ISWC 2026 Trip Report

Between the 25th and 29th of October 2026, the 25th edition of ISWC took place. The first two days were workshop days, and since 2024, ISWC has introduced Dagstuhl-like workshops. I really enjoy these workshops, as it brings together a group of people for deep discussions on emerging topics or ongoing challenges in the field. This year, I attended the CHARGED workshop, which focused on challenges of multimodal data in cultural heritage, and fidelity: loss of information with digitisation. We also discussed concept drift, a topic that I find very interesting due to its complex nature in combination with the use of ontologies, as concepts change over time. For instance, occupations shift in meaning over time, or systems for determining geolocation. Additionally, we discussed uncertainty: a lot of information is lost in historical sources, how can we still best represent them given this missing knowledge?

On the second day, the Sci-K workshop (which is now in its sixth edition) that I co-organised took place. Silvio Peroni gave a great keynote, advocating for sustainability plans for open scholarly infrastructures, that are under pressure from funding cuts and geopolitical instability. In terms of submissions, we saw returning topics such as tools for documenting ontologies and representing scholarly discourse more richly (LODE, SUDO), and LLM-based extraction pipelines for scientific metadata. This year, we also had two papers discussing how invalidations/retractions should be handled in scholarly knowledge graphs. From the submissions and discussions during the workshops, I think a broader question is one about the forms scientific knowledge should be communicated in, now that LLMs generate, extract, and consume scholarly knowledge at scale. Should papers combine human-readable text with machine-readable structured data (e.g., [1]), ideally with data and interactive visualisations, or should they perhaps be published as interactive systems (e.g., [2])?

The main conference opened on Tuesday with Frank van Harmelen's keynote "It was never about the Web: rethinking semantics at 25". Frank first celebrated what the Semantic Web succeeded at over the last 25 years: data integration, knowledge graphs in biotech, and interoperability, so semantics! But, the talk could not all be celebratory: he argued that the vision of the intelligent web agents, and mostly the 'Web' part, did not materialise. He talked about Robert Dijkgraaf's "curse of applicability", and how that hit our community as well: it became so good at building useful things, perhaps spending less attention to the fundamental questions. Therefore, Frank discussed the scientific lessons we've learnt over the last 25 years: (1) graphs are the natural forms of information, and (2) a little semantics goes a long way. The second part of his keynote looked forward, arguing for rethinking semantics. Rather than negotiating meaning before the problem-solving process, between a knowledge engineer and a domain expert, why not involve a problem-solving agent as an active participant in the meaning-negotiation process? So, instead of knowledge engineering as a separate activity that is decoupled from the problem-solving process, Frank suggests a shift towards knowledge engineering as an ongoing, interactive meaning negotiation process between domain experts, knowledge engineers, and AI agents, that is driven by the needs of downstream problem-solving tasks. What I like about this vision is that it might help address some of the difficulties we face when publishing static knowledge artifacts, given the dynamic nature of knowledge. If meaning negotiation becomes an ongoing, interactive process, changes, errors, and differences in conceptualisations across disciplines could be identified and addressed as they arise during problem solving, rather than requiring separate knowledge management activities (and it relates a little to [2] again).

In the afternoon, Jim Hendler's keynote revisited the original 2001 Semantic Web vision, evaluated what had actually been achieved over the past 25 years, and argued for recovering that vision using today's technologies, particularly LLMs and AI agents. His keynote ended with a sketch about the industry rediscovering the Semantic Web in venture capital language (with 'Global Data Handles' (URIs), and 'Enterprise Data Intelligence Mesh' (a KG)), starring Juan Sequeda as "Chad Stack". The keynote was followed by a panel with questions from early career researchers and the audience, about trust erosion and self-reinforcing biases and errors with LLM-built KGs; questions about the role of humans in knowledge generation (essential, I think, since knowledge is grounded in a reality that keeps changing); and accountability in public knowledge infrastructure.

I greatly appreciated the visionary track, and (even though I understand it would perhaps be too often), wouldn't mind seeing it back every year. This year, the talks I saw were 'unrethinking the semantic web', which argued that after 25 years, the point is not to rethink the semantic web, but to reclaim it, given the state of current technologies. A second vision paper argued for work on brain-computer interfaces that make human knowledge stored in the brain interoperable with knowledge on the Web (raising as many ethical questions as technical ones). Lastly, a paper about going from a Web of documents to a Web of claims argues that the web has moved from documents to data, and the next layer is claims: the Semantic Web community solved the representation problem, but the infrastructure was never populated at scale. Generative AI makes population cheaper, but throws away the claims every run. However, as we saw from Silvio Peroni's keynote during the Sci-K workshop, the infrastructure to store and accumulate claims already partially exists (the Open Research Knowledge Graph, OpenAIRE, and others) but we still lack good ways to populate it at scale.

Lastly, kudos to the general and local chairs, Valentina Tamma and Claudia D'Amato, and the rest of the organising committee; both the scientific program as well as the social program were great.

To close, some papers that I found interesting from this year's ISWC:

References

  1. Beek, W., Mourits, R., and Rijpma, A. "FAIR Paper: Applying FAIR to Academic Publishing." CEUR Workshop Proceedings, Vol. 3724. CEUR WS, 2024.
  2. Miao, J., Davis, J.R., Zhang, Y. et al. Reimagining research papers as interactive and reliable AI agents. Nature 658, 1272–1280 (2026). https://doi.org/10.1038/s41586-026-11044-y