of interesting work. Integrating uncertainty in the Semantic Web Stack, a Web-based heritage platform built on Linked Data, Using machine-learning for Ontology matching and much, much more. An award for best posters went to Maximilian Zocholl, Piere-Henri Paris and Valentina Carriero, Congratulations! I was lucky to receive an honorably mention for creativity for my Poster, thank you!

Talks

Keynotes were given by Enrico Motta about Data Analytics, Marta Sabou about Rigour and Relevance with Design Science, and Sebastian Rudolph about a Logician’s view of the Semantic Web. The keynotes covered more high level concepts and lessons learned. They were very inspiring and gave a good perspective on a researcher’s life. Additionally three tutorials were given, which focused more on concrete techniques. In the tutorial given by Maria-Esther Vidal and Sebastian Rudolph, we learned about the basics of Reasoning and SPARQL query execution. Claudia d’Amato and Michael Cochez talked about Machine Learning. John Domingue talked about Blockchain and decentralization (Ruben Verborgh was supposed to give the talk as well, but unfortunately he couldn’t make it to the summer school).

As a third kind-of talk, the summer school offered so called in-depth-pills. Aldo Gangemi talked about ontology design patterns, Marieke van Erp about Natural Language Processing and Marta Sabou about crowdsourcing. Overall, each talk gave a lot of insights and the numerous questions from the audience were answered.

Research Task Force work

This was probably the most interesting part of the summer school. Based on a bigger subject, which was Linked Open Data Validity, each tutor presented research questions covering his/her point of view on the subject. Participants could indicate preferences regarding tutors and then Research Task Forces (RTF’s) were built. The assignment of participants to tutors took place after dinner at the evening of the second day. Each RTF was assigned a name, following a scientifically sound methodology, a speaking hat!

Thus we ended up with the Teams

Hufflepuff (John Domingue), validity in a decentralized disintermediated world
Ravenclaw (Harald Sack), how does validity relate to context? (my task force)
Gryffindor (Aldo Gangemi), validity patterns and how to establish metrics
Mordor (Claudia d’Amato), can validity be formalized? learning validity
The 42’s (Marieke Van Erp), validity for incomplete information or fluid definitions
The Deloreans (Valentina Presutti), validity and common-sense
the Jedis (Maria-Esther Vidal), validity in SPARQL federated querying
Dragons (Sebastian Rudolph), logical LOD validity
Hobbits (Michael Cochez), detect anomalies with e.g. Machine Learning

Each RTF had to prepare an 8-page research report, a 10 minutes presentation and a one minute funny video. The best presentation was awarded to the Hufflepuffs and the Deloreans! Again the Hufflepuffs were awarded with the best research report, this time together with the Jedis.

An important part of the summer school was fun! To achieve that we had multiple events like slide karaoke (40 seconds time to present a random power point slide), dinner in Bertinoro, an excursion to Rimini with subsequent dinner at the beach and a Gala dinner with DJ and dancefloor. Usually one could also find multiple groups and/or tutors in a close-by gelateria or bar in the evening.

My Conclusion

Go there! I haven’t been to other summer schools, but I’ve heard that they can be boring. But this one is an exception! The Semantic Web Community is quite small, here you will meet your peers, people you will work together with, see on conferences and so on. As stated above, I really enjoyed it and learned a lot. But notice, if you go to Italy, bring a power adapter.

ISWS 2018 Poster

Mon, 02 Jul 2018 10:20:27 +0200

For the ISWS summer school I prepared an interactive poster. Read in his post what it is about and see how it was made.

My summer this year will start with the attendance of the International Semantic Web Research Summer School (ISWS) in Bertinoro, Italy. For that we were invited to bring a poster, describing our PhD research. My current research tasks involve so called shapes, describing data within a graph of information. Since graphs and shapes are very tangible concepts, and one of the main goals of the poster is to gather feedback, we had the idea of making the poster a bit interactive. My supervisor suggested post-its, but as I have a lot of creative energy (and probably procrastinate writing papers), I decided to go for a small Knowledge Jigsaw Puzzle.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

What is the whole context?

Within the Semantic Web, information are modeled using vocabularies of terms, relationships among these terms and also constraints, things which aren’t allowed. The combination of the mentioned vocabularies, relationships and constraints is usually called an ontology and can be expressed in a formal language. That formalization allows computer programs then to “understand” the data, but also to infer new knowledge. For example, if I model persons and say that there is a relationship called “knows” which means that both people know each other, and I have the Information that Sven knows you, then the information that you know Sven can be inferred.

