Miguel Couceiro – Universidade de Lisboa
From analogies to analogical pruning: A few-shot approach to Domain-Specific KG retrieval
Large-scale, generic Knowledge Graphs (KGs) such as Wikidata offer vast amounts of premium data, but their sheer size and heterogeneity pose significant scalability and relevance challenges when building domain-specific applications. To retrieve highly focused KGs, it is necessary to extract relevant subgraphs from these massive sources while avoiding topical drift.
In this talk, we will present an application of analogical reasoning to the KG retrieval process. Firstly, we will introduce a novel few-shot analogical pruning approach that traverses the neighborhood of seed entities and decides whether to keep or prune them using proportional analogies (i.e., A is to B as C is to D). By leveraging a convolutional neural network on pre-trained KG embeddings, this method dynamically captures the relative similarities and dissimilarities between entities. Our empirical evaluations demonstrate that this analogy-based model outperforms traditional classifiers like LSTMs with a drastically lower number of parameters and exhibits superior generalization capabilities in transfer learning settings.
Secondly, we will showcase KGPRUNE, a Web Application and API that makes this analogical pruning algorithm accessible to the wider community. By simply providing seed entities and properties of interest, users can automatically extract tailored, high-quality subgraphs. We will explore the practical impact of KGPRUNE through two distinct real-world use cases: the construction of Enterprise KGs for the IT domain, and the extraction of specialized networks to track looted artworks and protect cultural heritage.
Finally, we will briefly survey upon our broader research into analogical competency and case-base maintenance. The results we will present appear in the scope of the ANR project AT2TA and the various collaborations therein, in particular, F. Badra, Z. Bouraoui, F. Cunha, F. Gandon, L. Jarnac, D. Leake, Y. Lepage, M.J. Lesot, F. Malaca, E. Marquer, P. Monnin, E. Mbengue, among others.
Bernhard Ganter – Technische Universität Dresden
Essential Concepts
Essential Concepts is an experimental format that is being tried out for the fi rst time. Instead of a one-hour presentation, I will be offering a series of short talks spread out over several days. I will try to highlight some important aspects of Formal Concept Analysis that, in my opinion, deserve more attention, and I hope to rekindle interest in them. If the Essential Concepts format proves popular, other topics may be covered in the future.
Andreas Hotho – University of Würzburg
From German Language Models to Conceptual Knowledge Structures
Large language models have become powerful tools for representing and generating language in chat-based systems, but the relation between their learned representations and explicit conceptual structures remains only partially understood. Their simple training objective is astonishingly effective and leads to a large amount of knowledge being stored in the weights of such models. A lot of research is currently being conducted to understand the implicit learned representations and conceptualizations of these models, as well as to connect internal representations with explicit concepts. In this talk, I will start from our recent work on LL=C3=A4Mmlein and ModernGBERT, two transparent German model families trained from scratch, and discuss how language-specific models can be built, evaluated, and used. These models provide a controlled basis for studying how language-specific representations emerge from data, training pipelines, tokenization choices, and model architecture, and allow us to explore the relation between emergent and explicit conceptual structures.
Recent work on attribution graphs suggests that large language models may form internal feature structures and computational pathways that resemble concept-like organization. Such analyses reveal intermediate representations, multi-step mechanisms, and abstract features inside the model. Concepts in language models are present, can be used and extracted, but they are not directly available as explicit, stable, inspectable objects or formal concepts. In the talk, I will show the relation between these learned representations and explicit concepts.
I will then connect this view to our work on enriching language models with knowledge graphs for computational literary studies. In this setting, concepts are not only factual entities, but historically changing, culturally situated, and domain-specific structures of meaning. Dictionaries, encyclopedias, lexicons, and literary texts can be transformed into graph-based representations, i.e., knowledge graphs, that capture semantic relations, diachronic concept alignment, character networks, plot structures, and other forms of conceptual knowledge. These resources can enrich language models and support semantically richer representations, especially in domains with limited computational resources.
The talk argues for a concept-centered perspective on language models: learned representations reveal that models organize language through internal concept-like structures, while knowledge graphs and related conceptual representations make selected parts of this structure explicit, inspectable, and reusable. Connecting both perspectives opens a path toward concept-aware language technologies that combine the flexibility of modern language models with explicit structures of meaning.
Fatiha Sais – Paris Saclay University
When AI Generates Facts: Knowledge Discovery for Trustworthy Knowledge Graphs
Knowledge Graphs play a central role in modern AI systems by providing structured, explicit and machine-actionable representations of knowledge. They support data integration, semantic search, question answering, recommender systems and, increasingly, hybrid approaches combining symbolic knowledge with Large Language Models. However, the rise of generative AI profoundly changes the nature of the problem. AI systems now generate large volumes of plausible factual statements, but these statements may be incomplete, ambiguous, outdated, inconsistent or simply false. In this context, ensuring the quality, reliability and temporal validity of knowledge has become a major challenge.
This keynote will present our research on knowledge discovery methods for improving the quality of Knowledge Graphs. I will focus on the discovery of symbolic knowledge such as rules, keys and temporal constraints, and show how these forms of knowledge can be used to validate, refine, align and enrich Knowledge Graphs. Rules can reveal regularities and support inference and keys help identify and align entities across heterogeneous sources; temporal constraints make it possible to reason about the validity, evolution and expiration of facts over time. Together, these discovered structures provide interpretable mechanisms for detecting inconsistencies, completing missing information and assessing the reliability of facts.
I will place these contributions in the broader context of trustworthy and hybrid AI. While Large Language Models are powerful generators of fluent and plausible content, they still have limited abilities to guarantee factual correctness, provenance awareness and temporal reasoning. Knowledge Graphs, enriched with automatically discovered constraints and symbolic patterns, can therefore play a key role in mitigating hallucinations, validating AI-generated facts and supporting explainable fact-checking pipelines. The talk will discuss how learning-based discovery and symbolic reasoning can be combined to build more robust, transparent and temporally aware knowledge systems.
The main message of the talk is that, in an era where facts are increasingly generated rather than only collected, knowledge discovery becomes essential not only for building Knowledge Graphs, but also for making them reliable instruments of trust.