Talks

Andrea Scharnhorst

Building knowledge bases - a human-centered AI supported approach for long-term archiving

Each collection is specific, brought together for specific reasons, and usually dedicated to specific ‘user communities’. From the field of knowledge organization (KO) we know how each work and each collection of works represent specific knowledge. With the choice of a Knowledge Organization System (KOS) to organize this content and to make it FAIR (Findable, Accessible, Interoperable and Reusable), specific aspects of this knowledge are brought to light. Digitisation and automatisation has put the spotlight on issues of scalability, machine interoperability and standardisation when it comes to indexing content. Still, the human need for specific approaches remains. This presentation demonstrates the dialectics of generic and specific approaches when it comes to the long-term presentation of artefacts. We look into the case of a repository which heals aspects of inclusion and accessibility while dealing with a variety of content from and around performing arts (the case of MuseIT project). We discuss how the almost eternal tension between the more general and more specific aspects of human knowledge is re-shaped when ‘machine knowledge’ comes into play. How does this tension or contradiction play out in the presentation of artefacts? Is it due to the specific selection of objects or things presented? Or is a specific meaning intended to be conveyed by the presentation of artefacts? We also discuss how AI can be used in a human-centred way, to offer support when it comes to navigate diversity without erasing diversity and specificity.

Marcia Zeng

The role of KOS in AI-supported semantic integration: disambiguation, identification, linking

Advanced AI-supported semantic integration tools can be very helpful in processing news and formal publications, exploring vast cultural heritage resources, or analyzing digitized historical documents. They can perform semantic analysis, content categorization, entity and relationship extraction, identity matching, and real-time processing. Moreover, they can enable effective communications through structured data, interoperable mind maps and concept maps, semantic reasoning, path inference, in addition to Q&A and content summary. Recently, embracing AI has become very important in cultural heritage where we can see how technology can transform the way we interact with the past. AI opens new avenues for research, improves accessibility, and enhances preservation efforts. At the same time, challenges and issues can be found when cross-linguistic and cross-cultural processing are involved. Through the exploration of various cases, this presentation discusses the important role of KOS in AI-supported semantic integration, especially when it comes to eliminating ambiguity, enhancing the quality of identification, and enabling the effective cultural heritage entity connections across scripts, languages and cultures.

Ziyoung Park

Co-creating KOS with GenAI and expert review: empowering KOS projects through design, automation, and recommendations

This talk demonstrates how generative AI, combined with expert review, can streamline the design, automation, and recommendation processes in building and maintaining the Korean Knowledge Organization System (K-KOS) registry. Based on over 600 Request for Proposal (RFP) documents collected from national procurement platforms, the project integrates structured metadata management in Notion with publication in a MediaWiki-based KOS registry. Generative AI supports three main tasks: (1) page and script design for MediaWiki using ChatGPT-4o and GPT-5, (2) automated conversion of Notion API outputs into ready-to-publish wiki content, and (3) structural summarisation of RFPs with Notion AI. Human experts define metadata schemas and mapping rules, review AI-generated drafts, and finalise updates. Future developments include direct MediaWiki API integration for automated publishing and selective updates based on change tracking. The talk will also share practical checkpoints for effective AI use, including context provision, output verification, and transparent labelling of AI- generated content. This case illustrates a sustainable and adaptable model for enhancing KOS registries in resource-constrained environments.

Ronald Siebes

Potemkin Understanding: the illusion of thinking

A recent article from researchers from MIT and Harvard University with the title "Potemkin Understanding in Large Language Models" shows that LLMs (at least the current 'big ones') are often excellent to explain concepts, but if you ask them to apply a perfectly explained concept to a given task, it fails in a way a human would not. In other words, an elaborate façade of competence that masks a deep, non-human-like void of true conceptual grasp. In this talk we explore some of these failures of an LLM applying knowledge it pretends to have. Also we emphasize that it is not so innocent when people are not aware of this, especially when applied in real life systems and research. Next, we will discuss the ways to expose Potemkin understanding from LLMs via a new benchmark paradigm and show some recent work that tries to improve LLMs towards 'real' understanding and its relation to symbolic AI, and Knowledge Graphs in particular.

Angelo Salatino

Knowledge organization systems of research fields: overview and automatic generation

Knowledge Organization Systems (KOSs), such as term lists, thesauri, taxonomies, and ontologies, play a fundamental role in categorising, managing, and retrieving information. In the academic domain, KOSs are often adopted for representing research areas and their relationships, primarily aiming to classify research articles, academic courses, patents, books, scientific venues, domain experts, grants, software, experiment materials, and several other relevant products and agents. These structured representations of research areas, widely embraced by many academic fields, have proven effective in empowering AI-based systems to i) enhance the retrievability of relevant documents, ii) enable advanced analytic solutions to quantify the impact of academic research, and iii) analyse and forecast research dynamics. In this talk, I will briefly present the outcomes of a recent survey paper in which we analysed and compared 45 KOSs of academic disciplines according to five main dimensions: scope, structure, curation, usage, and links to other KOSs. Our results reveal a highly heterogeneous scenario in terms of scope, scale, quality, and usage, underscoring the need for more integrated solutions to represent research knowledge across academic fields. Then, I will present our ongoing activities for developing such a comprehensive and integrated ontology of research topics.