What has this now to do with shapes?

Well, constraints modeled in an ontology are used to infer new knowledge, they are not meant to use for validating data, e.g. that a person must have at least one last name. There are various reasons for that, which are out of scope for this blog post. Important is now, that recently the W3C committee released a standard to express constraints as shapes on data used i.a. for validation. It’s all quite new, so there are no clear methodologies yet, when to use shapes of data, when to use constraints in the ontology or when to use a formal rule language (yes these exist as well).

Okay okay, enough background, what’s the interactive poster now?

I am currently working on a project, which is about improving the customer journey when using public online services. An example: You want to move, so you have to do certain tasks at the city you are currently living in, like e.g. sign-off and provide your new address. At the city you are moving to, you have to provide certain information as well, e.g. if you’d like to request a parking spot for your car. These are all steps you should do, and they can be understood as a workflow. The aim of the project is to build customized user interfaces across cities. So that it seems you are only interacting with one system, whereas in the background multiple forms of different websites are filled in.

The customization here is the key point! Beside simple optimization like asking you only once for your old address (or even reuse it if the government has that information already), we can think of more personalization.

That’s were shapes come in. Given one or multiple workflows, containing different steps, your preferences can be modeled as shape(s) and be used to select the optimal steps for you. Similarly, context information like the device you are using or a certain life situation you are currently in, are somewhat constrains which can be used for optimal steps selection.

What I did now for the poster is to model the steps to do (blue), conditions need to be met for that step (grey) a transaction (green) with a certain system (orange). In the simple puzzle setup there are shapes for secure transactions (https) and conditions.

That is of course only a toy example of one possible way of modeling such steps.

Research-wise it is interesting to investigate in the use of shapes, finding out when they are useful and when it is better to use other means of representing constraints. If shapes are used it is also interesting to investigate if shapes shall be only applied one after each other or if they can be combined somehow. Can we derive new shapes? Also Reasoning becomes interesting, given different context constraints and other user constraints, can we infer new shapes?

A lot of research to be done! I am looking forward to get a lot of interesting feedback in the poster session.

WebSci 2018

Sat, 23 Jun 2018 11:27:27 +0200

At the end of May I had the chance to attend the WebScience conference. A venue where computer scientists and social scientists meet! ACM Turing Lecture by Sir Tim Berners-Lee, adaptive learning analytics, roots of radicalisation on Twitter, internet regulation in Russia. These are only some of the very interesting topics. Continue reading for the full travel report.

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

Rather than chronologically list all the talks, I will focus on one or two talks per day which I liked the most and the social activities. Let’s first start with a bit of context, what is the WebSci conference?

The WebScience conference is all about the Web, the largest socio-technical network in human history, and aims to bring together researchers from various disciplines like computer science, sociology, economics or psychology. This year’s WebSci was already the 10th edition and one special agenda point was the ACM Turing Lecture held by Sir Tim Berners-Lee the inventor of the Web (beside the 150 conference attendees, 1,500 guests were expected).

What I liked the most The interdisciplinary approach of the conference. Algorithms and data meet sociological and psychological use-cases.

What I didn’t like Each session had a scheduled Q&A panel, but none of the sessions I attended made use of it and additionally some talks were out of sync across sessions which made it difficulty to switch talks within a session.

LILE workshop

I already attended the program on Sunday, as I had to present our work about Linked Data Generation for Adaptive Learning Analytics Systems in the Linked Learning workshop (LILE). Past editions were held together with semantic web conferences (ESWC or ISWC) and the WWW conference. This year focused more on interdisciplinary approaches (hence together with WebSci), which was also reflected in the scoring for the best paper award which was won by Simone Kopeinik and colleagues and their work about the adaption of an open source social bookmarking system to observe critical information behaviour. Congratulations!

A keynote was given by Inge Molenaar about Self regulated Learning, who early in the beginning of the talk also raised the question: “Which processes we want computers to take over?”. Later in the presentation I learned that self report is not a really good method to measure self regulation.

There was also a Keynote by John Domingue about Using Blockchains to Open Higher Education. I learned among other things about Uber style universities, which are basically virtual universities where a Blockchain is used to record and verify students ‘attendance’ in online courses, grades and diplomas.

