The learning objects recommending systems designed in the last decade employ different recommendation techniques, which mainly use collaborative filtering and content-based filtering. This paper aims to analyze a new recommendation approach, that has not been yet applied to learning objects communities, called critiquing-based recommendation. This approach gathers the users preferences through an interaction based on examples and critiques. This feedback helps the system to recommend the most suitable learning resources. In this work we propose an hybrid recommender system of open educational resources that combines collaborative filtering and critiquing-based techniques. A prototype using MERLOT data and considering just one round of interaction was built in order to evaluate the recommendation results amongst a small group of volunteers. Preliminary results pointed out that the relevance of recommended learning resources increases as the system receives the users feedback and generates new recommendations based on it.
STEaMINg : modèles sémantiques qui évoluent dans le temps pour l'industrie 4.0: Dans l'industrie 4.0, les machines des usines sont équipées de capteurs qui collectent des données pour une surveillance efficace de l'état des équipements. C’est une tâche difficile car elle nécessite l’intégration et le traitement de données hétérogènes provenant de différentes sources, avec des résolutions temporelles et des significations sous-jacentes différentes. Les ontologies apparaissent comme une méthode pertinente pour traiter l’intégration des données et pour représenter la connaissance de manière interprétable par les machines grâce à la construction de modèles sémantiques. De plus, la surveillance des processus industriels dépend du contexte dynamique de leur exécution. Dans ces circonstances, le modèle sémantique lui-même doit évoluer afin de représenter dans quelle(s) situation(s) se trouve(nt) la ou les ressources pendant l’exécution de ses tâches pour soutenir la prise de décision. Cette thèse étudie l’utilisation des méthodes de représentation des connaissances pour construire un modèle sémantique évolutif qui représente le domaine industriel, en mettant l’accent sur la modélisation du contexte pour fournir la notion de situation.
Document-level relation extraction (DocRE) aims to identify semantic relations between entity pairs by aggregating evidence that may be dispersed across multiple sentences. Large language models (LLMs) offer a flexible alternative to supervised DocRE architectures through zero- and few-shot prompting, yet they remain prone to hallucinated relations, incomplete evidence aggregation, and unstable multi-hop reasoning. This paper presents a structured empirical evaluation of LLM-based DocRE strategies, spanning direct prompting, ontology-guided retrieval, knowledge graph–extended RAG, and planner-driven agentic reasoning, using a large-scale knowledge-rich benchmark as a representative use case and focusing on the development split for reproducible analysis. We compare prompting baselines with progressively grounded approaches that inject structured semantic constraints and external knowledge during inference, while maintaining a shared preprocessing pipeline, output contract, and micro-averaged evaluation protocol. All methods are implemented within a unified open-source framework that standardizes prompts, retrieval modules, and scoring. The comparison establishes strong prompting baselines and demonstrates that planner-driven semantic grounding can yield measurable gains, particularly through improved recall under controlled and auditable tool use in document-level relation extraction.
The use of sensor data to support monitoring and decision-making has become widespread across many application domains. Continuous data streams enable early detection of relevant situations, such as anomalies, but identifying and explaining their underlying causes remains a significant challenge. This paper proposes a unified framework that integrates stream reasoning and concept induction to enable real-time detection of abnormal situations and automated root cause identification. The framework relies on semantic enrichment and continuous reasoning over data streams to detect anomalies and to derive human-interpretable explanations. Although the approach is domain-independent, we demonstrate its applicability using a smart farming sensor dataset. The experimental results show that the proposed framework successfully detects abnormal environmental conditions using key variables, such as rainfall and humidity, and provides semantically grounded explanatory concepts that contextualize the detected conditions.