Mayukh Bagchi

Generative knowledge organization via human-LLM collaboration

Knowledge organization systems, including taxonomies, metadata schemas, ontologies and now knowledge graphs - have consistently underpinned information source, systems and services in information institutions. In the era of Artificial Intelligence (AI), these models are even more critical for high quality and explainable knowledge and data representation, yet their reusability and interoperability remain constrained by the assumption that a small set of monolithic, universal models can serve all domains and goals. This talk advances a generative AI-driven approach to knowledge organization that reconceives these models as multi-perspective, multi-level representational systems. It shows how manifold, overlapping representations create representation entanglement, and proposes a generative AI-human collaboration framework to disentangle them. By aligning knowledge organization with adaptive, context-sensitive and AI-relevant principles, the presentation outlines a path toward more flexible infrastructures like the recently proposed Knowledge Organization Ecosystems (KOEs).

Julaine Clunis

From black boxes to transparent AI systems: balancing innovation and responsibility using knowledge organization systems

The interplay between artificial intelligence (AI) and knowledge organization promises exciting innovations such as automated ontology development, AI assisted cataloging and cross-lingual classification, intelligent metadata generation, automated subject indexing and more. However, it also raises urgent ethical questions. This talk will examine how AI-driven systems can inadvertently perpetuate bias and discrimination, for example by reinforcing sensitive social biases in classification and recommendation algorithms. Concurrently, opaque “black box” AI models pose challenges for transparency and accountability, making it difficult for stakeholders to trust or understand automated decisions. However, ontologies, knowledge graphs, and other knowledge organization systems (KOS) can serve as transformative tools that reduce algorithmic bias and enable models to explain their outputs and reasoning in human-understandable terms. We will explore three key factors: (a) fairness and bias – how algorithmic discrimination in AI can marginalize certain groups, and how inclusive knowledge organization practices can help counteract these biases; (b) transparency and explainability – how the use of ontologies and knowledge graphs can turn black-box models into more “glass box” systems by providing context, semantic reasoning, and traceable decision paths; (c) accountability and human oversight – why human knowledge professionals remain essential in the loop to ensure ethical standards, cultural sensitivity, and quality control. Drawing on recent research and tangible cross-domain examples, this talk will highlight how knowledge organization principles can both mitigate the risks and enhance the benefits of AI, supporting the development of systems that are not only intelligent but also fair, transparent, and trustworthy.

Joane Casenave, Widad M ElHadi & Thibault Grison

AI, KO & information mediation on the Web: from ethical dimensions to social responsibility

While traditional documentary environments organize knowledge using established classification systems, knowledge organization on the web follows new and diverse paths. In our presentation, we will begin by addressing the ethical challenges posed by automatic classification and the use of algorithms in web-based knowledge organization. We observe that, in this context, knowledge organization involves both automated and participatory elements, which are not isolated but rather interact and strengthen each other. These intertwined practices can disrupt the functioning of knowledge organization systems, giving rise to various forms of bias, including intersectional biases. Consequently, they can perpetuate forms of epistemic violence such as discrimination, exclusion, and marginalization. We will illustrate these dynamics with examples from our current research. The second part of our talk focuses on information mediation on the web, specifically examining how AI-driven personalization and recommendation systems can restrict users’ exposure to diverse viewpoints. Techniques like filter bubbles and echo chambers, used in applications such as opinion polling, for instance, can narrow users’ choices. It is essential to recognize the limitations of AI algorithms and the potential negative impacts they may have on information mediation. We also stress the importance of the social responsibility of designers of information access systems and tools. Dealing with these issues is crucial, requiring us to design, deploy, and manage AI in ways that reflect societal values and promote the common good, while carefully considering and qualifying the risks and unintended effects on individuals and society as a whole. Social responsibility is a dimension of ethics that has become necessary in the face of modern times, which in turn have brought new techniques, new objects, consequences, and social relations that the old framework could not accommodate. The question of social responsibility, although largely debated in our domain, has also come to be seen in traditional professions such as archival science and library and information science, giving rise to the need to reflect on professional and social ethics from a new dimension: that of responsibility. The question of social responsibility, although largely debated in our domain, is especially resonant and relevant when it comes to the use of AI algorithms and generative AI.

Tony Russell-Rose

Building and Deploying LLMs for Search and Retrieval

This talk explores the use of language models to support academic search and systematic review, focusing on the generation of query suggestions through knowledge-based methods, context-free models, and large language models (LLMs). Drawing on two complementary studies – an offline evaluation using real-world search strategies, and an online user study - we compare the effectiveness and user perceptions of various approaches. We also reflect on the practical challenges of deploying NLP systems in production, and share insights from our ongoing migration from custom-built models to hosted LLM services, with implications for scalability, cost, and development practices.

Joseph Busch

The case for general purpose categorizers as part of the AI ecosystem

AI developers are simply not equipped to deal with the intricacies of assessing and selecting KOS and integrating them appropriately into general purpose applications. It has been left to consultants or researchers to develop custom applications for niche industries where the benefits of more precise retrieval outweigh the difficulties and costs of KOS integration and development. However, I think there has always been an interest in general purpose categorizers for common business functions, just as there is an interest in universal classification systems for organizing general content collections. In this talk I discuss the origins and evolution of today’s AI ecosystem and argue that general purpose categorization systems would be welcome as part of the AI ecosystem assuming that they are intuitive, easy to use, kept up to date, and free.