Monday

The first official conference day started directly with the “Best of Web Science” session, where Miriam Fernandez talked about understanding the roots of radicalisation on Twitter. Their approach: “Don’t try to detect whether someone is radical, but measure the influence on the micro, meso and macro level”. Interesting fact is, that the terror organization ISIS has a weekly magazine which is now used by researchers to learn word embeddings for further research.

Another talk from that session focused on third party tracking in the mobile ecosystem. Reuben Binns pointed out that parent companies relationships often are data sharing relationships and that one app usual has 5 trackers installed. Their x-ray transparency tool is available at GitHub proudly presented to you by Mircosoft 😛 (which by the way was also mentioned in the talk as one most prevalent tracker company).

Later that day I mostly attended talks in the “Flow, Information and News” session. Interesting there, was the investigation of effects of Google’s result pages for evaluating the credibility of online news sources given by Emma Lurie. I learned that snopes.com performs fact-checking and that some study participants value social media signals like facebook ranking or tweet recency. More information can be found here.

Monday was also the poster session day. On the Web Science spectrum everything from the online media discourse in authoritarian elections in the case of 2016 St. Petersburg electoral campaign to Linkflows: enabling a web of linked semantic publishing workflows, everything was covered. The best poster award was given to Pip Thornton and her poster about {poem}.py. Famous poems or texts printed on a receipt next to their current monetary value determined by GoogleAdWords.

The day was rounded out with a social gathering and dinner for PhD students sponsored by Web Science Trust. It was tasty and fun, thank you!

Tuesday

Tuesday was in the shadow of the ACM Turing Lecture held by Sir Tim Berners-Lee (the inventor of the web). Next to the 150 conference participants around 1,500 more guests were expected. The difference between the Internet and the Web, according to Tim: When retrieving an server error while requesting a website, one can clearly see that the Internet works, but the web part has issues. An important take-home message for me was the call to make use of our rights and protest against initiatives which are threaten the Web, like cancellation of the net neutrality.

In the session “Methods and Practice” Alexander Darer was talking about the automatic discovery of internet censorship by web crawling. His colleagues and him tried to make transparent how countries block the Internet. One recommendation from the audience regarding the performed research was to take the initial seed list of blocked websites which was taken from Wikipedia and try to find similar sites using techniques like cluster analysis. Also a deeper analysis of the distribution of types of blocked websites was suggested (e.g. gambling, news, entertainment). This social science viewpoint on the data, which can improve the performed research, is in my opinion one of the main benefits of events like the WebSci!

Tuesday was also the day of the official social event of the conference. We had a nice boat tour (unfortunately with a bit of rain) through the canals of Amsterdam and ended up having dinner in the restaurant Ij-kantine, where also the awarding of best papers happened.

Wednesday

The last keynote of the conference was held by John Domingue and was about the future of semantics on the web. After going through some history of semantics, he talked about real-world issues regarding provenance and centralization. The windrush generation in the UK are British citizens who came to the UK from the Commonwealth. Their rights were guaranteed in the Immigration Act of 1971. But now according to a new law, these people have to prove continuous residence in the UK since 1973. A task which is hard to achieve if you don’t have detailed provenance of what you did in those years. The Home Office also destroyed the landing cards (which were centralized stored), which makes it even harder. The keynote was also about machine learning algorithms and how they can use RDF triples in a vector space model.

Beside the mentioned facts, my take home message from the talk was the position of semantics to sit in between machines and humans. Also that even though - as pointed out by Miriam Fernandez - semantics are subjective, the bias is at least visible and provenance can be expressed to see where the bias comes from.

Another interesting talk that day was Internet Regulation Media Coverage in Russia: Topics and Countries given by Anna Shirokanova, where I learned that a small blogger sphere went to a huge politics sphere in Russia. In the beginning of the 2000s a lot of Blogs existed which were also critically talking about the Government and that later on multiple laws were introduced to e.g. make bloggers recognize themselves as media companies (with regulations and obligations involved) or laws regarding the storage of data of Russian citizens and an anti VPN law.

Gefion Thuermer was talking about online participation within the Green Party of Germany. According to a survey they made, young people think that online participation is a good thing for elderly, as they are not that mobile anymore. Similarly elderly think that online participation is a good thing for younger people as their are digital natives anyway. Future research will be a panel survey before and after the online participation tool gets introduced.

That was the WebScience conference from my perspective. As already pointed out, a great venue where Computer Science and Social Sciences meet. I met a lot of interesting people and hope to attend next year as well!

Transparency in Legislation, a must-have

Thu, 13 Jul 2017 00:00:00 +0000

The cabinet of Germany has the plan to publish the drafts of laws and associated position papers of the lobbies. This should be done for more than 600 laws of the past 4 years. This transparency can reveal a lot. For instance, which Lobby influenced which law actively. Nearly 17,000 documents are planned to be published. But how can one handle this huge amount?

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

4 years, over 600 laws, nearly 17,000 documents. These numbers are currently everywhere in the German media. The law “Informationsfreiheitsgesetz”, enables all citizens to request administrative data of the government. This include documents of the origin of law, like the first draft, the position papers of the lobbies regarding this draft or the later draft of the government, which considers the position papers. Over 1,600 citizen requests were just received within a week, after the portal fragdenstaat.de (ask the state), launched the campaign #GläserneGesetze (transparent laws). That are more requests, than received in the whole past year. The administrative overhead seems overwhelming, according to Netzpolitik.org, the German cabinet plans to gradually publish all law related documents of the past legislative period, which is 4 years.

Why transparency is a must-have?

In a democracy, the citizens are voting politicians to office. These politicians have to make decisions on a daily basis, which directly affects the citizens due to the created laws. But a politician is also just a human and cannot be expert in anything. After a first draft of a law is created, various interest groups, lobbyists, give their input. The respective groups are usually closer to the matter of the law. For instance the group BITKOM (the federal association for information technology) is active for subjects around IT. Besides represent direct members, these groups also represent companies. And as the name interest group suggests, these groups have interests. This is where transparency comes into play. Was a law improved, due to the technical competence of the lobby, or were financial interests the main reason and the citizens have to suffer?

How to handle 17,000 documents?

Who, did when, how participate in the development phase of a law. These meta information, which is called provenance, provide a real added value to assess a law. Provenance is the focus of my research. If the documents of the law development process are published in a digital format, they can be further processed. With techniques of the semantic web, raw data can be semantically enriched.

One with specific questioning is then able to search the data efficiently with a computer program. The sentence “The apple is red” for instance, can be semantically enriched, by saying an apple is a fruit and red is a color. Assumed the whole dataset (all sentences) is semantically enriched with the same technique, specific questions can be answered. E.g." Give me all sentences, in which red fruits occur" or “Give me all sentences which are about fruits”.

With these techniques, simpler or more complex questions regarding legislation and the involved organizations can be answered.

Preparing Travels Abroad

Wed, 28 Jun 2017 00:00:00 +0000

You are traveling abroad to a conference and you think you are prepared? Think again! There is always something you probably will forget about and if some unfavorable circumstances following each other you may find yourself in real trouble. Check out my list, learn from my mistakes and extend it, in case I missed something!

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

In most of the cases a good preparation seems time consuming and who need all this, nothing will happen, right?! On a recent trip to Seattle in the United states the worst case happened, I lost my passport. A lot of steps needed to be done, but outside of the usual environment obvious, self-evident things become challenging. Let’s go through my use-case to check what is needed. A small remark, I heed some of the mentioned points, but for some I didn’t even saw the problem arising.

Carry copies of important documents

Having copies of all important documents. Depending on what disappeared, copies can save your life. Always make copies of your passport, your ID, driver license, Visa or other travel authorization documents like ESTA and carry it with you in printed form or digital. In my case I still had my ID to identify myself and a printed version of the approved ESTA application. This document included my passport ID and allowed me to log-in online, in order to find out other important passport information like its expiry date. This information can help speed up the process of requesting a temporary passport from your country’s diplomatic mission.

A short introduction to diplomatic missions. Usually your country has an embassy in the capitol of other countries. But you don’t have to travel to the capitol in order to get a temporary passport. Your country may operate a hand full of consulates distributed across the foreign country. In my case the next consulate was in San Francisco which is still around 1.300 km away! There could still be some honorary consulates nearby, which act in behalf of your country, google for that! In my case there was one in Seattle, but unfortunately they were on vacation! That’s why I had to travel to Portland, which gave me time to write this blog post. Ask also your Airline, for example the German Lufthansa used to take German citizens back to Germany just with the ID.

Phone calls

In any case, you have to make some calls in order to arrange everything, e.g. the consulate or the airline. We are living in the 21st century, nearly everybody has some sort of a phone. Conversely, there is no need for phone booths to make calls anymore. But abroad you may have to make calls, but your provider may charge significant fees! The amount you are usually paying for your whole monthly subscription will may be due for a single minute! (For me it was around 6 Euro/min) Load money on your Skype account, WIFI is nearly everywhere. Of course you can ask around if you can use someone else’s phone, at your hotel’s desk or at random stores, but being independent is much more helpful, especially if you are around in the city!

Activate your debit/credit card’s fully functionality

Money rules the world! In general, having cash is very useful. Your bank might deactivate your debit card’s functioning in foreign countries. So, activate your debit card in advance for the use in the foreign country! I fortunately did this.

In case you your consulate is not around the corner, you might have to undertake a travel there. Trains or especially intercity buses are booked online. Sometimes online is the only way of purchase, or if not, you might have to pay a much higher price at the station or from the driver. Also sometimes, credit cards are the only possible payment method. Both, Visa and MasterCard offer fraud detection techniques like “Verified by Visa” or “MasterCard SecureCode”. If you aware of it, it can be your life vest. If you are not aware of it, this life vest might strangle you! If these functionality isn’t activated beforehand, you cannot use the cards in online shopping environment, like your ticket purchase. The activation can take multiple days, so activate your credit card’s online shopping functionality beforehand. Luckily I had friends and colleagues willing to helping me out!

Summarized preparation

Make copies of important documents beforehand
Load money on your Skype account
Activate debit card for the duration of your stay for the foreign country
Activate your credit card’s online shopping capabilities

Please comment if you can think of something I maybe missed.

What is Provenance?

Fri, 07 Apr 2017 00:00:00 +0000

Having Provenance or not having it, that could be worth than several hundred million dollar. Where does something come from, how was it changed since it exists, that is provenance. These informations help human beings to asses things. Which questions need to be answered and what is my role as computer scientists in this research field?!

If you want to get notified about more content or want to receive bi-weekly updates on FAIR and Linked Data, please consider subscribing to my bi-weekly Newsletter.

In 2015 the painting Les Femmes d’Alger from Picasso was sold for unbelievable 179 Mio Dollars. The painting La Bella Principessa from Leonardo da Vinci was sold for surprisingly cheap 20,000 Dollars (see Ludäscher). But what is the main difference, why did the first painting obtained such a high price compared to the second? The answer is, that there is a peace of paper which documented every owner and former places of custody of the first painting. The same information for the second painting was incomplete. This documented history of ownership is called Provenance.

Why is this Provenance so important to us?

Well, let’s say I sell a car and you want to buy it. If I can provide you a service record which documents how often the car was checked in a garage the last years you have more trust in its functionality. In the same manner is it also from interest how the car was changed, which parts are newer because they were replaced for example. Of course there is no direct trust involved, but the provenance helps to form an opinion and allows one to draw conclusions about the object. Other properties than quality can be important as well, the value of the car might also be higher if the former owner was a celebrity.

The car example shows the importance of provenance and the millions of dollars price difference of the paintings that it is definitely worth to capture! There are plenty of use cases where provenance is helpful, just think about the food you are eating, where does it come from? How was it produced? Was the cooling chain never interrupted? All these information influence your buying decision. Providing these information is therefore also in the interest of the producer. There is actually a start-up company called Provenance, which tries to make the supply chain of products transparent. Information if the food comes from regional producers or information if the employees were treated fair justify a higher price. Whereas this kind of provenance help to address a more selective client base other provenance like the consistent cooling chain is enforced by law and you have to proof it in case of an audit.

Provenance and Computer Science

I am a computer scientist and not a quality assurance officer, where is the link to Provenance? We are living in the 21 century, no one is using pen and paper to record the Provenance. Machines of the supply chain, Smart Sensors of the Internet of Things (IoT) are measuring and transmitting the provenance information. The provenance needs then to be stored in a queryable way . Why? Continue reading … Imagine you are a pharmaceutical company, you keep track of the bottling of the medicine and to which drugstore each bottle was sold. Let’s say you have all the provenance just split into hundred of thousands of files somewhere on an external hard drive in an custom data format, nobody knows about. After encountering a possible contamination of some of the produced drugs, you need to know where they end up so you can take selective only the affected bottles from the market. You don’t want to waste time and money to instruct one employee to carefully search for the information within the files, which can take multiple days. Having the Provenance stored in an machine understandable format which is queryable, will give you the answer within seconds! Using an appropriate data storage instead of an external hard drive ensures availability and prevents data loss. Having the provenance in an standardized data format allows the usage of standard software instead of investing time and money to reinvent the wheel. o

Overall, standards are important! Imagine your smartphone runs out of battery, you are at a friends place, but your charger doesn’t fit into the electrical socket because every house has another socket type. Thankfully there is a standard and every electrical socket within a country or larger region is built upon a standard. The same applies for digital standards, you can probably open PDF documents on every of your devices. For Provenance the World Wide Web Consortium (W3C) released a standard called PROV.

Speaking of the world wide web, everybody of us produces and/or consumes so much information. Just think about all the blogs and news sites producing content every minute. Journalists normally investigate carefully and check the reliability of all their sources. But classical media are nowadays only one possible source from which we get information from. Social media like Twitter or Facebook allows everybody to create content and offer it to a broad variety of people. How do you trust if it is real what is written there? Political oriented groups can spread so called fake news to influence the opinion of a lot of people. Who actually takes the time to really read an article and check its sources?! A dramatic headline together with some emotional pictures will get much more response than an article based on facts from boring statistics. Additionally, such articles also generate a lot of ‘clicks’, meaning that people click on them, visit the site and are seeing some special ads and the site operator gets a lot of money. Reason enough for some people to just make up stories, only interested in the money they will get.

This whole process of information spreading only partially produces provenance, but the provenance is somehow there. It is implicitly hidden in the timing, the textual similarity or the social connections between the participants of the spreading. The provenance can partially be reconstructed. That brings us directly to the last part, the research about provenance. As a computer scientists I do actual science. Research about what kind of provenance can be reconstructed from social media and to which extend is this useful? This is only one example, focusing on social media. Also general aspects about provenance are not fully explored right now. How to handle provenance, how to secure the provenance records itself against forgery. Provenance sometimes means total transparency, to which extend is this desirable, at which point the privacy of individuals gets affected? How can new technologies help to answer some of those questions?

You hopefully have now an idea what provenance is. What do you think, where else could provenance be useful? Leave a comment!

Posts | Sven Lieber

ENDORSE 2025

Context

Data stewardship in the Age of AI

Interoperable Public Services and Data Spaces

Provenance and Agreement

MetaBelgica: central vs decentralized data management

Final remarks

References

CoRDI 2025

Context

Keynotes

Rob Finn - Life Science data

Cathrin Stöver - Research and Education Infrastructure

Persistent identifiers, Authority data & Knowledge Graphs

PIDs: Small links, big impact

Authority data in the humanities

Knowledge Graphs and Ontologies

AI

Infrastructure

Governance

Global governance

Governance in practice

Training & Communities

Reflections & Takeaways

References

Clustering Book editions

Context

Better user experience and better data quality due to identified works

Alright, you convinced me, how can I do it?

Nice idea, how do I write a program to do this?

First attempt: reusing the Python package sklearn

Second attempt: implementing the algorithm myself

Third attempt: inverted index for clustering in no time

Why is it so fast?

Final remarks

CoRDI 2023

Context

Keynotes

Christine Borgman – Knowledge Infrastructures

Julia Janssen – data art installations

Mark D. Wilkinson – FAIR principles

Poster session(s)

MetaBelgica, Wikibase and NFDI4Culture

Knowledge Graphs at NFDI

Selection of presentations

Social sciences

Data spaces

Physics

Scholarly Communication

Research Data Management

Networking

Final remarks

RDF named graphs

Context

Example of named graphs for provenance information

Illustrations of different ways to query named graph data

Final remarks

Knowledge Graphs for Data Integration - UHasselt 2023

Context

Keynote

Explaining predictions of ML models

A logic-based query language for ML models

Foundations

Graph compression

Effective reasoning

Reliable query languages

Experience and exemplars

Data dependencies

Rapid Prototyping

Technology stacks Knowledge Graph

Law as code

Processing and mining

Mining software engineering patterns

More efficient SPARQL queries

More on query optimization

Final remarks

DHBenelux 2023

Context

Data platforms and usability